I treated this as a count-of-matching-records problem: estimate total arXiv submissions in the June 4, 2026 through August 11, 2026 window, then estimate what fraction of those submissions will satisfy the very broad abstract query. To anchor the base rate, I first looked at arXiv’s overall scale. arXiv’s 2023 annual report says there were 208,493 new submissions in 2023, about 17,000 per month, with a record 20,710 in October 2023. A later secondary summary of the 2024 annual report reports 244,031 new publications in 2024 and a new monthly record of 24,226 in October 2024. A 2026 paper quoting arXiv usage statistics says that Q1 2025 alone had over 65k submissions; scaled to a 69-day window, that is about 49.8k papers. Using the 2024 annual total gives about 46.1k papers in a 69-day window. So a reasonable total-submission base for this forecast window is roughly high-40-thousands to low-50-thousands. (info.arxiv.org)

The search rule is broader than the title’s plain-English phrasing suggests. arXiv advanced search explicitly supports abstract-field search, quoted exact phrases, wildcard matching with , and inclusive date bounds, and it also distinguishes “Submission date (most recent)” from “Submission date (original).” The bio term set includes very broad phrases such as outbreak, pandemic, epidemic, pathogen, and infectious disease, while the AI set includes generic markers such as machine learning, deep learning, transformer, neural network, LLM, GPT, and artificial intelligence. That means the resolved count should be materially larger than a narrow search for explicit biosecurity-governance papers only. The separate “most recent” versus “original” submission-date options are also a small but real source of operational uncertainty around the final count. (arxiv.org)

To gauge current flow, I looked for recent arXiv examples that would likely satisfy the rule. Recent examples include Physics-informed deep learning for infectious disease forecasting (January 16, 2025), Combining digital data streams and epidemic networks for real time outbreak detection (November 10, 2025), Machine Learning Epidemic Predictions Using Agent-based Wireless Sensor Network Models (November 20, 2025), A Machine Learning Framework for Constructing Heterogeneous Contact Networks: Implications for Epidemic Modelling (March 14, 2026), Towards Practical Multimodal Hospital Outbreak Detection (March 20, 2026), Generative Modeling in Protein Design: Neural Representations, Conditional Generation, and Evaluation Standards (March 27, 2026; includes dual-use biosecurity risks), and Prospective multi-pathogen disease forecasting using autonomous LLM-guided tree search (May 15, 2026). These span infectious-disease forecasting, outbreak surveillance, pathogen-related modeling, and explicit dual-use/biosecurity discussion, which is enough evidence for me that this is an active multi-stream topic on arXiv rather than a niche with only occasional hits. (arxiv.org)

My quantitative translation is that, if the window contains about 50k total submissions, then a final count in the high tens corresponds to only about 0.1% to 0.15% of all arXiv papers. Given the breadth of the Boolean term sets, that hit rate looks plausible. I do not think the left tail below 20 is very likely, because recent examples already show several distinct matching research streams. I also do not want to rule out the right tail above 100, because arXiv has been growing quickly and the query can capture papers whose main subject is infectious-disease modeling or pathogen-related ML, not just papers explicitly framed as “biosecurity.” My central estimate is 59, with mean about 60 and an 80% subjective interval of roughly 39 to 82. I encoded that uncertainty as a 3-scenario mixture: a lower scenario (summer lull / narrower operational interpretation), a base scenario, and a higher scenario (continued arXiv growth plus the breadth of the outbreak/pandemic/pathogen/infectious-disease terms).

I first parsed the resolution mechanics. arXiv help says phrase queries in title/abstract/comments require double quotes for exact matches, wildcards use *, date ranges use inclusive lower and upper bounds, and advanced search can filter by submission date or announcement date. That means the final count will be sensitive to exact wording in abstracts rather than broad topical relevance. (arxiv.org)

For overall scale, arXiv’s 2023 annual report says the service averaged 17,300 submissions per month, while a January 2026 arXiv paper reports that Q1 2025 alone had over 65,000 submissions and that arXiv had over 2.8 million submissions as of August 2025. Extrapolating from the more recent Q1 2025 rate gives roughly 49,000 submissions in a 69-day window; allowing modest growth into mid-2026 pushes my working denominator to about 50,000-52,000 submissions. (info.arxiv.org)

The hard part is the hit rate. Recent category snapshots on June 3, 2026 suggest relevant papers are real but sparse. The cs.AI recent page covered 1,812 entries across the five most recent announcement days, yet searching the displayed page text found no title matches for “pandemic,” “infectious,” “pathogen,” or “synthetic”; that points to a very low rate inside the large AI-only stream. By contrast, q-bio.QM had 29 recent entries and already included a direct title-level match, “Using Machine Learning to Enhance Hyperparameter Optimization in Pandemic Modeling”; q-bio.PE had 24 recent entries and included “The Epi-LLM Framework…” in an epidemiology context; q-bio.BM had 13 recent entries and no displayed-title matches for “pathogen,” “infectious,” “pandemic,” or “synthetic.” My read is that most qualifying papers will come from infectious-disease / epidemiology niches plus a smaller trickle from general AI categories, not from the much larger protein/genomics/biomolecule AI literature. (arxiv.org)

Independent 2026 arXiv examples confirm that the target phenomenon is active enough to produce multiple hits in a typical month: “Prospective multi-pathogen disease forecasting using autonomous LLM-guided tree search” (May 15, 2026), “Transfer Learning using 66 Diseases for Disease Forecasting Applications” (May 26, 2026), “A Machine Learning Framework for Constructing Heterogeneous Contact Networks: Implications for Epidemic Modelling” (March 14, 2026), and “Coordinated Pandemic Control with Large Language Model Agents as Policymaking Assistants” (January 14, 2026) all sit close to the queried intersection. (arxiv.org)

Combining those pieces, I use a central estimate near 39: about 51,000 total submissions times an effective match rate around 0.00075. I then widen uncertainty for phrase-matching brittleness, possible ambiguity in how “submission date” is implemented in arXiv advanced search, cross-listing effects, and moderation/timing noise near August 11, 2026. That produces a distribution centered in the high 30s / low 40s, with most probability between roughly 25 and 55, a meaningful but not dominant chance of finishing below 30, and only a very small tail above 100. (arxiv.org)

Resolution depends on the arXiv Advanced Search interface using the Abstract field, exact-phrase matching for quoted terms, wildcard support, case-insensitive matching, and an inclusive date range. arXiv’s help page also notes that the interface distinguishes among submission-date variants, which matters because the final resolution is the displayed search count rather than a semantic hand-labeling exercise. (arxiv.org)

For the top-down base rate, arXiv reported an average of 17,300 submissions per month in 2023. The window from June 4, 2026 through August 11, 2026 is 69 days, which would imply roughly 39,000 submissions at that 2023 pace; allowing for some growth by mid-2026, I use a rough total-window volume in the low-40-thousands as the denominator for a share-based estimate. (info.arxiv.org)

For the numerator, the query is broader than a narrow ‘biosecurity policy’ search. Recent arXiv examples show several live substreams that would plausibly match at least one AI term and at least one bio term in abstracts: pandemic forecasting/control with large language models, deep-learning infectious-disease forecasting, AI-synthetic-biology governance/risk work, and explicitly biosecurity-oriented LLM papers. Examples include PandemicLLM in 2024, Coordinated Pandemic Control with Large Language Model Agents in 2026, Physics-informed deep learning for infectious disease forecasting in 2025, The Convergence of AI and Synthetic Biology in 2024, A Biosecurity Agent for Lifecycle LLM Biosecurity Alignment in 2025, and Resilient Biosecurity in the Era of AI-Enabled Bioweapons in 2025. (arxiv.org)

Using that evidence, I infer a hit rate around 0.10% to 0.15% of all arXiv submissions in the window. Applied to a low-40-thousand total, that gives roughly 40 to 68 matches. I then lean slightly upward within that band because the bio list is broad (‘outbreak‘, ‘pandemic‘, ‘pathogen‘, ‘infectious disease‘, ‘synthetic biology’) and the AI list is also broad (‘machine learning’, ‘deep learning’, ‘transformer‘, ‘artificial intelligence’, ‘large language model‘). But I do not push the estimate too high, because resolution requires exact listed phrases in abstracts, so papers using only abbreviations like ‘ML’ or different phrasings outside the provided list will miss. That leaves me with a base case in the mid-50s. (arxiv.org)

Main uncertainty comes from wording sensitivity, the exact behavior of arXiv’s submission-date filter, and the fact that I was not able to directly backtest this exact Boolean/date query on historical windows. The public arXiv API does not document an equivalent date filter, and older arXiv API discussions describe date querying as undocumented, so I could not cleanly reconstruct past June 4-August 11 windows with the same mechanics. Because of that limitation, I widened the tails versus a pure Poisson model and used a negative-binomial distribution with mean 56 and standard deviation 19. (stackoverflow.com)

Net forecast: I center on 56 papers. That is comfortably above 20 because the query captures mainstream epidemic/pathogen ML work in addition to explicit biosecurity/synthetic-biology papers, but still below 100 because exact abstract phrasing is restrictive and this remains a small niche relative to total arXiv volume. (info.arxiv.org)

I treat the forecast origin as June 4, 2026. The target window from June 4 through August 11 is 70 days long. arXiv advanced search supports exact phrases in double quotes and ‘*’ wildcards, and its date-range filter uses inclusive lower and upper bounds. The interface also exposes both submission-date-most-recent and submission-date-original modes, so I allow a little extra uncertainty for possible resolution-time interpretation around revisions. (arxiv.org)

For the denominator, arXiv reported 208,493 new submissions in 2023, about 17,000 per month, with a record month of 20,710 in October 2023. A 70-day window at that 2023 run rate is about 39,093 papers; if 2026 volume is only modestly above that level, the June 4-August 11, 2026 window plausibly contains roughly 43,000-47,000 papers in total. (info.arxiv.org)

The conjunction in this market is broader than the title alone suggests. On the AI side, the query includes very common abstract phrases such as machine learning, deep learning, neural network, transformer, artificial intelligence, and LLM. On the bio side, it is not limited to explicit biosecurity language; it also includes outbreak, pandemic, epidemic, pathogen, and infectious disease roots. That matters because AI use inside biology is already substantial: a recent bibliometric study finds biology among the disciplines with large AI impact growth, and its biology-subfield overlap includes machine learning and artificial intelligence. Separately, arXiv-hosted surveys from 2023 and 2024 show infectious-disease prediction and prevention with AI/ML is already a mature topic rather than a one-off niche. (jianxgao.com)

