An article published in the Sunday Times on 1 November 2026, AI 'erased evidence on natural birth harm', raises issues that those of us working with AI agents have known about for a while.
What happened
In June, the national review of maternity care in England, led by Baroness Amos, concluded that “normal birth ideology” was not widespread in NHS maternity services. Dr Bill Kirkup, a former adviser who left the review team, believes that this conclusion was skewed by how the evidence was analysed by an AI programme, rather than reflecting a lack of evidence.
More than 10,500 submissions, including transcripts of interviews and panel discussions, testimonies and reports, were fed into Microsoft Copilot. NHS England confirmed that it was asked to identify themes and produce summaries. Kirkup says the expert advisers never saw the original transcripts or written evidence. Their work was based entirely on the AI’s output.
In a debate on the issue, former Health Secretary Sir Jeremy Hunt described the allegations as deeply troubling, "...evidence from families may have been missed because the AI prompts used to analyse it did not look for it".
Why this can happen
AI tools are very good at producing confident, readable summaries. That’s what makes them useful, and it’s also the risk. A well-written summary looks complete, even when it isn’t.
Depending on how the tool is set up, asking an AI to “identify themes” can be a misleading question. It may assess the importance of content by how often words and phrases appear, rather than by what people have actually said. Unless the tool is designed to search by meaning (semantic search), and given the different ways people might describe the same (in this case, traumatic) experience, those accounts may never be connected to the issue the review was examining.
Furthermore, depending on the version of Copilot used and how the files were provided, may have contributed to its failure to retrieve and process every submission in full.
In my experience, the prompt alone is unlikely to be the only cause. How the tool was set up, what content it was given, what it was asked to look for, and whether anyone checked its output against the original evidence, all played a part.
So who is responsible: the AI or the people using it?
The prompt didn’t decide to rely on the summaries. People did. And from what has been reported, no one compared what the AI produced with what mothers actually said in their original submissions.
How to avoid this
Prepare the content
Make sure the agent can read everything. Format and organise your content so it can be found and processed, and confirm that every document has been reviewed and understood. A large body of content, particularly if it includes interview transcripts, may be hard for the agent to interpret given the difficulty for interviewees to describe sensitive situations.
Design the agent
Build an agent, not just a prompt. Give the agent specific instructions and robust guardrails on what to extract and how to analyse it, including the different words people might use to describe the same issue (synonyms), including what to do with evidence that doesn't align with the main themes.
Ask neutral questions. Check that your instructions aren’t leading the agent towards a particular answer, and that it covers the issues you need to explore in an objective, unbiased manner.
Test before you scale
Treat the agent like any Agile software product. Build it iteratively, one step at a time. Agree acceptance criteria up front, so you know what a good result looks like before you see the output. Start with a small sample, and only scale when the agent consistently meets those criteria. Whenever you add new content or change the instructions, run regression tests to check that what worked before, still works.
Ask the agent to show its workings. Every theme should be traceable to specific submissions, so you can see what has been included and what has been left out.
Keep a human in the loop
Read some of the original evidence yourself and sense-check the conclusions. Being the human in the loop means checking the output against the source, not just reading the summary. Analyse a sample of your content and don't hand your judgment over to the agent. Challenge yourself on whether the conclusion reached by the agent is defensible. This is your user acceptance testing, and it can’t be delegated.
Conclusion
So, can an AI prompt be held responsible? No. A prompt is an instruction written by a person, run on a tool chosen by a person, producing output that a person decides to rely on. AI can analyse 10,000 submissions, but it can’t decide what matters in them. That judgement, and the accountability for it, stays with you.
Related Services