I am enthusiastic about what AI can bring to healthcare market research. Faster delivery, less administration and more opportunities to explore our data are all welcome.
But enthusiasm needs judgement.
Some applications save time while strengthening the work. Others simply make it easier to produce something convincing without establishing whether it is true.
Here are the five uses I would be most concerned about.
1. Manufacturing evidence instead of collecting it
Synthetic respondents can be useful for exploring hypotheses, rehearsing questions or testing possible scenarios.
The problem begins when generated answers are presented as evidence of what real patients or healthcare professionals think.
A simulated patient has not lived with a condition. An invented quotation is not testimony.
Any synthetic work needs clear labelling, appropriate validation and honest limits. It should never quietly become a substitute for the real experiences a project claims to investigate.
2. Replacing sensitive conversations without considering the consequences
There may be appropriate roles for automated interviewing. But choosing it purely because it costs less misses an important question: what does this conversation require?
A patient discussing a frightening diagnosis may need reassurance, flexibility and the opportunity to explain something unexpected.
An experienced moderator can recognise uncertainty, respond sensitively and decide when to pause or probe.
We should assess the topic, participants and support arrangements before deciding how much of that interaction to automate.
3. Outsourcing the final interpretation
AI can produce a fluent explanation of research findings.
Fluency, however, does not establish accuracy.
A summary may exaggerate agreement, overlook a minority perspective or imply that one factor caused another without sufficient evidence.
Being involved in fieldwork helps researchers challenge these conclusions. It does not remove the need to verify them.
Final interpretations and recommendations should receive expert review, with claims traced back to the underlying material and limitations made clear.
4. Uploading confidential information without checking the arrangements
Putting a transcript into an AI tool can feel like an administrative step. It is also a decision about where information goes and how it may be used.
Before uploading, researchers need to establish that the tool and proposed use are approved, with appropriate arrangements for access, retention and confidentiality.
Participant information, client agreements and permitted data uses all matter.
Convenience is a poor reason to skip those questions, particularly when handling sensitive health information.
5. Leaving safety and compliance to AI
AI may help flag potential adverse events, product complaints or other issues requiring attention.
It should not become the sole basis for deciding that nothing needs reporting.
Established procedures, trained people and clear accountability remain necessary, including arrangements for reviewing and escalating potential cases.
The same principle applies to consent and participant welfare: technology can support the process, while responsibility stays with people.
The common thread is delegation without sufficient scrutiny.
I want AI to help us do better research. That requires knowing its limits, checking its outputs and retaining responsibility for the work.