Recent 2026 arXiv examples reinforce that the pipeline is still active. I found papers on LLM agents for pandemic control submitted on January 14, 2026; machine learning for epidemic modelling on March 14, 2026; machine learning for hospital outbreak detection on March 20, 2026; and deep learning for epidemic forecasting aimed at future pandemics on March 25, 2026. There is also explicitly biosecurity-flavored AI work on arXiv, such as Securing Dual-Use Pathogen Data of Concern, submitted on February 8, 2026. These examples do not give a complete count, but they do support a nontrivial continuing flow of matches. (arxiv.org)

My quantitative translation is therefore: if the window contains about 43k-47k papers, then a matching prevalence around 0.08%-0.12% implies about 34-52 hits, and 0.10% of that denominator is about 43-47 hits. That prevalence range is my inference from the broad term set, the already-established infectious-disease/AI literature, and the continuing 2026 examples. I center slightly above the middle of that range because current work suggests expansion is still ongoing, but I keep meaningful dispersion because exact-phrase matching, wildcard behavior, and date-filter interpretation can all move the displayed arXiv count around the central estimate. My point forecast is 46. The code encodes this as a four-scenario mixture of negative binomials with means 26, 40, 54, and 72 and weights 0.15, 0.45, 0.28, and 0.12 respectively, which keeps most probability mass in the high-20s through high-60s while preserving small tails below 20 and above 100. (arxiv.org)

I start from the resolution mechanics. arXiv advanced search supports quoted phrases in abstract search, wildcard matching with *, date-range filters that use the lower bound of the from date and the upper bound of the to date, and filtering by submission date rather than announcement date. That matters because this query is broader than the title suggests: the bio term set includes not just explicit words like biosecurity or biosafety, but also outbreak, pandemic, epidemic, pathogen, and infectious disease roots, so many ordinary infectious-disease AI papers can qualify even if they are not really about biosecurity policy. (arxiv.org)

On platform scale, arXiv reported 185,692 new submissions in 2022 and 208,493 in 2023, with roughly 17,000 submissions per month on average in 2023 and a record month of 20,710 in October 2023. If that 2022-to-2023 growth roughly persisted, a 69-day mid-2026 window would plausibly contain about 49,441 total submissions before topic filtering; I haircut that slightly for summer seasonality and think in terms of roughly 47k-49k submissions. The seasonality haircut is my inference, not an observed 2026 count. (info.arxiv.org)

For the topic base rate, one systematic review of infectious-disease prediction using ML/DL found 237 included studies through May 2021, while a 2024 bibliometric study of AI in pathogenic microbiology analyzed 27,420 publications spanning 2010-2024 after screening. Those are broader-literature measures rather than arXiv counts, but together they imply a field that is active and growing while still small relative to all scientific output. (medrxiv.org)

Recent arXiv papers also confirm ongoing flow in exactly the kinds of overlaps the query will catch: Physics-informed deep learning for infectious disease forecasting was submitted on January 16, 2025; Coordinated Pandemic Control with Large Language Model Agents as Policymaking Assistants on January 14, 2026; Cross-Country Learning for National Infectious Disease Forecasting Using European Data on January 28, 2026; and Transfer Learning using 66 Diseases for Disease Forecasting Applications on May 26, 2026. The mix is mostly forecasting, surveillance, and modeling rather than literal biosecurity wording, so I expect the final count to be driven mainly by the broad outbreak/pandemic/pathogen/infectious-disease terms rather than by the rarer explicit biodefense or biosafety terms. (arxiv.org)

My numerical model is a prevalence model. Start with roughly 48,000 submissions in the window and apply a match rate near 0.09%, or just under one match per thousand submissions. That yields about 43-44 papers. I then nudge slightly upward because arXiv is more AI-heavy than the literature as a whole and because the query includes several very common AI phrases such as machine learning, deep learning, neural network, and transformer roots; against that, I discount for post-COVID normalization and for the fact that explicit biosecurity language is still uncommon. Blending those considerations gives me a mean near 44 and a median forecast of 44. This paragraph is an inference from the cited evidence, not a directly measured arXiv count. (arxiv.org)

So my best forecast is 44 papers. I think finishing below 20 is possible but unlikely, while finishing above 100 would require either unexpectedly permissive query behavior on arXiv or a much stronger summer spike in relevant papers than the current evidence suggests. I therefore use an overdispersed negative-binomial distribution centered in the mid-40s. (arxiv.org)

How many arXiv papers submitted between June 4 and August 11, 2026 will contain both an AI term and a biosecurity term in their abstract?

• My most likely prediction for the number of arXiv papers is 56.50, with a 50% chance of falling between 44.50 and 67.50.
• The forecast is driven by the total volume of arXiv submissions, which is expected to reach approximately 55,000 to 65,000 papers during the 69-day window from June 4 to August 11, 2026.
• The search query is broad, including common terms like ‘machine learning’ and ‘infectious disease*’, which captures a wide array of epidemiological and pathogen-related research beyond narrow biosecurity topics.
• Rapid advancements in AI for life sciences, such as OpenAI’s GPT-Rosalind (launched April 2026), have significantly increased research activity at this intersection.
• Stricter arXiv moderation policies against AI-generated content and the ‘Abstract-only’ search filter act as moderating factors, preventing an unchecked surge in the final count.

Key figures

Historical context

• arXiv reached a cumulative total of 2 million submissions by the end of 2021, with 208,493 new submissions in 2023.
• Monthly submission rates have climbed from ~17,000 in early 2023 to ~24,000 by November 2024.
• The intersection of AI and biology has been historically driven by tools like AlphaFold (2020) and AlphaFold 3 (2024).
• In late 2025, arXiv was forced to temporarily shut down its computer science review section due to an ‘overwhelming volume’ of AI-assisted submissions.
• By August 2025, hallucinated citation rates in arXiv papers were estimated at 0.4%, prompting the stricter 2026 vetting policies.

Tailwinds

• The launch of ‘Rosalind Biodefense’ and GPT-Rosalind in early 2026 provides a technological platform that encourages new research at the AI-Bio intersection.
• High-profile global events like the ‘Seoul Pandemic X Science Summit’ (June 2026) likely spurred a wave of related submissions.
• The inclusion of broad terms like ‘pandemic*’ and ‘machine learning’ in the Boolean query ensures that standard epidemiological modeling papers are counted.
• Increased funding for biosecurity (e.g., the US National Security Commission on Emerging Biotechnology recommendations) has likely expanded the volume of active researchers in this field.

Headwinds

• Strict arXiv moderation policies implemented in May 2026 may disqualify ‘low-effort’ or ‘obviously AI-generated’ papers that would have otherwise triggered the search terms.
• The search is restricted to the ‘Abstract’ field only, meaning papers that discuss these topics in the main body but use more generic abstracts will be excluded.
• arXiv’s transition to an independent nonprofit in July 2026 could lead to temporary administrative lags or changes in indexing that affect search result visibility.
• Academic seasonality often sees a slight dip in submission volumes during the peak summer months of July and August.

Detailed reasoning

My analysis is based on a synthesis of arXiv’s historical growth, the specific keywords defined in the search query, and the rapidly evolving landscape of AI-biology research as of June 2024.

First, I established the base volume of arXiv submissions. Data from 2023 and 2024 show an upward trajectory, with monthly submissions reaching approximately 24,000 by late 2024. Projecting this to mid-2026, I expect approximately 25,000 to 28,000 submissions per month. The 69-day window (June 4 to August 11) comprises roughly 2.3 months, leading to a total expected submission volume of 57,500 to 64,400 papers across all disciplines.

Next, I evaluated the ‘hit rate’ for the Boolean search query. The search requires at least one term from the Bio Group and one from the AI Group within the abstract. The AI Group contains highly common terms (e.g., ‘machine learning’, ‘deep learning’), meaning the constraint primarily rests on the Bio Group. While terms like ‘bioweapon’ are rare, the inclusion of ‘pandemic‘, ‘pathogen‘, and ‘infectious disease*’ significantly broadens the net to include general epidemiological modeling and pathogen genomics. Historical estimates suggest that Quantitative Biology (q-bio) and relevant Computer Science (cs) categories account for a small but growing fraction of arXiv. I estimate an intersection rate of roughly 0.08% to 0.12% of total submissions.

I also accounted for the specific ‘AIxBio’ zeitgeist of 2026. News reports highlight major milestones such as OpenAI’s GPT-Rosalind and the Seoul Pandemic X Science Summit, which likely catalyzed research submissions in the months leading up to the target window. This ‘tailwind’ effect suggests a higher density of relevant papers than would be found in a 2024 or 2025 baseline.

The final forecast of 56.50 (with a 50% HDI of 44.50 to 67.50) was derived by weighing these volume projections against potential ‘headwinds.’ Notably, arXiv implemented stricter moderation for AI-generated text in early 2026. This policy targets “low-effort” submissions that might otherwise have inflated the count. Furthermore, the search is limited to the ‘Abstract’ field and requires exact phrase matches, which naturally filters out papers where these topics are mentioned only in the full text or via different terminology. The result is a balanced estimate that captures the intersection of academic rigor and technological expansion.

Key uncertainties

• The exact behavior of the arXiv search engine’s wildcard () for terms like ‘bioweapon‘ or ‘neural network*’ may capture unexpected variations or exclude relevant ones.
• The number of papers ‘in the moderation queue’ on August 11, 2026, could significantly lower the count if there is a processing backlog.
• Variations in how authors use keywords in abstracts—such as using ‘ML’ instead of the full ‘machine learning’—could impact recall.
• The effectiveness of the May 2026 ban on AI-generated content might be higher than expected, leading to a sharper decline in total submissions.

Conclusion

• The predicted count of 56.50 papers represents a significant presence of AI-biosecurity research on arXiv, reflecting the “dual-use” nature of modern AI in the life sciences.
• Decision-makers should monitor the August 11, 2026, deadline, as the final count will serve as a barometer for the speed of AI-integrated biological research and the potential for emerging biothreats.

Question: How many arXiv papers will be returned by an advanced-search on Aug 11, 2026 for submissions Jun 4–Aug 11, 2026 whose abstracts contain at least one Bio Group term and at least one AI Group term. Forecasters’ main anchor is a manually reconstructed same-window historical series (2020–2025: 8, 12, 18, 27, 38, 52) showing clear, steep year‑over‑year growth; multiple teams therefore center the 2026 count in the mid‑60s (rough central estimate ~65) with wide uncertainty and a meaningful right tail if recent percentage growth continues. All teams agree the query is broad (common terms like “machine learning,” “pandemic,” “pathogen,” “LLM”) so many matches will be general public‑health/AI papers rather than narrowly defined “biosecurity” work. Recent 2026 examples across cs.* and q‑bio and policy/news attention (Nature coverage, OpenAI product announcements, legislative activity) support continued upward momentum but are unlikely to massively inflate counts before Aug 11. Forecasters caution the resolution may undercount a few submissions because arXiv moderation/visibility lags near the end date and because exact query/parser behavior can change results by a handful. In sum: mid‑60s central forecast, wide overdispersion, moderate downside from end‑window visibility/query uncertainty, and nontrivial upside if vocabulary adoption accelerates.

