Thirty co-authors from institutions across Europe and the US just published what amounts to a 24-page intervention letter for the AI-in-biology community. Their message, landing in Nature Methods this March: we built this incredible machine, and now half the parts are missing screws.
The Trust Problem Nobody Wants to Talk About
AI in life sciences has been on an absolute tear. AlphaFold predicted 200 million protein structures and won a Nobel Prize (Hassabis & Jumper, Chemistry 2024). Drug discovery pipelines run on neural networks. Genomic analysis that used to take weeks now takes hours. By every measure, AI is the best thing to happen to biology since the microscope.
But here's the awkward part: a huge chunk of published AI research in the life sciences can't be reproduced. One review of 257 ML publications in biomedical research found that only 16 shared their data or used publicly accessible datasets. Sixteen. Out of 257. That's a 6% sharing rate, which is roughly the success rate of asking strangers on the subway for their WiFi password.
Farrell, Tosatto, and their 28 co-authors call this an "erosion of trust," and they're not being dramatic. If you can't rerun someone's model and get the same results, you don't have science - you have a very expensive anecdote.
Three Flavors of Broken
The paper breaks the problem into three buckets, each one spicier than the last.
Reusability is the first headache. Most published AI models arrive like IKEA furniture missing the Allen wrench, the instructions, and half the dowels. Training data? Undisclosed. Preprocessing steps? Vibes-based. Model weights? Behind a login wall that hasn't worked since 2023. The Model Openness Framework tried to grade AI models on transparency, and the results were... not great. Many so-called "open" models practice what researchers politely call "openwashing" - slapping an open-source label on a locked box.
Reproducibility is reusability's meaner sibling. Even when code is shared, ML models are exquisitely sensitive to training conditions, random seeds, hardware quirks, and the phase of the moon (okay, not that last one, but sometimes it feels close). Semmelrock et al. (2025) catalogued the barriers: incomplete documentation, limited compute access, nondeterminism baked into GPU operations, and the uncomfortable reality that many researchers simply don't budget time for making their work reproducible.
Sustainability is the one that makes climate scientists wince. Training GPT-4 consumed an estimated 50 gigawatt-hours of energy - enough to power a small city for a year. Global data center electricity consumption hit 415 TWh in 2024 and is projected to double by 2030. AI-optimized servers alone could consume 432 TWh annually by decade's end. Meanwhile, most papers don't even report how much energy their model training used. It's like running a restaurant and refusing to tell anyone where the food comes from.
The 300-Piece Toolkit
Rather than just complaining (though the complaining is well-earned), the team mapped nine concrete recommendations across more than 300 components of the AI ecosystem. Think of it as a checklist for not being part of the problem. The recommendations span everything from sharing preprocessing pipelines and using containerized environments (Docker, Singularity) to measuring and reporting energy consumption per training run, including hardware specs and operational carbon emissions.
The framework aligns with FAIR principles - Findable, Accessible, Interoperable, Reusable - but extends them specifically for AI workflows. And it's built on community consensus, not one lab's opinion, which gives it a better shot at actually being adopted. For anyone trying to navigate this sprawling ecosystem of tools, standards, and best practices, visual mapping tools like mapb2.io can help untangle the web of dependencies and decision points that the authors describe.
Why You Should Care
If AI in biology keeps growing without guardrails for openness and sustainability, we end up with a field that produces impressive demos but unreliable science - powered by the carbon footprint of a small nation. The authors aren't anti-AI. They're pro-AI that actually works twice, costs less to the planet, and doesn't require a decoder ring to understand.
The paper's real contribution isn't pointing out the mess. It's handing researchers a mop and a map.
References:
- Farrell, G., Adamidi, E., Andrade Buono, R. et al. Open and sustainable AI: challenges, opportunities and the road ahead in the life sciences. Nature Methods (2026). DOI: 10.1038/s41592-026-03037-6
- White, M. et al. The Model Openness Framework. arXiv (2024). arXiv:2403.13784
- Semmelrock, S. et al. Reproducibility in machine-learning-based research. AI Magazine (2025). Wiley
- Huerta, E.A. et al. FAIR for AI: An interdisciplinary perspective. Scientific Data (2023). DOI: 10.1038/s41597-023-02298-6
- IEA. Energy demand from AI. Energy and AI Report (2025). IEA
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.