Medical publication policies and guidelines offer a framework for best practices, but there may be situations when more than one approach seems reasonable. The primary purpose of “What Would You Do?” is to explore examples of such situations. With the limited information provided to interpret the scenarios, you may find yourself agreeing with one, more than one, or none of the proposed actions. And that’s the point ‒ you should debate, contemplate, and communicate (with a comment) before selecting your “best” answer.
Now let’s find out how you responded and read through some commentary to the below scenario.
You are a writer at a medical communications agency supporting development of the first draft of a manuscript reporting primary results from a phase 3 clinical trial. At your client’s request, you are using the pharmaceutical company’s internally developed generative AI tool as part of the drafting process.
Initially, the tool performs well, but as you develop the manuscript, you notice that it occasionally misinterprets study findings, overstates the clinical relevance of certain results, and inconsistently summarizes safety data. These issues can be corrected with careful human review, but doing so substantially reduces the expected efficiency gains. The publication timeline is tied to an upcoming scientific congress, and any significant delay could affect the planned communication strategy.
Your client explains that the AI platform has been approved for internal use and is expected to be used on all publication projects. Reverting entirely to manual drafting would likely affect project timelines and budget. The authors have not yet reviewed the draft and are unaware of the issues you have identified.
What Would You Do?
The latest poll posed a realistic and increasingly common dilemma facing medical communications professionals: what should a writer do when a client-mandated AI tool introduces substantive scientific inaccuracies that require extensive human correction to maintain publication quality? Responses revealed a profession that is largely aligned on the importance of quality and transparency but divided on the best path forward when those principles must be balanced against operational realities.
Among the 57 respondents, two response options emerged as clear front-runners and were selected with nearly identical frequency. The most selected response was to escalate concerns through both agency leadership and the client’s publication or compliance teams (42%; 24/57). Close behind was the option to continue using the AI tool while implementing enhanced human review, documenting errors, and recommending future improvements (40%; 23/57).
Escalation and Governance
Respondents favoring escalation emphasized that the client should be made aware of problems with its AI tool rather than treating individual errors solely as a manuscript-level quality-control issue. One respondent described enhanced review alone as a “band-aid solution.”
Comments also pointed to wider implications. If correcting AI-generated content substantially increases effort, agencies and clients may need to consider effects on resources, budgets, timelines, and other publications using the same tool. This perspective reflects a concern that AI-related quality issues are not simply project-level execution challenges but potential governance issues affecting scientific integrity, compliance, operational planning, and risk management across programs.
Managing the Immediate Publication
Respondents favoring enhanced review focused on delivering an accurate manuscript within the congress-driven timeline while still meeting the client’s expectation that the AI tool be used. Several stressed that publication quality remained the primary objective.
Documentation was also a recurring theme. Recording AI-generated errors and the effort required to correct them could provide evidence of performance gaps, support discussions with the client, and inform improvements to the tool.
A Hybrid Approach
The comments suggest that the choices were not necessarily viewed as mutually exclusive. Several respondents who selected enhanced review said they would also escalate concerns, while others favored combinations of review, documentation, escalation, or possible suspension of the tool.
The practical response, therefore, may involve parallel actions: protect the quality of the current manuscript, communicate emerging risks, and gather evidence to inform future decisions about AI use.
The Quality–Efficiency Tradeoff
Respondents also highlighted a fundamental tension in AI adoption. They noted an apparent conflict between two legitimate objectives: producing a high-quality, scientifically accurate manuscript delivered on time, and achieving the efficiency gains that motivated adoption of the bespoke AI tool in the first place.
If substantial human review is required to identify and correct AI-generated errors, those expected efficiencies may diminish. Yet publication teams may still face fixed timelines and organizational expectations to use AI, creating additional complexity in balancing publication quality with operational expectations.
Takeaway
Overall, the poll results point to several shared priorities: maintaining scientific accuracy, communicating meaningful AI performance concerns, and ensuring appropriate human oversight. Where respondents differed was less on these underlying principles than on the emphasis and sequence of actions needed to address the problem.
The near-even split between escalation and enhanced review reflects the challenge of balancing immediate publication needs with broader governance concerns. Importantly, the open-text comments suggest that these approaches are not mutually exclusive: protecting the integrity of the current manuscript may need to occur alongside escalation, documentation, and evaluation of the tool’s performance.
As AI becomes increasingly integrated into medical communications workflows, organizations will likely need frameworks that support both objectives—maintaining the quality and integrity of scientific publications while ensuring that AI tools are appropriately evaluated, monitored, and improved.
This article was prepared by members of the ISMPP Insider Committee using a combination of human drafting, review, and oversight and AI-assisted revisioning. The opinions expressed within are the authors’ own and do not necessarily reflect the views of their employers or of ISMPP.