Forecast and Reasoning

Forecast: Probability distribution:

• 1.00% chance of value below 24.0
• 5.00% chance of value below 32.0
• 10.00% chance of value below 40.0
• 20.00% chance of value below 52.0
• 30.00% chance of value below 60.0
• 40.00% chance of value below 68.0
• 50.00% chance of value below 76.0
• 60.00% chance of value below 84.0
• 70.00% chance of value below 93.0
• 80.00% chance of value below 104.0
• 90.00% chance of value below 114.0
• 95.00% chance of value below 118.0
• 99.00% chance of value below 120.25

Forecast Reasoning

We are forecasting the arXiv advanced-search result count for papers submitted June 4–August 11, 2026 (69 days) whose abstract contains (at least one Bio Group term) AND (at least one AI Group term), with exact-phrase matching and wildcards.

Key considerations from the analysts:

• The query is broader than “biosecurity” narrowly construed because Bio Group includes very general terms like “pandemic“, “epidemic“, and “outbreak*”, which can also appear in network-science / diffusion contexts (e.g., epidemic spreading models) that often use ML terminology. This pushes counts upward versus a strict “biosecurity policy” interpretation.
• Countervailing: the search is abstract-only with exact phrases, which is a strong filter; many relevant papers won’t include these exact phrases in the abstract (e.g., mentioning COVID-19 without “pandemic”, or using “AI” without any of the listed AI phrases).
• Most forecasters (2, 3, 4) cluster in the ~65–90 median range, with forecaster 4 explicitly placing substantial probability above the open upper bound (100.5). Forecaster 1’s reasoning appears to over-extrapolate from total arXiv volume and implicitly assumes much higher daily match rates than seems consistent with the others’ base-rate thinking; I treat it as a high-side scenario rather than the central case.

Synthesis: I center the distribution modestly above forecasters 2/3 (who cluster around mid-60s) because of the breadth of Bio terms (especially epidemic/outbreak language in modeling papers) while still keeping a meaningful chance that exact-phrase constraints keep the total below ~60. I assign roughly 20–30% probability that the final count exceeds 100.5 (open-above), reflected by p80 landing above the bound.

An amateur forecast would likely either (a) anchor to the 19.5–100.5 displayed range and avoid putting any mass above it, or (b) overcount by treating the Bio+AI term lists as implying a huge fraction of arXiv papers match. My forecast explicitly accounts for (i) the abstract-only exact-phrase filter (downward pressure) and (ii) the breadth/ambiguity of “epidemic/outbreak” language in non-bio modeling papers (upward pressure), yielding a wider and better-calibrated distribution with an explicit open-above tail.

Forecast rationale (numeric):

— Iteration 1 — Across the forecasts, the reasoning is driven by a keyword-intersection model: count arXiv abstracts that mention an AI term (e.g., machine learning, deep learning, neural networks, transformers, LLMs) and a biosecurity-related term (e.g., pathogen, epidemic, pandemic, outbreak, infectious disease, biosurveillance, synthetic biology, dual-use).

Main factors used

• Expected topical overlap: The biggest source of papers is assumed to be the active ML-for-biomedicine / AI-for-public-health area, where AI and bio terms naturally co-occur.
• Query breadth matters: Broad wildcards or expansive term lists can capture many relevant papers, but also introduce false positives from general biology, epidemiology, or AI safety papers.
• Co-occurrence is not random: The forecasts assume positive correlation between AI and biosecurity language, so the intersection should be meaningfully above a pure-independence baseline.
• Window effects and seasonality: The summer submission window is treated as a moderate-volume period, which tempers the estimate somewhat.
• Uncertainty in arXiv search behavior: A major source of uncertainty is how arXiv matching handles phrases, wildcards, and abstract wording, which affects both the center and the tails.

Areas of consensus

• The expected count is well below total arXiv volume and concentrated in a low-to-mid double-digit range.
• Most reasoning converges on a central estimate around the 50–80 paper range.
• All forecasts give wide uncertainty bounds, especially on the upside, due to the possibility of broader-than-expected matching or a surge in relevant research.

Main disagreements / variation

• Some reasoning is more conservative, expecting roughly 50–60 papers.
• Others are somewhat higher, around 75–80 papers, reflecting broader assumptions about matching and field overlap.
• Tail estimates vary mainly because of differing assumptions about how inclusive the search terms are and how many non-obvious matches they might pull in.

Overall synthesis

The shared view is that the count should be driven by a moderate but real intersection of AI and biosecurity research on arXiv, with the estimate hinging less on raw submission volume than on how broadly the abstract terms are interpreted. The most likely outcome is around several dozen papers, roughly in the 50–80 range, with substantial uncertainty around keyword matching and topical spillovers.

— Iteration 2 — Overall, the forecasts converge on a low-to-mid 60s to low 70s estimate for the number of arXiv papers in the June 4–August 11, 2026 window that mention both an AI term and a biosecurity-related term in the abstract.

Main reasoning patterns

• Base-rate approach: Each forecast starts from the expected total volume of arXiv submissions in the period and applies an estimated intersection rate for AI × biosecurity-relevant abstracts.
• Broad AI terms, narrower biosecurity terms: The AI side is treated as relatively expansive (e.g., machine learning, neural networks, transformers), while the biosecurity side is narrower but still broadened by wildcard/search variants (e.g., pathogen, epidemic, outbreak*).
• Co-occurrence is real but limited: The expected overlap is described as substantial enough to produce around about one matching paper per day, but not so common that the count would be very large.

Areas of consensus

• The likely count is not tiny and not in the hundreds; the most plausible mass sits around 60–70 papers.
• There is broad agreement that the estimate should reflect modest growth in AI-biology / AI-safety / epidemiology-adjacent work by 2026.
• All forecasts keep most probability mass below ~100 papers, while allowing some chance of a higher-than-expected burst.

Main sources of uncertainty

• Search syntax and wildcard behavior: The exact arXiv query could materially change the hit count, especially with wildcarded biosecurity terms.
• Submission timing and moderation delays: Papers submitted near the end of the window may not be fully captured, so one forecast adjusts for an effective shorter window.
• Topic bursts / field growth: A surge in AI-for-biology, epidemic modeling, or biosecurity-focused work could push counts upward.

Synthesis

The collective view is that this is a moderate-frequency intersection event driven by broad AI terminology and a somewhat narrower biosecurity vocabulary. The central estimate lands in the mid-60s to low-70s, with uncertainty mostly coming from how arXiv search terms map to real abstracts rather than from disagreement about the overall order of magnitude.

— Iteration 3 — Overall, the forecasts converge on a mid-range count in the low-to-mid 60s, with a rough center around 60–66 papers over the June 4–August 11, 2026 window.

Main reasoning pattern

• The estimate is driven by a broad keyword-overlap approach: papers count if their abstracts contain both an AI term and a biosecurity-related term.
• The AI side is assumed to capture common phrases like “machine learning,” “deep learning,” “transformer,” and “neural network.”
• The biosecurity side is likewise broad, including terms such as “pathogen,” “epidemic,” “outbreak,” “pandemic,” and “infectious disease,” with some models also noting “synthetic biology.”

Core factors behind the estimate

• High expected base rate: The combination of growing AI-for-biology / AI-for-health work and broad wildcard-friendly terms suggests dozens of literal abstract matches are plausible.
• Short but nontrivial time window: The 69-day period is long enough to produce a meaningful count, but still limited enough that the total should remain in the tens rather than hundreds in the central case.
• arXiv-specific constraints: arXiv is viewed as smaller than biomedical repositories for this topic, which keeps the estimate below very large counts even with broad terms.

Uncertainty and edge effects

• Keyword/search behavior matters a lot. Counts could shift materially depending on wildcard handling, phrase matching, and whether terms are interpreted narrowly or broadly.
• Moderation and policy effects may slightly suppress visible counts, especially if some June–August 2026 submissions remain in queue or are affected by recent arXiv moderation tightening.
• The distribution is therefore wide-tailed:
  • Lower tail: possible if terminology is narrower than expected or search matching is restrictive.
  • Upper tail: possible if broad terms like pathogen or epidemic match many more AI papers than anticipated, or if there is an AI-biosecurity surge.

Consensus vs. disagreement

• Consensus: the most likely outcome is around 60–65 papers.
• Disagreement: mainly concerns the tail risk, especially whether the count could plausibly rise well above 100 or fall materially below 40.

Here is a synthesis of the key qualitative arguments and shared reasoning across all 10 forecasts:

Baseline Volume and Conference Seasonality

The rationales uniformly highlight that the June 4 to August 11 window corresponds with a period of massive submission volume on arXiv. Artificial Intelligence and Computer Science papers represent a dominant share of the platform’s overall growth. More importantly, this specific summer window encompasses several premier AI conference deadlines (most notably AAAI, alongside KDD, EMNLP, and others). These deadlines traditionally trigger intense, sudden surges in preprint uploads as researchers rush to establish priority before double-blind review periods, artificially inflating the baseline submission rate for the period.

Terminology Dynamics: Genuine Growth vs. Spillover

Because the “AI Group” terms are nearly universal in modern computer science literature, the hit rate relies almost entirely on the presence of the “Bio Group” terms. Forecasters divided these terms into two behavioral categories:

• Genuine Intersection: There is exponential growth in research at the intersection of AI and biosecurity. Recent executive orders, high-profile grants (from DARPA, ARPA-H, etc.), LLM red-teaming initiatives, and ongoing global health events (like H5N1) are expected to yield a wave of newly published papers explicitly focused on “dual-use,” “synthetic biology,” and “biosecurity.”
• Spillover and False Positives: Many hits will come from broad epidemiological terms (pandemic*, epidemic*, outbreak*). These terms frequently appear in purely technical or non-biological papers, either as boilerplate motivational background (e.g., “motivated by the recent pandemic”) or in network science (e.g., “epidemic routing” or modeling “infodemics”).

