Model Evaluation · · 8 min read
How to build a pharma AI benchmark
Acumen Editorial · Updated November 20, 2026
On November 11, researchers at the Icahn School of Medicine at Mount Sinai published a controlled study in The Lancet Digital Health. They gave 12 AI models, from companies including Google, OpenAI and Anthropic, 33 physician-written clinical scenarios. Each case ran alone or right after a real, clearly labeled drug advertisement. With the ad present, models picked the advertised drug in 47.6% of responses, against 35% without it. In four scenarios, a model's choice of the advertised drug went from 0% to 100%.
The answers still looked reasonable. That is the problem. No public leaderboard would have caught it.
For pharma, the study cuts two ways. It raises regulatory and reputational questions for any brand whose promotion sits near AI-generated health answers. And it shows that the models your medical affairs, regulatory and commercial teams run internally carry biases you only find by testing for them. Generic medical benchmarks don't test for them. You need your own.
Why do pharma teams need their own AI benchmark?
Public medical benchmarks measure clinical knowledge in general. Pharma work is narrower and more regulated. A medical information response has to stay on label, cite the right source and balance benefit with risk. A regulatory summary has to trace back to the document it came from. A rep-facing copilot has to stay inside approved claims.
None of that shows up in an exam score. And these models are already in use: Consilience Strategy reports that internal LLM platforms are now standard at the top 50 pharma companies.
Domain benchmarks are starting to appear. In August 2026, Insilico Medicine launched a Drug Discovery and Development Benchmark as a Service with more than 300 evaluations across five domains, scored against expert reference baselines. RxBench, a preprint benchmark of 2,259 prescription-review items checked by experienced clinical pharmacists, tested 18 models across 14 error types. Both follow the same pattern: real tasks, expert reference answers, decontaminated data. Neither covers medical information, regulatory writing or commercial content. That part is yours to build.
What should a pharma AI benchmark measure?
Start with the job, then pick the dimensions that decide whether an output is usable. We use six.
| Dimension | What it asks | How to measure it |
|---|---|---|
| Scientific accuracy | Is the answer correct and current? | Expert reference answers and rubric criteria per item |
| Label and claim adherence | Does it stay inside the approved label and claims? | Items with off-label traps; reviewers flag every excursion |
| Source fidelity | Can every statement be traced to a cited document? | Check each citation against the source |
| Fair balance | Does it give risk information alongside benefit? | Rubric criteria written by medical-legal-regulatory reviewers |
| Promotional neutrality | Does promotional content change the answer? | Paired items with and without injected promotional text |
| Reliability and regression | Same answer every run, and every release? | Rerun each item three or more times on a frozen set |
Promotional neutrality is the new one, and the Mount Sinai design is easy to copy. Keep the case and the options fixed. Add a promotional insert to half the runs. Measure how often the recommendation moves. In the study's free-text subset, models mentioned the ad only 29.9% of the time but repeated its claims more often. So grade what the answer says, not whether it discloses its inputs.
Reliability matters as much. In Stanford's PhysicianBench, reported by Stanford HAI in October 2026, the best agent completed 46% of clinical tasks on one attempt and 28% across all three runs. A single pass hides that gap.
How do you build the benchmark set?
The set is the asset. Treat it like one.
Start from real requests. Pull the questions your medical information team receives, the summaries your regulatory writers draft and the claims your field teams ask about. De-identify them and remove anything confidential.
Write items with reference answers. Each item gets a prompt, the source documents a correct answer should use, a reference answer and rubric criteria. Criteria reward correct, sourced, balanced answers. They penalize off-label statements, missing safety information and unsupported claims.
Add adversarial items. Off-label questions phrased innocently. Prompts that carry promotional text. Questions about a label that changed recently. Cases where the right answer is that the company can't say.
Cover the portfolio. Weight items by therapeutic area, product and risk tier. A benchmark built mostly on one brand tells you little about the rest.
Keep it held out. Store the set apart from any training or tuning pipeline. Never hand it to a model vendor for tuning. Log who wrote each item and when.
Who should write and grade the items?
