Life sciences face a 'scientific content crisis' in AI adoption
AI is moving into R&D, but many models are running on thin evidence. A new poll from the Pistoia Alliance's US conference points to a growing scientific content crisis: too little data, unclear provenance and weak governance are limiting trust and results.
More than 170 experts from pharma, tech and academia met in Boston to discuss these gaps and what it will take to fix them. The message was clear: without complete, licensed and traceable content, AI will underperform where it matters most-patient outcomes and research decisions.
The signal in the data: what's missing
27 percent of life science professionals don't know what scientific content their AI or LLM systems use, or they rely only on titles and abstracts. That's thin context for high-stakes work.
Only about 36 percent are feeding internal documents into models. The result is incomplete datasets, low confidence in outputs and slower adoption across teams who need evidence they can cite and audit.
Governance, licensing and risk
"Many AI models are not yet drawing on the full range of scientific evidence needed to deliver authoritative results. Many organisations are still in a learning phase on data and governance and, given the stakes for patient safety, that cannot be ignored," said Neal Dunkinson of the Copyright Clearance Center.
38 percent of respondents said copyright and licensing policies are unclear or not enforced. That invites legal exposure and needless rework. As Dunkinson noted, datasets must be AI-ready: properly structured, licensed and transparent.
Standards for AI agents
Half of respondents flagged the lack of shared verification standards as the top barrier to adopting AI agents. In response, Robert Gill, Agentic AI programme lead at the Pistoia Alliance, encouraged organisations to join the Alliance's new agentic AI project to develop standards for safe, scalable deployment and full visibility into training data.
Learn more about the Alliance's work and collaboration opportunities at Pistoia Alliance.
AI in practice: what's working
Conference sessions showed concrete applications: EPAM on streamlining clinical operations with AI; the Michael J. Fox Foundation on using knowledge graphs to accelerate Parkinson's research; and AbbVie on improving pharmacovigilance.
A roundtable hosted by Elsevier with leaders from Eli Lilly, Pfizer, Bayer, J&J and Takeda reached a simple consensus: adoption sticks when tools are problem-led, intuitive and fit neatly into current workflows. No friction, no extra clicks.
Change management is still the bottleneck. Echoing Pistoia's Lab of the Future survey, 34 percent cited a shortage of skilled talent as a key barrier to adoption.
What you can do now: a workable playbook
- Build a full content inventory: publications, internal reports, protocols, lab notes, safety data, real-world evidence. Track sources and access rights.
- Make data AI-ready: structure, normalize, version, de-duplicate and watermark provenance; attach licenses; capture consent where required.
- Governance that actually runs: clear policies, enforced access controls, usage logs, human-in-the-loop review and a named model owner.
- Benchmark before scale: define tasks, gold-standard datasets and acceptance criteria; monitor citation accuracy and hallucination rates.
- Agent safety: run agents in sandboxes, limit tools and data scopes, require audit trails for every action and source.
- Procurement checklist: license verification, indemnity, content coverage maps, red-teaming results and data-retention terms.
- Upskill the people: train scientists, data stewards and legal on prompt practices, evaluation methods and licensing basics.
Metrics that matter
- Evidence coverage: percent of required evidence types included per use case.
- Provenance completeness: percent of outputs with source citations and license status.
- Policy adherence: number of models with cleared licenses and documented data lineage.
- Model quality: hallucination rate, citation accuracy and task-level success against gold standards.
- Cycle time: time to provenance proof for a given answer or recommendation.
A shared path forward
"The same concerns around AI trust, transparency and skills were raised at both our US and European conferences. These issues are clearly universal across the life sciences community," said Dr Becky Upton, President of the Pistoia Alliance.
Collaboration on standards, data quality and practical implementation will decide whether AI becomes dependable in labs and clinics. The Alliance will continue these discussions at its spring meeting in London.
Skill up your team
If talent gaps are slowing your AI projects, explore current training mapped to roles and needs: AI courses by job.
Your membership also unlocks:
