What are the limits to biomedical research acceleration through general-purpose AI?

_{Hebenstreit, K.†, Convalexius, C.†, Reichl, S.†, Huber, S., Bock, C., & Samwald, M. (2025). Scientific Reports (in publication). · †Equal contribution · Preprint: arXiv:2508.16613}

---

General-purpose artificial intelligence (GPAI) is widely expected to transform scientific discovery, but in biomedicine its real-world impact remains uncertain. This study provides a systematic analysis of how much current biomedical research could realistically be accelerated by GPAI, and where fundamental limits remain.

Research tasks We mapped the biomedical research lifecycle into nine major tasks: Cognitive (blue): information processing, analysis, decision-making Physical (red): lab procedures, experiment execution	Acceleration estimates Scoping review of 16 publications reveals a bimodal distribution: Next-level: ~2x (current, partial automation) Maximum-level: ~100x cognitive, ~25x physical

Biological time constants

Large task-level speed-ups do not translate into equivalent reductions in overall project duration. Many biomedical projects are constrained by biological processes that cannot be compressed (cell growth, organism development, disease progression).

Modeling a hypothetical 3-year project with 3 months of incompressible biological processes:

Project duration	Physical: No GPAI	Physical: Next-level (2x)	Physical: Max-level (25x)
Cognitive: No GPAI	36 months	32 months	27 months
Cognitive: Next-level (2x)	24 months	20 months	15 months
Cognitive: Max-level (100x)	12 months	7.7 months	3.6 months

Even with maximum acceleration, the lower bound is ~3.6 months (10x overall), with incompressible biological processes dominating.

Expert elicitation

Eight senior biomedical researchers participated in an expert survey.

How is time distributed across research tasks?

Project durations Experts reported average project durations of ~6 years for high-impact publications.

Are maximum-level acceleration estimates plausible?

Acceleration plausibility Experts considered strong acceleration plausible for manuscript preparation and publication processes, but were skeptical about hypothesis generation, experiment design, and execution.

What limits acceleration potential?

Limiting factors All experts identified scientific community assimilation as a moderate to crucial bottleneck for realizing acceleration benefits.

Realizing the full potential of GPAI-driven research acceleration will require coordinated investments in automation infrastructure, improved data accessibility, and reforms in research organization and publication practices.

Code & data

This repository contains the data and R scripts used to generate the figures:

Figure	Description	Data	Script
1	GPAI capability framework	—	plot_capability_model.R
2	Major research tasks	—	(graphical software)
3	Acceleration factors	acceleration_factors.csv	plot_accelerations.R
4	Project time durations	project_times.csv	plot_project_times.R
5	Plausibility estimates	plausibility.csv	plot_plausibility_estimates.R
6	Limiting factors	limiting_factors.csv	plot_limitation_estimates.R

Citation

If you find our work useful in your research, please cite:

Scientific Reports (2025)

Hebenstreit, K.†, Convalexius, C.†, Reichl, S.†, Huber, S., Bock, C., & Samwald, M. (2025). What are the limits to biomedical research acceleration through general-purpose AI? Scientific Reports (in publication).

ArXiv Preprint (2025)

doi: 10.48550/arXiv.2508.16613

† Equal contribution