Trusted Evidence in the Age of AI

What Makes AI Trustworthy in Pharmaceutical R&D?

At Wiley, we recently surveyed thousands of researchers on AI adoption. A year and a half ago, 57 percent reported using AI in their work. That number is now 84 percent. The striking detail is that of those 84 percent, the vast majority cited serious concerns about the quality, accuracy, and trustworthiness of the outputs they were relying on. It’s likely that they’re using these tools anyway because the productivity gains are too compelling to ignore, but they don’t fully trust what’s coming back. In fact, their concerns have increased, not decreased over the past year.

For general-purpose applications, that’s a manageable tension. In pharmaceutical R&D and clinical practice, it is not. When I talk with R&D leaders about building AI they can actually rely on, I come back to three things: quality, accuracy, and attribution.

Quality, in this context, means grounding AI outputs in peer-reviewed, version-of-record content from authoritative repositories – the sources your scientists already know and cite. The “deep research” function in most general-purpose AI tools is really doing a broad sweep across whatever the model can reach on the web, which often means using incomplete information. Real depth means full-text articles from authoritative repositories. AI teams at pharmaceutical companies consistently find that public sources like PubMed are a reasonable starting point but insufficient for high-stakes applications. Sooner or later, they need the full-text literature.

Accuracy means grounding AI system responses in current evidence in real time – typically using retrieval-augmented generation (RAG) – rather than relying on model weights. Model weights return probabilistic results, not deterministic, which is the source of hallucinations. The analogy I always use: model weights are a closed-book exam. In pharmaceutical R&D, you always want open book. This matters especially given how fast the literature moves; in some medical specialties, knowledge is doubling every 73 days.

Attribution means full citation back to the original source so that the user can follow the chain of evidence all the way back. AI that cannot show its work is a liability in scientific contexts. When a system can point to the specific study behind a recommendation and a researcher can click through to verify it, you build a feedback loop of credibility and reproducibility.