LLM Novice Uplift on Dual-Use, In Silico Biology Tasks
| Authors | Chen Bo Calvin Zhang et al. |
| Year | 2026 |
| Field | AI / Agents |
| arXiv | 2602.23329 |
| Download | |
| Categories | cs.AI, cs.CL, cs.CR, cs.CY |
Abstract
Large language models (LLMs) perform increasingly well on biology benchmarks, but it remains unclear whether they uplift novice users -- i.e., enable humans to perform better than with internet-only resources. This uncertainty is central to understanding both scientific acceleration and dual-use risk. We conducted a multi-model, multi-benchmark human uplift study comparing novices with LLM access versus internet-only access across eight biosecurity-relevant task sets. Participants worked on complex problems with ample time (up to 13 hours for the most involved tasks). We found that LLM access provided substantial uplift: novices with LLMs were 4.16 times more accurate than controls (95% CI [2.63, 6.87]). On four benchmarks with available expert baselines (internet-only), novices with LLMs outperformed experts on three of them. Perhaps surprisingly, standalone LLMs often exceeded LLM-assisted novices, indicating that users were not eliciting the strongest available contributions from the LLMs. Most participants (89.6%) reported little difficulty obtaining dual-use-relevant information despite safeguards. Overall, LLMs substantially uplift novices on biological tasks previously reserved for trained practitioners, underscoring the need for sustained, interactive uplift evaluations alongside traditional benchmarks.
Engineering Breakdown
Plain English
This paper measures whether novice users actually perform better at complex biology tasks when given access to LLMs compared to just having the internet. The researchers ran a controlled human study where participants tackled eight biosecurity-relevant biology tasks with up to 13 hours per task, splitting groups between LLM-assisted and internet-only access. They found dramatic uplift: novices with LLMs were 4.16 times more accurate than controls (95% CI [2.63, 6.87]), and on tasks with expert baselines, LLM-assisted novices approached or matched expert performance. This matters because it directly addresses whether AI accelerates scientific research for humans and simultaneously raises dual-use risk concerns—unskilled people can now do dangerous biological work more effectively.
Core Technical Contribution
The core contribution is empirical rather than algorithmic: a rigorous, large-scale human-in-the-loop evaluation framework that measures genuine user uplift on real-world biosecurity tasks, not just benchmark performance. Prior work showed LLMs score well on biology benchmarks, but this paper is the first to measure whether that translates to actual human performance improvement in time-constrained, open-ended problem solving. The novelty is methodological—the authors designed a controlled study comparing human+LLM workflows against human+internet-only workflows across multiple models and tasks, with expert baselines for validation. This moves beyond synthetic task metrics to measuring what matters: can non-experts actually accomplish dangerous biology tasks better with AI assistance?
How It Works
The experimental design is straightforward but rigorous: recruit novice participants with no biology background, assign them to either LLM-assisted or internet-only control groups, and measure task completion accuracy on eight biology benchmarks spanning DNA analysis, protein design, drug synthesis, and other biosecurity-relevant domains. For each task, participants receive the same problem statement and have up to 13 hours to solve it (with realistic time budgets per task type); LLM participants can query the model iteratively while controls use search engines and standard biology resources. The researchers measure accuracy against ground-truth solutions and compare results statistically, calculating uplift ratios and 95% confidence intervals. For validation, they compare novice+LLM performance against expert baselines (measured on internet-only resources) to see if AI closes the expertise gap. The key technical detail is holding time budgets constant—this ensures the uplift signal comes from LLM quality, not from unlimited iteration time.
Production Impact
For engineers building AI systems, this paper has sobering implications: it's concrete evidence that LLM-assisted workflows genuinely enable non-experts to perform dangerous tasks at near-expert levels. In production terms, this means deploying LLMs for biology tasks creates real dual-use risk—you cannot assume safety through expertise barriers anymore. For legitimate scientific research, this suggests LLM integration into biology pipelines (drug discovery, protein folding, synthetic biology workflows) will substantially improve output quality and speed for teams without deep domain expertise, justifying investment in specialized LLM fine-tuning and domain-specific RAG systems for biology. The trade-off is critical: integrating LLMs dramatically reduces the expertise barrier, making both beneficial research and harmful research more accessible. Engineers deploying biology-focused LLMs must implement strong access controls, audit logs, and task filtering—the cost of deployment now includes infrastructure for detecting and blocking dangerous use cases.
Limitations and When Not to Use This
The study assumes novice participants behave like real-world threat actors, which they don't—the lab setting, ethical oversight, and time constraints don't capture how an adversary would actually use an LLM with fewer constraints. The paper also doesn't measure what happens with more capable future LLMs; the study uses models from ~2025-2026, and capabilities are rapidly improving, so uplift ratios could be even higher. The benchmarks are in-silico (computational biology tasks) and don't measure physical wet-lab competence, so there's a gap between solving a DNA design problem and actually building dangerous pathogens in practice. Finally, the paper doesn't deeply explore mitigation strategies—it documents the risk but offers limited guidance on how to build LLM systems that maintain the benefits (scientific acceleration) while reducing dual-use harm.
Research Context
This paper sits at the intersection of LLM capability evaluation and AI safety research, building on earlier work that showed LLMs have broad biology knowledge but extending it to measure real human impact. It relates to benchmark research (BioGPT, SciBench, etc.) that evaluates LLMs on biology tasks, but goes further by asking the crucial question: do benchmark scores translate to human productivity gains? The work also connects to emerging dual-use AI safety literature, which flags that general-purpose models may inadvertently enable harmful applications. This study opens up a research direction on measuring human uplift across other dangerous domains (chemistry, bioweapons design, cybersecurity attack planning) and exploring which safeguards (filtering, access control, model editing) preserve uplift for beneficial work while blocking misuse.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::