Sanctuary strategies

A defence built on what AI can't do has an expiry date

‘Sanctuary strategies’ is my name for the ways health professionals and educators protect their sense of value from AI, usually by asking what humans can do that AI can’t and sheltering in the answer. The grouping is my own reading of responses I keep meeting, with no study behind it, and I offer it because I think it helps us see what they have in common.

What are sanctuary strategies?

One-sentence definition: A sanctuary strategy is a response to AI that secures the value of human practice by pointing to something AI can’t do, or by keeping AI away from it, so that the claim lasts only as long as the limit does.

I started using the term in talks to clinicians in 2025, after noticing how often the same reassurance came back when I showed what AI can now do in clinical settings. Most of it was an answer to the question of what humans can do that AI can’t. In 13 of 15 studies comparing chatbots with health professionals, for example, raters judged the AI’s written responses more empathic (Howcroft et al., 2025). In the talks I described four responses, and going back through my notes since then I’d add two more:

• Denial: “It can’t replace the therapeutic relationship.” In some contexts, on some measures, it already has.
• Retreat: “Focus on what only humans can do.” The list gets shorter with each model release.
• Restriction: “Don’t let students or clinicians use it.” They will, because it’s useful.
• Resignation: “Our roles will inevitably diminish.” This sounds like realism and works like helplessness.
• Raising the bar: “It still makes mistakes.” True, and so do we, since diagnostic error is a well-documented source of harm (Graber, 2013). Holding AI to a standard the existing system has never met keeps it out without anyone having to argue that it can’t do the work.
• Jurisdiction: “The regulator will never allow it.” Registration, liability, and professional regulation do keep people in the loop, often long after AI could do the task, and so the shelter feels solid. It lasts as long as the rules do, and rules change as the evidence and patients’ expectations change.

What they have in common

All six share an assumption that human value lies wherever AI is absent, whether because it can’t do the work or because we won’t let it. That makes each of them a bet against the next model release, or against the next revision of policy, and the god of the gaps¹ shows how those bets tend to go: its territory shrinks every time science explains something new. The pull is easy to understand, since AI reaches the expertise, status, and autonomy that professional identity rests on (Jussupow et al., 2018). The cost, as I see it, is that our attention stays on ourselves and drifts away from the patient and the student.

Two kinds of retreat

Retreat is the strategy I meet most often, and it takes two forms that are harder to spot than “what only humans can do”.

The first is a retreat to what can’t be measured: the experienced nurse who senses a patient is going off despite normal observations, or the educator who reads the mood of a seminar group. These are real forms of expertise, and assessment has rarely captured them. The trouble is that “unmeasured” describes our instruments, and instruments change. Nurses’ unease already shows up in what they document, and a trial across 74 hospital units found that an AI reading patterns in that documentation lowered the risk of in-hospital death (Rossetti et al., 2025).

The second is a retreat to values: “we’re human because we care.” I think that’s true, and it does little work. Care at the level of one clinician and one patient is partial and unevenly given, and on its own it doesn’t tell a programme team what to teach or a service what to fund. It’s also where the evidence on perceived empathy lands hardest, because care is judged by how it’s received, and that’s where the AI responses in those studies were rated well.

Both forms point at something that matters, and both put it somewhere I expect AI, or our measures of practice, to reach.

What survives the next release

A speech and language therapy team might move its assessment onto reflective writing because AI can’t reflect on a student’s own placement, then find within a year that markers can’t tell AI-assisted reflections from the rest. Insisting the writing is still authentic, banning AI from the task, or giving up on assessing reflection all leave the more useful question unasked: what was the reflection for, and was a written piece ever the best evidence of it?

Some claims about human work don’t depend on AI’s limits. Ethan Mollick describes “Just Me Tasks” that we keep for ourselves for reasons of meaning or ethics, whatever AI can do. Accountability is a reason of the same kind: someone has to answer for a clinical decision regardless of who proposed it, and that holds when the AI is excellent. It becomes jurisdiction when the rule is used to keep AI out of work it could do well. The test I use is whether the argument would survive the next model release. In place of asking what humans can do that AI can’t, I’d ask what clinicians and AI can reach together; distributed cognition, context sovereignty, and the essay Taste and judgement develop that answer.

Sources

• Graber, M. L. (2013). The incidence of diagnostic error in medicine. BMJ Quality & Safety, 22(Suppl 2), ii21–ii27. https://doi.org/10.1136/bmjqs-2012-001615
• Howcroft, A., et al. (2025). AI chatbots versus human healthcare professionals: A systematic review and meta-analysis of empathy in patient care. British Medical Bulletin, 156(1). https://doi.org/10.1093/bmb/ldaf017
• Jussupow, E., Heinzl, A., & Spohrer, K. (2018). I am; we are: Conceptualizing professional identity threats from information technology. In Proceedings of the International Conference on Information Systems (ICIS 2018), San Francisco.
• Rossetti, S. C., et al. (2025). Real-time surveillance system for patient deterioration: A pragmatic cluster-randomized controlled trial. Nature Medicine. https://doi.org/10.1038/s41591-025-03609-7