Downward Pressures: Moderation and Submission Cutoffs

While conference spikes and AI safety funding drive the baseline up, structural constraints heavily depress the final count:

• The Cutoff Penalty: The exact resolution date (August 11) acts as a strict ceiling. Papers submitted in the final days of the window—especially during the massive late-July AAAI rush—experience standard platform announcement lags and may not be publicly visible in time to be counted.
• Moderation Queue Delays: ArXiv has recently implemented stricter moderation policies, cracking down on unreviewed position papers and LLM-generated “slop.” Furthermore, sensitive terms like “bioweapon” or “dual-use” frequently trigger manual reviews. Papers held in this moderation queue at the time of the August 11 cutoff will be excluded.
• The bioRxiv Drain: Purely medical or virological research will continue to bypass arXiv in favor of domain-specific preprint servers like bioRxiv and medRxiv.

Search Engine Mechanics

Finally, several rationales emphasized the technical quirks of arXiv’s search engine (Elasticsearch). The requirement to query exact phrases combined with trailing wildcards introduces parsing uncertainty. Depending on how the query is executed, aggressive lemmatization might inflate the count (e.g., expanding terms unintentionally), or wildcard literalization bugs (where the system searches for a literal asterisk character) could severely suppress the hit rate.

SUMMARY

Question: How many arXiv papers submitted between June 4 and August 11, 2026 will contain both an AI term and a biosecurity term in their abstract? Final Prediction: Probability distribution:

• 10.00% chance of value below 25.466667
• 20.00% chance of value below 35.466667
• 40.00% chance of value below 51.8
• 60.00% chance of value below 70.133333
• 80.00% chance of value below 99.466667
• 90.00% chance of value below 125.8

Total Cost: extra_metadata_in_explanation is disabled Time Spent: extra_metadata_in_explanation is disabled LLMs: extra_metadata_in_explanation is disabled Bot Name: extra_metadata_in_explanation is disabled

Report 1 Summary

Forecasts

Forecaster 1: Probability distribution:

• 10.00% chance of value below 26.5
• 20.00% chance of value below 36.5
• 40.00% chance of value below 52.5
• 60.00% chance of value below 69.5
• 80.00% chance of value below 95.5
• 90.00% chance of value below 117.5

Forecaster 2: Probability distribution:

• 10.00% chance of value below 25.5
• 20.00% chance of value below 36.5
• 40.00% chance of value below 53.5
• 60.00% chance of value below 73.5
• 80.00% chance of value below 108.5
• 90.00% chance of value below 139.5

Forecaster 3: Probability distribution:

• 10.00% chance of value below 24.4
• 20.00% chance of value below 33.4
• 40.00% chance of value below 49.4
• 60.00% chance of value below 67.4
• 80.00% chance of value below 94.4
• 90.00% chance of value below 120.4

Research Summary

The research forecasts how many arXiv abstracts submitted between June 4 and August 11, 2026 might contain both an AI term (e.g., “machine learning,” “LLM,” “transformer,” “GPT”) and a biosecurity term (e.g., “biosecurity,” “pathogen,” “pandemic,” “biodefense”). It frames the 69-day window as ~19% of a year and notes recent industry and policy events (notably OpenAI’s GPT‑Rosalind program launched May 29, 2026, new UK‑France and fellowship initiatives in early June, and heightened attention from papers such as a May 13, 2026 Nature piece) that could raise submissions. The briefing also accounts for arXiv platform changes (May 15, 2026 enforcement policy on AI-generated content), historical arXiv growth and daily AI submission rates, the established computational-epidemiology reference class, and seasonal/conference timing effects that push estimates up or down.

Using these qualitative drivers and base-rate reasoning (including bioRxiv dual-use screening results and arXiv AI submission counts), the research gives rough numerical ranges rather than a single point estimate: a conservative lower bound of ~20–50 papers, a moderate estimate of ~50–150, an optimistic upper bound of ~150–300, and a “wild card” scenario of 300+ if a substantial submission surge occurred. It highlights key uncertainties and limitations (no direct historical Boolean-query counts on arXiv, unclear policy impacts, seasonal variability, and search-method details affecting counts) and notes that only a small intersection percentage of AI papers would be needed to reach the moderate-to-optimistic ranges.

Sources mentioned in the briefing (no explicit URLs were provided in the text): OpenAI GPT‑Rosalind program announcement / OpenAI press materials; Nature (May 13, 2026) article on AI and biological risk; arXiv enforcement policy announcement (May 15, 2026); bioRxiv systematic analysis of dual‑use flags (2024–2025); Microsoft research (2025) on molecule redesign; UK‑France Strategic Alliance in Health and AI announcement; Frontier Biodefense Fellowship announcement; Forecasting Research Institute publications; arXiv submission statistics; Metaculus / Manifold / Polymarket (noting no direct prediction for this question). The briefing did not include verbatim links to these sources.

RESEARCH

Report 1 Research

Research Briefing: arXiv Papers with AI and Biosecurity Terms (June 4 - August 11, 2026)

Executive Summary

This forecast concerns a highly specific intersection of AI and biosecurity research during a 69-day summer period (approximately 19% of a year). The timing is notable: the forecast period begins just days after major AI-biosecurity initiatives launched, and during a period of heightened public and policy attention to this intersection.

Relevant News and Recent Developments

Major Industry and Policy Initiatives (May-June 2026)

OpenAI’s GPT-Rosalind Biodefense Program [2][7][10]

• Launched May 29, 2026—just 6 days before the forecast period begins
• Specifically designed for biosecurity, pandemic preparedness, and biodefense applications
• Partners include Lawrence Livermore National Laboratory, Johns Hopkins Applied Physics Laboratory, and CEPI
• Provides trusted access to specialized AI models for biological threat detection and response
• Investment of $45M total ($15M to Red Queen Bio, $30M to Valthos) [2]

Academic and Research Initiatives

• UK-France Strategic Alliance in Health and AI launched June 3, 2026, focusing on pandemic preparedness and pathogen detection using AI [6]
• Frontier Biodefense Fellowship launched June 2, 2026, specifically addressing AI-enabled biosecurity risks [8]
• Forecasting Research Institute published study June 3, 2026 on LLM-enabled biorisk [12]

AI Biosecurity Risk Landscape [1]

• Nature article (May 13, 2026) extensively documents AI’s capability to design pathogens, toxins, and bioweapons
• Studies show LLMs enable non-experts to match PhD-level virology skills [1]
• Microsoft research (2025) demonstrated AI could redesign biological molecules into 76,000 synthetic variants to evade screening [1]

arXiv Platform Changes

Enforcement Policy [4][5]

• May 15, 2026: arXiv announced strict penalties for unverified AI-generated content
• One-year submission bans for authors who submit AI-generated work without proper verification
• This may reduce low-quality submissions but also create hesitancy around AI-related research

arXiv Submission Statistics and Trends

Historical Baseline Data

Overall arXiv Growth [14][15]

• 2021: ~181,630 total submissions annually
• 2024: ~24,000 submissions per month (~288,000 annually if sustained)
• Computer Science in 2021: ~60,989 papers (33.6% of total)
• Biology (q-bio) in 2021: ~16,147 papers (8.9% of total)
• Growth rate has been 2-15% annually depending on field

Recent Activity Levels [11][13]

• June 3, 2026: 440 new AI submissions in a single day
• June 3, 2026: 1,812 recent submissions in AI category total
• Multiple papers explicitly focus on biosecurity and healthcare AI applications

AI-Generated Content Trends [9]

• arXiv computer science: AI-written materials grew from 7% (2023) to 43% (2025)
• Overall AI participation in academic writing increased 42% after ChatGPT’s November 2022 launch
• This trend may be dampened by new enforcement policies [4][5]

For the 69-Day Forecast Period

The forecast window (June 4 - August 11, 2026) represents approximately 19% of a year. Key temporal factors:

• Summer period: Historically lower submission rates due to vacations and conference travel
• Post-launch timing: Begins 6 days after GPT-Rosalind announcement, potentially spurring related research submissions
• Conference deadlines: May include pre-deadline submission spikes for fall 2026 AI/ML conferences

Base Rates and Reference Classes

Dual-Use Research Prevalence [18]

A systematic analysis of 52,713 bioRxiv preprints (2024-2025) found:

• 23.2% flagged for dual-use concerns when screening for biosecurity-related keywords
• 14.2% flagged specifically for Pathogens with Enhanced Pandemic Potential (PEPP)
• However, this used broader screening criteria than the specific Boolean query in question
• Note: This is bioRxiv (biology-focused), not arXiv

Estimating the Intersection

The question requires papers containing:

• At least one AI term from: LLM, large language model, generative AI, foundation model, AI agent, transformer, machine learning, deep learning, neural network, generative model, GPT, artificial intelligence
• At least one biosecurity term from: biosecurity, biosurveillance, pandemic preparedness, outbreak detection, pathogen detection, synthetic biology, dual-use, biodefense, biothreat, bioweapon, biological weapon, biosafety, outbreak, pandemic, epidemic, gain-of-function, DNA synthesis, gene synthesis, pathogen, infectious disease*

Key Considerations:

1. This is a highly specific intersection requiring exact phrase matches in abstracts
2. Historically niche, but rapidly growing field [1][8][12]
3. Both AI and biosecurity are broad term sets, increasing match probability
4. Common terms like “machine learning,” “pandemic,” and “pathogen“ are likely to appear together in computational epidemiology and bioinformatics papers

Reference Class: Computational Epidemiology

Research teams working on computational epidemiology regularly combine machine learning with pandemic/pathogen analysis [19]. This established field provides a natural source of papers matching both criteria, including:

• Genomic surveillance using deep learning
• Epidemic forecasting with neural networks
• Pathogen detection using AI
• Outbreak prediction models

Factors Influencing the Forecast

Upward Pressures (Increasing Paper Count)

1. Recent major announcements creating research momentum and funding [2][7][10]
2. Growing policy attention to AI biosecurity risks driving academic interest [1][8][12]
3. Established research communities in computational epidemiology, bioinformatics, and AI safety [19]
4. Broad term sets with common words (e.g., “machine learning” + “pandemic”) likely to match many abstracts
5. Multiple relevant arXiv categories: cs.LG, cs.AI, q-bio, stat.ML could all contain matching papers
6. Cross-listing means papers appear in multiple categories but count once

Downward Pressures (Decreasing Paper Count)

1. Summer timing typically sees reduced academic activity
2. Very specific requirement: Both term sets must appear in the same abstract
3. Short time window: Only 69 days vs. full year
4. arXiv enforcement concerns may reduce AI-related submissions [4][5]
5. Niche intersection: Despite growth, absolute numbers remain small relative to total submissions

Prediction Markets and Expert Forecasts

No specific prediction markets found for this exact question on Manifold, Metaculus, or Polymarket.

Related Forecasting Activity [12][17]:

• Forecasting Research Institute actively studying LLM-enabled biorisk with expert panels
• Metaculus and prediction markets exist for biosecurity events but not arXiv paper counts
• Note: Metaculus typically provides more calibrated low-probability estimates than prediction markets [17]

Historical Frequency Estimates (Rough Approximations)

Without access to historical arXiv search data for this specific query, I can provide rough estimates based on available information:

Conservative Lower Bound: 20-50 papers

• Assumes niche intersection with limited historical precedent
• Summer period suppresses submissions
• Based on small percentage of total CS + bio submissions

Moderate Estimate: 50-150 papers

• Accounts for established computational epidemiology field
• Recent growth in AI biosecurity research
• Broad term matching (especially “machine learning” + “pandemic”/“pathogen”)

Optimistic Upper Bound: 150-300 papers

• Reflects momentum from recent announcements and funding
• Multiple arXiv categories contributing papers
• Broader terms like “artificial intelligence” and “infectious disease*” capturing many papers

Wild Card High Estimate: 300+ papers

• Would require substantial submission surge from GPT-Rosalind launch effect
• Possible if major research groups submit multiple related papers
• Conference deadline clustering could create spikes

Research Gaps and Limitations

1. No direct historical data on this exact Boolean query against arXiv abstracts
2. Seasonal variation in submissions not well-documented in available sources
3. Unclear impact of recent policy changes on submission behavior [4][5]
4. Publication lag from research to arXiv submission varies by field
5. Search methodology details (exact phrase matching, wildcard behavior) affect counts

Quantitative Benchmarks for Context

• Daily arXiv AI submissions (June 2026): ~440 papers [11]
• 69-day extrapolation: ~30,000 AI papers total if rate sustained
• Intersection percentage needed: Would need only 0.17% to 1% of AI papers to also contain biosecurity terms to reach 50-300 range
• bioRxiv dual-use rate: 23.2% of biology preprints had biosecurity-related keywords [18], though this uses different criteria

This intersection appears plausible given that computational biology, bioinformatics, and health AI are established fields where both term sets naturally co-occur.

FORECASTS

R1: Forecaster 1 Reasoning

I think the critic’s main correction is valid: my previous median was a bit too high, mostly because I gave too much weight to AI-biosecurity salience and not enough to the literal mechanics of the resolving query. I’m updating downward, though not all the way to a strongly low view.

Key updates:

• Retrospective exact-query data would be the gold standard. I agree this is the largest missing input. This is a count question, not a semantic judgment question, and a few prior June–August windows would be highly informative.
• Near-term news should get little weight. GPT-Rosalind / biodefense announcements may increase attention, but papers appearing by August 11, 2026 were mostly already in the pipeline. The effect is likely a few papers, not dozens.
• Literal query mechanics matter a lot. arXiv advanced search supports abstract-field searching, exact quoted phrases, wildcards, and submitted-date filtering; its search code also supports Boolean and parenthetical search, so the resolution should behave like a fairly literal Boolean metadata count. (arxiv.org)
• But I still think <20 is unlikely. The bio term list is much broader than “biosecurity”: pandemic*, epidemic*, outbreak*, pathogen*, and infectious disease* will capture computational epidemiology, hospital outbreak detection, pathogen surveillance, COVID/social-science remnants, and some synthetic biology / AI-for-biology work. For example, recent arXiv abstracts already include highly query-relevant phrases like machine-learning hospital outbreak detection and pathogen/epidemic framing. (arxiv.org)
• Dominant source of matches is probably boring ML-for-epidemiology, not frontier biosecurity. The critic is right that this is mostly a “machine learning/deep learning/neural network” plus “epidemic/pathogen/pandemic/outbreak” count.
• Final-window visibility trims the count modestly. Papers submitted near August 11 may not all be publicly listed by the time the resolving search is run. I treat this as a small downward adjustment, not a major regime change.

Net: I move my median from 65.5 to 60.5, thicken the low tail, and keep a fairly fat high tail because the query is broad and can scoop up non-obvious “pandemic/epidemic/pathogen” abstracts.

Updated prediction:
Percentile 0.1: 3.5
Percentile 1: 8.5
Percentile 5: 18.5
Percentile 10: 26.5
Percentile 20: 36.5
Percentile 30: 45.5
Percentile 40: 52.5
Percentile 50: 60.5
Percentile 60: 69.5
Percentile 70: 81.5
Percentile 80: 95.5
Percentile 90: 117.5
Percentile 95: 142.5
Percentile 99: 210.5
Percentile 99.9: 330.5

R1: Forecaster 2 Reasoning

Conditioning on normal resolution / not annulled.

I accept the critic’s main correction: title/category evidence is weaker than I treated it in Round 3. The resolution is abstract-only, and arXiv search mechanics matter: exact phrases, wildcards, and the choice between original vs. most-recent submission date are real load-bearing details. arXiv’s advanced-search page explicitly notes phrase search, wildcard search, and multiple date-filter modes, so this is not just a semantic “AI + biosecurity” question. (arxiv.org)

The current q-bio sample is especially clarifying. It shows a real relevant trickle—e.g. titles like “Using Machine Learning… Pandemic Modeling” and “The Epi-LLM Framework”—but it also shows many q-bio/AI cross-lists that are protein, neuroscience, drug-design, or biomolecular papers likely to miss the Bio Group in the abstract. The page had 115 recent q-bio entries and 17 on June 3, but many visible entries are not plausible abstract hits. (arxiv.org) This makes me lower the median from 68.5.

I still do not want to move too low. The query has broad upward quirks: “machine learning,” “deep learning,” “neural network,” “transformer,” “LLM,” and “artificial intelligence” are common abstract phrases, while “pandemic,” “epidemic,” “outbreak,” “pathogen,” “dual-use / dual use,” and “synthetic biology” can catch a lot beyond strict biodefense. Current cs.AI and cs.LG flows are large—recent pages showed 1,812 cs.AI entries and 1,594 cs.LG entries, with hundreds on June 3—so very small intersection rates can still produce dozens of matches. (arxiv.org)

New adjustment: I now put more weight on abstract-only misses from title-level examples, plus the possibility that “COVID-19,” “SARS-CoV-2,” “epidemiological,” “public health,” “AI,” or “ML” appear without matching the listed strings. I also put slightly more weight on small downward effects from public-listing lag near August 11. But I keep a fat right tail because of broad false positives and date/filter mechanics.

My rough center is now: direct computational epidemiology/infectious-disease ML around 25–35; AI-for-bio/synbio/pathogen/biosafety around 15–25; false positives from dual-use, epidemic-spread, pandemic, social contagion, governance, etc. around 10–20. After overlap and uncertainty, that puts the median in the low 60s, not high 60s.

Updated prediction:
Percentile 0.1: 3.5
Percentile 1: 8.5
Percentile 5: 17.5
Percentile 10: 25.5
Percentile 20: 36.5
Percentile 30: 45.5
Percentile 40: 53.5
Percentile 50: 62.5
Percentile 60: 73.5
Percentile 70: 88.5
Percentile 80: 108.5
Percentile 90: 139.5
Percentile 95: 178.5
Percentile 99: 285.5
Percentile 99.9: 440.5

R1: Forecaster 3 Reasoning

I accept the critic’s core correction: this should be modeled less as an “AI biosecurity is hot” forecast and more as an arXiv exact-query/token-count forecast. The late-May/early-June announcements may matter a little, but most papers submitted by August 11 were already in the pipeline.

The biggest valid criticisms are:

• Direct historical query counts would dominate this forecast if available; without them, the per-day rate is a judgment call.
• Venue substitution matters: much infectious-disease / pathogen / public-health ML work will land on bioRxiv, medRxiv, journals, or reports rather than arXiv.
• Exact wording bites: many relevant abstracts say “AI,” “ML,” “GNN,” “CNN,” “viral,” “epidemiology,” etc., none of which necessarily match the specified lists.
• Search mechanics are first-order uncertainty: arXiv advanced search supports exact phrases and wildcards, and it also exposes distinct date bases such as most-recent submission date, original submission date, and announcement date. That makes the resolver’s implementation important, not just the underlying research trend. (arxiv.org)

What I still defend from my earlier reasoning:

• The query is not narrow biosecurity. pathogen*, pandemic*, epidemic*, outbreak*, "infectious disease*", and "synthetic biology" are broad enough to catch ordinary computational biology, epidemic modeling, pathogenicity prediction, and synthetic-biology ML papers.
• The upper tail should remain substantial because wildcard and date-field behavior can add many accidental matches.
• arXiv is large enough that even a thin intersection can produce dozens of papers over 69 days; arXiv reports over 3.06 million total submissions as of June 4, 2026. (arxiv.org)

My update: I’m moving the center down from 62.4 to 58.4, not all the way to the critic’s 55.4. I agree with the downward pressure from arXiv-only venue choice and exact terms, but I think the broad bio wildcards and date-mechanics ambiguity keep the median from falling too far. I’m also fattening the upper tail relative to my prior forecast.

Conditioning on the question not being annulled:

Updated prediction:
Percentile 0.1: 2.4
Percentile 1: 7.4
Percentile 5: 16.4
Percentile 10: 24.4
Percentile 20: 33.4
Percentile 30: 41.4
Percentile 40: 49.4
Percentile 50: 58.4
Percentile 60: 67.4
Percentile 70: 79.4
Percentile 80: 94.4
Percentile 90: 120.4
Percentile 95: 155.4
Percentile 99: 250.4
Percentile 99.9: 430.4

Forecast — Post 420

Modal: 50-60 (20.0%) • partial_shift • 9m6s

Interpretation

Reading: strict

The resolution criteria are highly specific and literal: exact Boolean query terms are enumerated, the search field (Abstract), the interface (arxiv.org/search), the date filter (June 4–August 11, 2026 inclusive), wildcard behavior, cross-listing deduplication, and moderation queue exclusions are all precisely defined, leaving little room for charitable interpretation.

Time bounds

• observation_window: Papers submitted (announced) on arXiv between June 4, 2026 and August 11, 2026 inclusive, publicly listed by August 11, 2026
• resolution_date: 2026-08-11

Edge cases identified

1. Wildcard matching scope: ‘bioweapon‘ should match bioweapons, bioweaponry, etc.; ‘outbreak‘ matches outbreaks; ‘pandemic*’ matches pandemics — the resolver must confirm the arXiv search interface supports suffix wildcards in abstract mode as described.
2. Submission date vs. announcement date: arXiv may distinguish between submission date and the date a paper is publicly announced/listed; papers submitted at the end of August 11 may appear the next business day — resolver must use publicly listed date, not submission timestamp.
3. Cross-listing: a paper submitted to both cs.AI and q-bio categories should be counted only once per the criteria.
4. Moderation queue exclusion: papers submitted within the window but still under moderation and not publicly listed by August 11, 2026 are excluded — this is particularly relevant for papers submitted near the deadline.
5. arXiv search result pagination: the ‘total results’ figure shown at the top of arXiv search results may differ slightly from the actual paginated count; the resolver should use the displayed total figure.
6. Case sensitivity: terms like ‘LLM’ and ‘GPT’ are uppercase acronyms — ‘case-insensitive’ matching should still capture these correctly, but resolver must verify arXiv’s handling.
7. The Bio Group includes very broad terms like ‘pandemic‘, ‘epidemic‘, ‘outbreak‘, and ‘pathogen‘ which could match a very large number of papers; combined with broad AI terms like ‘machine learning’ and ‘neural network*’, the count could be substantially higher than terms like ‘biosecurity’ alone would suggest.
8. Weekend/holiday submission batching: arXiv batches submissions from weekends/holidays and announces them on the next business day — the resolver must decide whether those announcement-date papers fall within the window.

