MHRA’s Inspectorate published a blog post on 29 June 2026 titled, with a candor regulators don’t always show, “Use of AI for GXP inspection responses: setting standards without stifling innovation.” It’s worth reading in full, and it’s worth taking seriously even though it’s a blog post rather than formal guidance, because it does something more useful than a ban would have: it names, specifically, what’s already gone wrong at sites using these tools, and sets a standard that has nothing to do with which tool was used.
What MHRA is actually saying
The core position is straightforward. MHRA isn’t prohibiting AI assistance in drafting inspection responses, and it isn’t singling AI out as uniquely risky compared to other drafting methods. Its stated standard applies identically regardless of whether a response was written by AI, pulled from a template, drafted by an external consultant, or written manually: every submission has to be factually accurate and verifiable, technically reviewed by people with genuine, appropriate experience, and signed off by someone who holds actual authority and accountability for what’s being submitted. Claims need evidence behind them, and the response needs to be appropriate to the specific regulatory context it’s responding to, not generic.
That’s a deliberately tool-agnostic standard, and the choice to frame it that way is itself informative. MHRA is signaling that the accountability question doesn’t change based on how a document was produced. What changes is how easy it’s become to produce a document that looks adequate without anyone having done the underlying work, and that’s exactly the pattern the post goes on to describe.
The failures MHRA is already seeing
This is the part of the post that should get more attention than the general accountability framing, because it’s concrete, and because it describes exactly the risk this argument has been making about AI-assisted GMP documentation generally: fluent output substituting for verified substance.
MHRA describes responses citing MHRA guidance documents that don’t actually exist, fabricated references presented with the same confident tone as real ones. It describes responses citing regulatory frameworks that don’t apply to the situation being addressed, misapplied with enough surface plausibility that the error isn’t obvious on a quick read. It describes responses to serious deficiencies that, in its own words, appear designed to mislead, language a regulator doesn’t use casually. And it flags excessive verbosity that fails to actually address the underlying issue, illustrated with a specific example: a 90-page response that took inspectors more than 20 hours to review, against a typical review time of around four hours for a properly scoped response.
None of these are AI failures in the narrow technical sense. They’re accountability failures that AI has made faster and cheaper to produce at scale. A fabricated citation is a fabricated citation whether a person or a tool generated it. What’s changed is how quickly a plausible-sounding, ultimately unverified response can now be assembled, and how much that shifts the burden onto whoever is supposed to be checking it before it goes out the door.
The voluntary disclosure option, and why it matters even though it isn’t mandatory
MHRA’s post also raises the option of voluntarily disclosing AI involvement in a response, a brief statement identifying which sections used AI assistance and confirming that a human verified the output. It’s explicitly optional, not a requirement, and it’s worth thinking about why a regulator would offer that option without mandating it.
An optional disclosure functions less as a compliance checkbox and more as a trust signal. A company willing to state plainly which parts of a response were AI-assisted, and to stand behind having verified them, is making a different kind of claim than a company that says nothing either way. It’s not that silence is itself suspicious. It’s that voluntary transparency, offered without being required, tends to correlate with the kind of internal discipline this whole post is actually asking sites to have: knowing what was checked, by whom, and being willing to say so.
Where this sits relative to other AI-in-GMP guidance
Worth being precise about scope here: this MHRA post addresses inspection responses specifically, not AI use inside day-to-day GMP operations generally, and it doesn’t reference EU or PIC/S draft Annex 22, the text most directly addressing AI within GMP operations more broadly, which remains in draft form as of this writing. The two are related in spirit, both circling the same underlying question of how accountability attaches to AI-assisted output in a regulated environment, but they’re separate documents addressing separate scopes, and it would be inaccurate to describe one as building on the other.
What connects them is the pattern, not a formal linkage: regulators are responding to real, observed failures faster than they’re able to finalize comprehensive guidance, which means the practical standard sites are actually being held to right now is closer to “can you defend what you submitted” than any specific finalized rule about AI tools themselves.
What this means for how a response actually gets built
A few practical implications follow directly from what MHRA has flagged, regardless of whether your own site uses AI tools in drafting inspection responses today.
Every factual claim and every citation in a response needs to be independently verified against a primary source before submission, not accepted because it reads fluently or matches the general shape of what a correct citation looks like. This was always good practice. It’s now the specific failure mode a regulator has publicly named.
Length isn’t a proxy for thoroughness, and MHRA’s own example makes the cost of that confusion explicit: a 90-page response consumed five times the normal review effort without necessarily addressing the deficiency any more effectively than a properly scoped response would have. A response that’s long because it’s padded, rather than long because the deficiency genuinely requires that much substance, reads to an inspector as evasive regardless of intent.
Whoever signs the response needs to be able to answer, specifically, what they verified and how, not just that they reviewed the document generally. That’s the same accountability standard this argument has made about AI-assisted investigations internally, applied now to the external-facing document a site sends back to a regulator.
Conclusion
MHRA’s post is a useful early marker precisely because it doesn’t try to regulate the tool. It names the failure pattern the tool has made easier to produce, and it sets a standard, verifiable accuracy, genuine technical review, real accountability at sign-off, that any site can measure itself against today, without waiting for AI-specific rules to catch up to the practice already happening in the field.
If you’re not confident your own inspection response process could survive the kind of scrutiny MHRA is describing here, that’s worth checking directly, and it’s exactly the kind of gap a Rapid Diagnostic is built to find.
Key Takeaways
MHRA set a standard, not a ban. Factual accuracy, technical review by qualified people, and accountable sign-off apply regardless of whether AI, a template, a consultant, or manual drafting produced the response.
The failures named are specific and real. Fabricated guidance citations, misapplied regulatory frameworks, and a 90-page response that took inspectors five times the normal review effort.
Voluntary disclosure functions as a trust signal, not a requirement. A site willing to state which sections used AI assistance is also signaling the internal discipline behind that statement.
This addresses inspection responses specifically, not GMP operations broadly. It doesn’t reference draft Annex 22, and the two shouldn’t be described as connected beyond sharing the same underlying accountability question.
Length is not a substitute for substance. A padded response reads as evasive to an inspector, regardless of whether the padding was AI-generated or not.