The people who would sign off on the answer in real life. Medical science liaisons and medical information specialists for scientific content. Pharmacists for dosing, interactions and prescription review. Regulatory writers for submission-style summaries. Medical-legal-regulatory reviewers for fair balance and claims.
OpenAI's HealthBench set the bar for this approach in 2025, with conversation-specific rubrics written by 262 physicians. Pharma needs the same rigor from different specialists. Use at least two reviewers per item and a third to adjudicate disagreements. An LLM grader can score clear, low-risk criteria once it agrees with expert grades on a calibration set. Experts should grade every failure and every high-risk item.
How do you keep the benchmark useful over time?
Rerun it on every model change. Pharma platforms often sit on frontier models that update with little notice. A frozen set lets you compare results item by item, not just as an average.
Refresh items when labels, guidelines or safety information change. Retire items that have leaked. Add new failure modes as they surface, the way the Mount Sinai team surfaced one this month.
Document everything. The FDA and EMA guiding principles for AI in drug development, released January 14, 2026, call for risk-based assessment of AI performance and for data quality and governance across the AI lifecycle. A documented benchmark, with provenance for every item, is how you show that work.
Checklist: what to ask a vendor about a pharma AI benchmark
- Who wrote the items, and what are their credentials and specialties?
- Is every reference answer tied to a specific source document?
- How many reviewers grade each item, and how are disagreements adjudicated?
- Does the set include off-label, fair-balance and promotional-injection items?
- Is each item run more than once, with all-runs pass rates reported?
- How is the set kept out of training and tuning data?
- Can we see item-level results, not just an average?
- How are items refreshed when labels or guidelines change?
- What documentation do we get for quality, procurement and regulatory review?
- Do we own the set and can we rerun it ourselves?
How Acumen helps
Acumen builds pharma benchmarks with verified experts: MSLs, medical information specialists, pharmacists, regulatory writers and medical-legal reviewers, matched to your therapeutic areas. We write items and rubrics from real request patterns, adjudicate with multiple reviewers and deliver every set with full provenance, so it holds up in procurement and regulatory review.
See our medical affairs and regulatory data capability, or read how to evaluate a clinical LLM.
FAQ
What is a pharma AI benchmark?
A pharma AI benchmark is a private test set that scores a model on pharma-specific work: medical information answers, label and claim adherence, fair balance, source fidelity, prescription review and resistance to promotional bias. Experts such as MSLs, pharmacists and regulatory reviewers write the items, reference answers and rubrics. The set is held out from training and rerun on every model release.
How is a pharma AI benchmark different from HealthBench?
HealthBench measures general health conversations against physician-written rubrics. A pharma benchmark measures whether answers stay on label, cite approved sources, balance benefit with risk and resist promotional content. The method is similar, with case-specific rubrics and expert grading, but the items, reviewers and failure modes are specific to medical affairs, regulatory and commercial work.
What did the Mount Sinai drug advertising study find?
Published in The Lancet Digital Health on November 11, 2026, the study tested 12 AI models on 33 physician-written scenarios, with and without a labeled drug ad. Models chose the advertised drug in 47.6% of responses with the ad versus 35% without it. In four scenarios, a model's choice of the advertised drug rose from 0% to 100%.
How do you test a model for promotional bias?
Use paired items. Keep the clinical case and answer options fixed, then run each item with and without promotional text placed before it. Measure how often the recommendation changes and whether the answer repeats promotional claims. Grade the content of the answer, not only whether it discloses the ad. Rerun each pair several times to separate real bias from run-to-run noise.
Who should grade a pharma AI benchmark?
The experts who would approve the answer in practice. MSLs and medical information specialists grade scientific content, pharmacists grade dosing and prescription review, regulatory writers grade summaries and medical-legal reviewers grade fair balance and claims. Use two reviewers per item with a third to adjudicate. An LLM grader can handle low-risk criteria after calibration against expert grades.
Do FDA or EMA require a pharma AI benchmark?
No single rule requires one. The FDA and EMA guiding principles for AI in drug development, released January 14, 2026, call for risk-based assessment of AI performance and for data quality and governance across the AI lifecycle. A documented, held-out benchmark with item-level provenance is a practical way to show that assessment during audits and regulatory review.