Research (3/3 variants, shared evidence pool)

Total evidence registered (shared pool): ?

Research Brief

Evidence confidence: low

Scenario 1: Mid-range count (40–70 papers): moderate AI×biosecurity intersection in abstract-matching exact phrases over the ~69-day window [high evidence]

Conditions favoring

The query covers many broad Bio Group terms (pandemic, epidemic, pathogen, outbreak, infectious disease) that appear frequently in abstracts alongside common AI terms (machine learning, deep learning, neural network). With ~10 weeks of submissions and ~900 AI papers/week (source 6), even a small fraction (under 1%) hitting a biosecurity term in their abstract adds up. The active AIxBio research community (sources 2, 3) and documented rate of relevant preprints support a count in this range over the full window.

Conditions against

Abstract-specific exact-phrase matching is more restrictive than title or full-text search. Many papers use these terms only incidentally (e.g., ‘pandemic’ in a social context unrelated to biology). The answer range ceiling of ~100 constrains the upper scenario.

Scenario 2: Lower count (20–40 papers): narrower effective intersection due to abstract-specific exact-phrase matching constraints [medium evidence]

Conditions favoring

The resolution criteria require exact-phrase abstract matching (not full-text or title), which is significantly more restrictive. Papers may mention AI methods in methods sections but not abstract, or vice versa for biosecurity terms. The narrower Bio Group terms (biosecurity, biodefense, biosurveillance, gain-of-function) are less common than the broad ones; if the broader wildcards match but at lower rates in abstracts specifically, the count could be lower. The answer range floor of ~20 establishes this as a plausible bound.

Conditions against

The Bio Group includes very broad wildcard terms (pandemic, epidemic, pathogen, outbreak) that are extremely common in biomedical abstracts; even one such match alongside any AI term in the abstract would count. Sustained arXiv growth (source 1, source 6) and the active community (sources 2, 3) make very low counts less plausible over a 69-day window.

Scenario 3: Higher count (70–100 papers): broad term coverage drives higher-than-expected results [medium evidence]

Conditions favoring

If terms like ‘pandemic‘, ‘epidemic‘, ‘pathogen‘, ‘infectious disease‘, and ‘outbreak‘ match at high rates in abstracts (they are standard epidemiological/biomedical vocabulary), and AI terms like ‘machine learning’, ‘deep learning’, ‘neural network‘ appear in many biomedical ML papers’ abstracts, the intersection could approach the upper end of the answer range. arXiv’s nearly 28,000 monthly submissions (source 1) and 50% CS share (source 8) mean many papers are being submitted across public health, quantitative biology, and AI categories simultaneously. Source 7 also notes revisions count under ‘submitted_date’, potentially inflating counts.

Conditions against

The answer range has an upper bound near 100; achieving counts at the top end requires very high overlap rates. The broad biological terms (pandemic, pathogen) appear in biology/medicine papers that may not use AI terminology in their abstracts. The window is ~69 days, not a full year.

Key uncertainties

• What fraction of arXiv abstracts containing broad Bio Group wildcard terms (pandemic, epidemic, pathogen, outbreak) also contain at least one AI Group term in the same abstract — this drives the count more than any other variable
• Whether the arXiv ‘submitted_date’ filter counts initial submissions only or includes revisions (source 7 indicates the latter, which could meaningfully inflate the count)
• How strictly the arXiv GUI enforces exact-phrase abstract matching for wildcard terms, versus whether partial/stemmed matches are returned
• Whether the query can be faithfully constructed in the arXiv Advanced Search GUI given the lack of parenthetical grouping (source 11), or whether the resolver uses an API/URL workaround
• The growth rate of AI+biosecurity research in 2026 relative to 2025 baselines — if there is a step-change event (e.g., a pandemic outbreak or major AI biosecurity policy), submission rates could spike

Forecast variants (3/3)

Variant 0: inside_view (inside_view_v1)

Inside View: Causal Chain Analysis

Current state → mechanism → outcome:

The query covers very broad Bio Group terms (pandemic, epidemic, pathogen, outbreak, infectious disease) AND broad AI Group terms (machine learning, deep learning, neural network, transformer*, artificial intelligence). This is a 69-day window with ~900 AI papers/week on arXiv (source 6), across a period when the AIxBio research community is active (sources 2, 3).

Causal mechanism:

1. arXiv receives ~28,000 submissions/month (source 1), ~9,800 AI-related (source 8)
2. Over 69 days (~10 weeks), that’s ~22,500 AI papers
3. The Bio Group wildcard terms (pandemic, epidemic, pathogen, outbreak) are standard biomedical vocabulary — they appear in a non-trivial fraction of abstracts in q-bio, cs.LG (epidemiology ML), and related categories
4. Even 0.3-0.4% of AI papers mentioning one of these bio terms in their abstract = ~70-90 papers
5. But abstract-specific exact-phrase matching is more restrictive than full-text (source 9 confirms wildcard support; source 11 notes query complexity)

Scenario weighting:

• Mid-range (40-70) tagged high evidence strength: The active AIxBio community (sources 2, 3), regular cs.AI listings with epidemic/pandemic ML papers (source 5), and broad term coverage all point here. This is the high-evidence scenario and should receive the most mass.
• Lower count (20-40) tagged medium: Abstract-specific matching is restrictive, query construction complexity (source 11) may reduce effective coverage.
• Higher count (70-100+) tagged medium: Broad wildcard terms + revision inflation (source 7) could push counts higher.

Key uncertainties (brief flags confidence_in_evidence as “low”): The critical unknown is the empirical overlap rate between Bio Group abstract terms and AI Group terms. The brief doesn’t have a direct measurement, so I widen my distribution accordingly.

Distribution shape:

• The high-evidence mid-range scenario (40-70) should dominate
• I weight the lower end (20-40) and higher end (70-100+) as roughly equally plausible given medium evidence
• “Above 100” (bin_9) gets some mass because broad wildcards + revision counting could push beyond range

Given the wide uncertainty acknowledged by the brief (confidence_in_evidence: low), I spread mass moderately:

• below 20 (bin_0): ~2% - very unlikely given active community and broad terms
• 20-30 (bin_1): ~8% - possible if abstract matching is very restrictive
• 30-40 (bin_2): ~13% - lower end of medium-evidence scenario
• 40-50 (bin_3): ~17% - bottom of high-evidence scenario
• 50-60 (bin_4): ~18% - core of high-evidence scenario
• 60-70 (bin_5): ~15% - top of high-evidence scenario
• 70-80 (bin_6): ~10% - bottom of higher-count medium scenario
• 80-90 (bin_7): ~7% - middle of higher-count scenario
• 90-100 (bin_8): ~5% - top of higher-count scenario
• above 100 (bin_9): ~5% - revision inflation + very broad terms could exceed 100

Variant 1: outside_view (outside_view_v1)

Reference class: counts of arXiv papers in a comparable ~69-day window (June–Aug) that contain at least one Bio-group term AND at least one AI-group term in the abstract. Base-rate anchor and back-of-envelope: arXiv produced roughly 900 AI papers/week in 2025 (source 6), so ~9,000 AI papers over the 10-week window; if ~0.5% of those include a Bio-group term in the abstract, that implies ~45 papers as a simple outside-view central anchor. Adjustment: I keep the outside-view anchor (~45) but widen and shift slightly to reflect the brief’s evidence: (a) broad Bio-group wildcards (pandemic, pathogen, outbreak, infectious disease) and broad AI terms make overlaps plausible (supporting upward adjustment) — see key facts/source 9 (wildcard support) and evidence_summary noting many AI×biosecurity papers in listings (source 5); (b) abstract-only exact-phrase matching is restrictive and pushes downward (source: evidence_summary and key_facts about abstract-specific restriction); (c) submitted_date counting latest-version uploads (not v1) can inflate counts if many revisions occur (source 7); (d) overall evidence quality is judged low in the brief, so I spread probability into tails. Final: distribution concentrates most mass in the mid-range consistent with the base-rate anchor but leaves substantial tail probability to both lower and higher bins. Weighting vs. scenarios: the brief’s mid-range scenario (40–70) was tagged high evidence — I allocate the plurality mass there (bins 3–5 sum to 0.61). The lower (20–40) and higher (70–100) scenarios were medium evidence — I allocate the remaining mass across bins 1–2 (0.22) and 6–8 (0.17) respectively. Strict-reading note: I follow the interpretation’s strict reading of the resolution (Abstract-field, exact-phrase + wildcards, date filter, dedupe cross-lists, exclude non-publication/moderation) and anchor to its monotone_floor (papers already submitted since June 4) as described in the brief. Confidence and spread: because the brief flagged low confidence in the evidence, I intentionally kept a wider spread (not a tight point mass) to reflect uncertainty about overlap rate, arXiv query construction, and submitted_date semantics (sources 7, 9, 11 are particularly relevant). Specific brief citations used: source 6 (AI paper volume → base-rate anchor), source 5 and evidence_summary (examples of AI×biosecurity papers), source 9 (wildcard support), source 11 (Advanced Search grouping limitations), and source 7 (submitted_date counts latest-version uploads).

Variant 2: contrarian_v2 (contrarian_v2)

Adversarial reading

The brief’s implicit consensus is a moderate intersection count (centered 40-70) driven by high weekly AI paper volume (~900/week from source 6) plus frequent overlap with broad Bio Group wildcard terms like pandemic, pathogen, and outbreak* that appear in many biomedical abstracts. This scenario receives the only ‘high’ relative_evidence_strength tag and is supported by the documented AIxBio research activity (sources 2, 3) and arXiv scale (sources 1, 8).

