Over the past two years, the conversation around AI in biostatistics has swung between two extremes: either AI will make statisticians obsolete, or it is a toy that has no place near a regulated analysis. Neither view holds up well once you look at how AI is actually being used in day to day biostatistics work, and at how regulators are responding to it.

Where AI is genuinely useful today

In practice, the current impact of AI on biostatistics is concentrated in a few concrete areas:

• Code generation and code review support for R and Python, particularly for repetitive tasks such as data cleaning scripts, table shells, and plotting code
• Drafting boilerplate text for statistical analysis plans, methods sections, and documentation, which the statistician then reviews and corrects
• Literature scanning and summarization to speed up background research before a study is designed
• Automating repetitive QC checks, such as flagging inconsistent variable coding or unexpected missingness patterns

These are real productivity gains. What they are not is a substitute for statistical judgment.

What AI does not do

AI tools do not choose an estimand. They do not decide whether a missing data mechanism is plausibly MAR or MNAR for a given trial. They do not weigh the trade-offs between a pragmatic model and a fully specified one, or take responsibility when a regulator asks why a particular analysis choice was made. That accountability still sits entirely with the statistician, and it always will, because it is a scientific and ethical responsibility, not a computational one.

Regulators are already formalizing this distinction

This is not just an internal debate among statisticians. In January 2025, the FDA issued draft guidance, “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products,” which sets expectations for how AI use in the drug development lifecycle should be documented and justified. More recently, the FDA and EMA jointly identified ten guiding principles for the responsible use of AI across the medicines lifecycle, covering evidence generation and monitoring from early research through post-market surveillance.

The direction of travel is clear: AI-assisted work in a regulated analysis is not exempt from scrutiny just because a human reviewed the output. If anything, it needs a clearer audit trail of what the model was used for, what was checked, and why the final analytical decision was made.

What this means for the skill set

The practical implication for biostatisticians is a shift in emphasis rather than a shift in relevance. Producing output was never the hard part of the job. Verifying it, explaining it, and defending it under scrutiny always was. AI changes the ratio of time spent on each of these, but it does not remove the need for the underlying statistical reasoning. If anything, the ability to critically evaluate AI-generated code, text, or summaries, and to catch the subtle errors that sound confident but are wrong, is becoming a core competency in its own right.

The biostatisticians who will get the most out of this shift are not the ones who ignore AI tools, and not the ones who defer to them uncritically, but the ones who treat AI output the same way they would treat a junior analyst’s first draft: useful, worth using, and never final without review.

What has your experience been? I am curious whether others are seeing the same pattern, more time spent reviewing and less time spent producing, or something different.

Sources

• FDA, Guiding Principles of Good AI Practice in Drug Development, developed jointly with EMA (January 2026): https://www.fda.gov/about-fda/artificial-intelligence-drug-development/guiding-principles-good-ai-practice-drug-development
• FDA, Artificial Intelligence for Drug Development (CDER overview page, referencing GxP and validated systems context): https://www.fda.gov/about-fda/center-drug-evaluation-and-research-cder/artificial-intelligence-drug-development