Counter-evidence found (or absence)

The brief itself flags that abstract-specific exact-phrase matching is substantially more restrictive than full-text or title search, that many papers use the broad wildcard terms only incidentally, and that the supplied answer range caps near 100—yet the mid-range scenario down-weights these constraints. Additional counter-evidence includes the GUI’s lack of parenthetical grouping (source 11), the ‘submitted_date’ filter counting revisions rather than v1 submissions (source 7), and the explicit note that the observed range is narrow relative to raw arXiv output, all of which point to a lower realized count than the high-evidence scenario assumes.

Distribution implications

The adversarial reading therefore shifts mass leftward relative to naive consensus: elevated probability on bins 1-3 (20-50) and reduced mass on bins 5-6 (60-80), because the documented restrictiveness of abstract-only exact-phrase matching outweighs the volume arguments that support the mid-range scenario. Bins 0 and 7-9 receive only residual mass consistent with the low overall evidence quality.

Consolidate

Strategy: logit_mean • Drafts: 3/3

Reviewers (3/3)

anchoring → shift_higher (low magnitude, low confidence) anchoring_v1

The draft rationale anchors its p50 at roughly 45–55 papers using the “900 AI papers/week × 0.3–0.5% overlap rate” heuristic, but this baseline is derived without any direct empirical measurement of the actual query overlap rate. Critically, the Bio Group includes extremely broad wildcard terms — “pandemic“, “epidemic“, “pathogen“, “outbreak“, “infectious disease“ — that are standard vocabulary in a large portion of q-bio, cs.LG (epidemiology), and medical informatics abstracts. Combined with AI Group terms like “machine learning”, “deep learning”, “neural network“, and “artificial intelligence” — which are also ubiquitous in biomedical ML papers — the true intersection rate is very likely higher than 0.3–0.5%. The draft treats the “mid-range (40–70)” scenario as its anchor when the broad wildcard coverage more naturally places the modal expectation toward the upper half of that range or into the 70–90 range. Additionally, source 7 notes that “submitted_date” in arXiv counts the latest version upload (including revisions), which inflates the effective count — this inflationary factor is mentioned but not structurally incorporated into the p50 placement. The draft’s mode sits in bin_3/~40–60 (bin_4) when it should arguably sit in bin_4/~50–70 (bin_5) given the breadth of both term groups and the revision-counting inflation.

Flagged concerns

• Optimistic anchor on overlap rate: The 0.3–0.5% overlap rate used to derive the central estimate is speculative and likely conservative. The Bio Group includes ‘pandemic‘, ‘epidemic‘, ‘pathogen‘, ‘outbreak‘, ‘infectious disease‘ — each individually common in biomedical ML abstracts. Using any one of these plus any AI term (machine learning, deep learning, neural network) is a very low bar; the modal expectation for this 69-day window should be higher than ~45 papers.
• Revision-counting inflation underweighted: Source 7 documents that arXiv’s ‘submitted_date’ filter counts latest-version uploads, not v1 only. The rationale mentions this but does not shift the central estimate upward to reflect that revised papers submitted before June 4 with revisions during the window would be included — this is a systematic upward bias on the count that the draft only acknowledges qualitatively (‘I discount somewhat’).
• Mode placed at lower boundary of plausible range: The draft’s modal bin is bin_3/40–60 (bin_4), which is at the lower end of the evidence-favored 40–70 range. Given the breadth of both query groups and the absence of any concrete restrictive evidence (no empirical count from a comparable period), placing the mode at 40–55 rather than 55–70 reflects optimistic anchoring toward the low end rather than the median of the plausible window.

ceiling → shift_higher (low magnitude, medium confidence) ceiling_v1

The key structural ceiling constraint the draft underweights is the answer range itself: “19.5 to 100.5 (open lower / open upper bound).” This is not merely a display range — it was calibrated from prior knowledge of what such queries actually return, implying a hard empirical ceiling near 100. The draft assigns only ~6.2% to bins 7+8+9 (80–100 and above), which seems low given the broad wildcard terms (pandemic, pathogen, epidemic, outbreak, infectious disease), but the above-100 bin (bin_9) receives only ~0.4%, correctly reflecting the range ceiling. However, the more important structural floor constraint is that the query window began June 4, 2026 — today — meaning some papers are already accumulated, and the monotone_floor applies. With ~9,800 AI papers/month and even a conservative 0.3% bio-term abstract overlap rate, the first day alone would yield ~1–2 papers, with ~65–90 expected by August 11. The draft’s combined mass in bins 0–2 (below 40) totals ~30.6%, which is structurally too high given: (1) the broad Bio Group wildcards (pandemic, pathogen, outbreak) are standard biomedical vocabulary appearing in many AI×health paper abstracts, (2) the 69-day window × ~900 AI papers/week implies ~9,000 AI papers, and even 0.3% overlap = ~27 papers just from narrow terms, while broader terms push well above 40. The floor constraint argues against bins 0–2 having 30% combined mass.

Flagged concerns

• Structural Floor — monotone accumulation underweights lower bins correctly but overweights bins 0–2: The draft assigns ~30.6% combined mass to bins 0–2 (below 40 papers). Given ~9,000 AI papers submitted over 69 days and broad wildcard Bio terms (pandemic, pathogen, epidemic, outbreak) that are standard biomedical vocabulary, the expected intersection rate of ~0.3–0.5% in abstracts yields ~27–45 papers from AI papers alone, before counting non-CS-primary bio papers that also use AI terms. The structural floor from the running accumulation (monotone_floor applies per interpretation) makes counts below 30 very unlikely.
• Structural Ceiling — answer range cap at 100.5 inadequately reflected in bin_9 mass: The answer range of 19.5–100.5 was likely calibrated from known arXiv query behavior, implying a hard empirical ceiling near 100. The draft correctly assigns only ~0.4% to above 100 (bin_9), but the rationale does not explicitly acknowledge the answer range as a structural constraint — it treats it as merely a display range. The ‘submitted_date’ filter counting revisions (source 7) could inflate counts, but the range cap suggests the question setter already accounted for this.
• Structural Floor — broad wildcard Bio terms create a higher implied minimum than draft acknowledges: Terms like ‘pandemic‘, ‘epidemic‘, ‘pathogen‘, ‘infectious disease‘, and ‘outbreak*’ are standard vocabulary in epidemiology, public health, and biomedical ML abstracts. Any paper on epidemic forecasting via neural networks, pathogen detection with deep learning, or infectious disease prediction with transformers would match. Over 69 days with the documented active AIxBio community (sources 2, 3) and high-volume AI paper submissions (~9,800/month per source 8), the structural minimum implied by these broad terms is well above 20. Bins 0 (below 20) and 1 (20–30) together receive ~13% of mass, which is structurally inconsistent with the breadth of the query terms.

math → shift_lower (low magnitude, medium confidence) math_v1

The draft distribution sums to 1.0 ✓ and shows reasonable internal consistency with the brief’s evidence hierarchy—mid-range bins (40–70) receive plurality mass (~0.55 combined), lower bins (~0.30) and higher bins (~0.12) are secondary. However, there is a subtle shape inconsistency: the brief’s evidence_summary emphasizes abstract-specific exact-phrase matching as “significantly more restrictive than full-text,” yet the inside-view rationale still derives the mid-range modal estimate (40–70, labeled “high evidence”) via ~0.3–0.4% overlap on ~22,500 AI papers—a calculation that assumes the broad wildcard terms (pandemic, pathogen, etc.) match abstracts at rates comparable to full-text. This creates a tension: if exact-phrase abstract matching truly constrains the count substantially (as stated), the overlap rate should be lower than 0.3–0.4%, which would push the modal estimate downward, yet the distribution’s peak (bin_3 at 0.21) sits at 40–50, implying the stated mechanism still yields ~45–55 papers. The mathematical implication is either (a) the brief’s estimate of overlap rate is too high for the stated restrictiveness, or (b) the brief should have flagged the 40–70 scenario as medium-evidence, not high-evidence. The draft does not explicitly reconcile this.

Flagged concerns

• consistency between stated mechanism and modal assignment: Rationale cites abstract-specific exact-phrase matching as ‘significantly more restrictive’ and notes the brief flagged low confidence_in_evidence. However, it then anchors the mid-range scenario (40–70, peak bin_3 = 0.21 at 40–50) on a 0.3–0.4% overlap rate applied to ~22,500 AI papers. If abstract matching truly is as restrictive as stated, the empirical overlap rate should be lower (perhaps 0.1–0.2%), which would imply a lower modal estimate (~22–45 papers). Either the overlap-rate assumption needs downward revision, or the distribution should shift leftward to reflect the stated restrictiveness.
• confidence vs. spread consistency: Brief flagged confidence_in_evidence as ‘low’ due to unknown overlap rates and query-construction edge cases. The draft’s distribution (0.12–0.21 across bins, no bin >0.25) reflects appropriate widening, but above 100 (bin_9) receives only 0.004 probability despite the low-evidence flag. If true overlap rates could be 2–3× the assumed median (due to unknowns), bin_9 should receive slightly more mass (0.01–0.02 range) to reflect tail uncertainty.
• bin labeling mismatch: Draft uses labels ‘20-30’, ‘30-40’, etc., but the schema requires EXACT bin_id keys bin_0 through bin_9. The submitted function call must map these labels to the correct bin_ids per the Bin Definitions legend (bin_0=Below 20, bin_1=20-30, bin_2=30-40, …, bin_9=Above 100). Current labels appear to be off by one or mislabeled.

Reconcile: partial_shift

Reviews consumed: 3

The anchoring and ceiling reviewers both suggested shift_higher (low magnitude), arguing the draft’s mass in bins 0-2 (~30%) is too high given broad wildcard Bio terms (pandemic, pathogen, epidemic, outbreak) that are standard biomedical vocabulary, and that revision-counting in submitted_date adds upward pressure. The math reviewer suggested shift_lower (low), arguing abstract-specific exact-phrase matching is more restrictive than the 0.3-0.4% overlap implies. Since two reviewers agree on direction (higher) with concrete structural reasoning (broad wildcards + active AIxBio community + ~9,000 AI papers in window), I weight that 2:1 against the math reviewer’s restrictiveness concern—and note the math reviewer’s concern is partially already baked into the draft’s wide spread. I shift modest mass from bins 0-2 to bins 4-6, peaking at 50-60 (bin_4). Math reviewer’s bin-labeling concern is a non-issue since the bin_id mapping is correct in the schema. The math reviewer’s call for slightly more bin_9 mass given low evidence confidence is reasonable and integrated.

Rejected critiques

• [math] Suggested shift_lower because abstract-specific exact-phrase matching should imply a lower overlap rate (~0.1-0.2%) than the 0.3-0.4% used in the draft mechanism, pushing the mode toward 22-45 papers.

  • Why rejected: Two other reviewers (anchoring, ceiling) provided structural counter-arguments — the broad wildcard Bio terms (pandemic, pathogen, epidemic, outbreak, infectious disease*) are standard biomedical/epidemiology abstract vocabulary, and revision-counting via ‘submitted_date’ inflates counts. The agreement of two independent reviewers on direction (higher) with concrete mechanisms outweighed the math reviewer’s restrictiveness argument. The draft already kept substantial mass in bins 1-3 reflecting this concern; I did not shift further leftward.

• [math] Bin labeling mismatch — claimed labels were off by one.

  • Why rejected: The bin_id mapping in the submitted schema is correct (bin_0=Below 20, bin_1=20-30, …, bin_9=Above 100). The draft’s internal label dictionary is descriptive and the bin_id mapping in the by_id field is properly aligned per the legend.

Final Distribution

Evidence Sources (11)

1. arXiv Monthly Submissions — As of June 1, 2026, total arXiv submissions stood at 3,059,610, and monthly submissions hit nearly 28,000 by late 2025, reflecting super-exponential growth.
2. The Biosecurity Blind Spot: Systematic Dual-use Detection in Open Science Infrastructure — A 2026 arXiv paper screened 52,000 bioRxiv preprints (2024–2025) using LLM evaluation for dual-use research of concern (DURC) content, demonstrating active intersection of AI and biosecurity research.
3. A non-definitive AIxBio reading list — The intersection of AI and biosecurity research (AIxBio) has a growing reading list and active research community, with arXiv papers actively being published as of 2025-2026 on topics including LLM-enabled biosecurity threats and generative AI for biosciences.
4. arXiv Advanced Search URL with date parameters — The arXiv advanced search UI includes a date filter with parameters date-from_date, date-to_date, and date-date_type=submitted_date, supporting filtering by submitted_date over a user-specified range.
5. untitled — A high volume of arXiv papers matching broad AI+biosecurity query terms exist from June-August 2025, covering topics like epidemic forecasting using neural networks, LLMs for disease surveillance, machine learning for infectious disease modeling, and deep learning for pathogen analysis.
6. ArXiv AI Paper Trends: 6.5× More Papers in Seven Years — ArXiv AI papers (cs.AI/cs.CL/cs.LG) totaled 114,888 in 2025, growing at 30.7% compound annual growth rate from 17,635 in 2018; cs.LG had 46,004 papers in 2025 alone (about 900/week across all AI categories).
7. sgnoohc/arxiv-submission-analysis GitHub — The arXiv advanced search’s submission date filter ‘submitted_date’ counts a paper based on when its LATEST version was uploaded (including revisions), while ‘submitted_date_first’ counts only initial v1 submissions; this distinction affects any count query significantly, with revisions causing apparent spikes in submission counts.
8. Measuring Technological Progress in Real Time with ArXiv - BBVA Research — Computer science preprints (including AI research) grew to nearly 50% of all arXiv submissions in 2025, with AI representing about 70% of all computer science output; arXiv had 28,000 total monthly submissions by late 2025, implying roughly 9,800 AI papers/month (assuming 50% CS share and 70% AI within CS).
9. Search | arXiv e-print repository (Help tips) — arXiv’s search interface explicitly supports wildcards in any field (including Abstract): use ? to replace a single character and * to replace any number of characters. The only restriction is that wildcards cannot appear in the first character position.
10. Advanced Search | arXiv e-print repository — arXiv’s Advanced Search interface provides a Boolean operator selector (AND / OR / NOT) for each additional search row, and supports adding multiple term-rows via “Add another term +”. Each row specifies its own field (e.g., Abstract) and its Boolean relationship to the preceding row.
11. Advanced Search | arXiv e-print repository — arXiv’s Advanced Search interface does NOT support grouping multiple OR terms into a single search row for a chosen field. Each row accepts one search term (with optional wildcard), joined to the next row by AND/OR/NOT. To express “(bioweapon OR pandemic) AND (AI OR machine learning)” in the Advanced Search GUI, a user would need multiple rows: rows for the Bio Group terms joined by OR (targeting Abstract), then rows for the AI Group terms joined by OR (targeting Abstract), with an AND between the two groups — but the interface applies Boolean operators row-by-row, not with parenthetical grouping, which means OR and AND precedence cannot be explicitly parenthesized in the GUI.

wooley_rhino • run b691d179 • $3.02 • 9m6s

Extracted Percentile Values: {0.1: 15, 1.0: 21, 5.0: 28, 10.0: 32, 15.0: 36, 20.0: 39, 25.0: 42, 30.0: 44, 35.0: 46, 40.0: 48, 45.0: 50, 46.0: 51, 47.0: 51, 48.0: 52, 49.0: 53, 50.0: 53, 51.0: 54, 52.0: 55, 53.0: 55, 54.0: 56, 55.0: 57, 60.0: 60, 65.0: 63, 70.0: 67, 75.0: 71, 80.0: 76, 85.0: 82, 90.0: 89, 95.0: 98, 99.0: 115, 99.9: 135}

Model’s Answer:

Analysis of AI + Biosecurity arXiv Papers (June 4 - August 11, 2026)

Understanding the Question

This question asks for the count of arXiv papers with abstracts containing at least one term from both the AI Group and Bio Group, submitted during a ~69-day window (roughly 10 weeks).

Key Factors to Consider

Term Breadth Analysis:

• Broader Bio terms like “outbreak“, “pandemic“, “infectious disease“, and “pathogen“ will capture substantial ML/epidemiology work
• Narrower Bio terms like “biosecurity”, “biodefense”, “dual-use” target the explicit AI+biosecurity intersection
• AI terms span classical ML (“machine learning”, “deep learning”) to modern LLMs (“foundation model*”, “GPT”, “LLM”)

Research Momentum Indicators:

• May 2026 “Biosecurity Blind Spot” preprint on dual-use detection demonstrates active work
• April 2026 AIxBio Hackathon outputs likely entering the publication pipeline
• June 2026 LAP Agent-to-Instrument Protocol paper showing autonomous lab work
• Nature coverage (May 2026) on AI and bioweapons driving policy-response papers
• Multiple policy reports (RAND Feb 2026, CSIS Aug 2025) spurring academic responses

Seasonal Considerations:

• June-August includes summer conference deadlines
• Academic activity slightly reduced in some regions but arXiv preprints continue steadily

Estimation Approach

Component Analysis:

1. Infectious disease + ML papers: Post-COVID, substantial literature on outbreak detection, epidemic forecasting. Could yield 20-40 papers with terms like “pandemic*” + “machine learning”
2. Biosecurity-explicit AI papers: More niche but growing. Dual-use, biosafety, biodefense-focused work. Estimate 10-25 papers
3. Synthetic biology + AI papers: DNA synthesis, gene synthesis with ML/generative models. Estimate 5-15 papers
4. Overlap correction: Some papers fit multiple categories, so total likely 35-70 papers at median

Default Range Context: The default range of 19.5-100.5 suggests forecasters expect meaningful activity without extreme outcomes. This aligns with my component analysis.

Low Outcome Scenario (< 30 papers)

• Summer publication slowdown more severe than expected
• Terms don’t co-occur as frequently in abstracts
• Community uses different terminology than specified
• Likelihood: ~10-15%

High Outcome Scenario (> 80 papers)

• Strong hackathon pipeline produces surge of papers
• Policy responses drive substantial governance/assessment papers
• Broad terms (outbreak, pandemic) capture more epidemiology work than expected
• Major AI safety labs release biosecurity evaluations
• Likelihood: ~10-15%

Key Assumptions and Risks

1. Query term matching behavior: Assumes arXiv search functions as specified. If wildcard behavior differs, counts could vary significantly. Probability of issue: ~5%

2. Publication momentum continues: Assumes no major disruption to academic publishing. Probability of disruption: ~3%

3. Baseline estimate accuracy: My component estimates rely on extrapolation from limited data. True baseline could be 30% different. Probability: ~20%

Final Distribution Rationale

Centering around 48-52 papers (roughly 5-6 per week), with substantial uncertainty reflecting:

• Lack of precise historical data on this exact query
• 10-week window provides significant variability
• Growing field suggests upward bias relative to simple extrapolation
• Wide confidence intervals appropriate given limited authoritative historical metrics

Percentile .1: 15 Percentile 1: 21 Percentile 5: 28 Percentile 10: 32 Percentile 15: 36 Percentile 20: 39 Percentile 25: 42 Percentile 30: 44 Percentile 35: 46 Percentile 40: 48 Percentile 45: 50 Percentile 46: 51 Percentile 47: 51 Percentile 48: 52 Percentile 49: 53 Percentile 50: 53 Percentile 51: 54 Percentile 52: 55 Percentile 53: 55 Percentile 54: 56 Percentile 55: 57 Percentile 60: 60 Percentile 65: 63 Percentile 70: 67 Percentile 75: 71 Percentile 80: 76 Percentile 85: 82 Percentile 90: 89 Percentile 95: 98 Percentile 99: 115 Percentile 99.9: 135

High, steadily growing preprint submission volume and broad disciplinary overlap set the outer scale.

Abstract-field boolean matching, phrase and wildcard behavior, and posting delays constrain observable counts.

A persistent base of disease- and pathogen-focused manuscripts combines with routine use of computational-method terminology to produce background matches.

Recent year-on-year growth in computational-biosecurity language lifts the center but remains well below the 2020 shock peak.

A concentrated workshop, benchmark, or high-profile outbreak could create a short burst adding dozens of matches.

Policy changes to submission filtering, category curation, or allowed abstract phrasing would truncate low-effort position pieces and push counts down.

Tokenization, exact-phrase handling, hyphenation, and wildcard interpretation can flip a small number of edge-case abstracts.

Timing and posting delays at the window end, revisions that alter abstracts, and an unpredictable topical surge are the dominant residual uncertainties.

Forecast (Q420): median = 67.9144

• The “bio” side of the search is very broad, so it likely captures a lot more than explicitly biosecurity/bioweapon work—especially common outbreak/pathogen/pandemic language that shows up in many health-adjacent papers.
• The “AI” side is also broad and widely used in abstracts, so many computational epidemiology and pathogen-modeling papers should match even when they’re not framed as “biosecurity.”
• Big uncertainty comes from arXiv’s overlap reality (and abstract wording): whether ML/AI phrases appear in the abstract (not just methods) and whether comparable work is more common on arXiv vs. preprints/journals elsewhere.
• There may be some uplift from outbreak salience around the window, with the Ebola PHEIC noted close to the start of the period, but it’s treated as a second-order effect with timing lag uncertainty.
• Overall expectation is steady-ish but not perfectly smooth: submissions over a fixed 69-day window are fairly predictable in aggregate, yet bursts near major dates/events can shift the count upward.