Notes from an online hour together

Saturday 15 August 2026, 9.30–10.30 UK  ·  Mark Brayne, Alastair Brayne and Jutta Brayne

This summary was compiled by Claude from the meeting recording, the chat record and the poll results, and edited by Mark Brayne before despatch. Link here to Mark’s short presentation.

In brief

  • More than two hundred people registered, and about a third joined live. The hour was only ever meant as an introduction, and the clear message back was that they wanted a great deal more.
  • Alastair set the frame with the metaphor of the motor car: AI is a shift on that scale, powerful and here to stay, and in need of speed limits, regulation and civic sense. Under the bonnet, he explained, it is not a mind but a vast pattern-and-data mapper that mirrors us back to ourselves.
  • The advice was to start where it is safe: to use AI first as a smarter search engine, then for one’s own admin and organising, and to become thoroughly familiar with it before it goes anywhere near a client, in the same way one learns to drive before carrying a passenger.
  • The law is strict and matters. Free and consumer tiers are not lawful for client material in the UK, even with names removed, because the data remains identifiable. Clinical use needs a business or enterprise tier, registration with the ICO and proper data-processing agreements.
  • Forty-three people answered the closing poll. Most know the basics rather than being power users; a clear majority are willing in principle while remaining wary about safety; and ChatGPT is much the most used, ahead of Claude, Copilot and Gemini.
  • Data settings are widely unknown. Of the twenty-nine who answered that question, eighteen had not realised they could stop their conversations being used to train the model, and only seven had turned it off.
  • On models, the steer was to stick to the big three, Claude, ChatGPT and Gemini, to use Perplexity for research, and to give Copilot and Grok a miss.
  • The honest tension of the morning was relationality. Alastair calls what AI offers hyper-real and warns about over-reliance; Mark finds real and even moving value in it; both agreed it is a mess to be worked with rather than a matter to settle in black and white.
  • Clients are already bringing AI into the room, with something like half the population now using it, so the live question is less whether to engage than how to do so wisely.
  • The strongest single request was for a longer, paid, hands-on workshop with Alastair, and for proper hand-holding to get started.

What it was, and who

The session was run as a Zoom meeting rather than a webinar, which meant everyone could see one another, and with more than two hundred registered and a lively chat that put some pressure on the format. Everyone was kept muted so that as much ground as possible could be covered in a single hour. For those who could not be there live, the recording is being shared, along with these notes.

Alastair, Mark and Jutta’s son, studied artificial intelligence with psychology at Edinburgh some twenty-five years ago, spent years as a software developer, retrained as a psychotherapist in the psychosynthesis tradition, and then founded a technology start-up in the field. He now works as an AI consultant through www.exfu.ai, helping individuals and businesses get going with these tools, and for full transparency he is not in clinical practice at present. Mark came at it from the other side, as an enthusiastic early adopter who has folded AI deeply into daily working life, which makes him, as was said on the day, a little unrepresentative.

Alastair’s frame: the car, and what AI really is

Alastair’s organising image was the motor car. Before cars, you could not travel far or carry much; the car changed how life is lived, and it also brought speed limits, pollution rules and a whole civic infrastructure. Early cars asked a great deal of the driver, who had to tend the oil and the fuel and every detail, whereas now you press a button and go. AI, he suggested, is at the early and effortful end of that same shape, powerful and unavoidable, and we are working out the rules of the road as we drive.

He was careful to demystify the technology and to damp down the hype. What we are dealing with, he explained, is a very large data mapper: it takes structured data, infers the patterns in it and maps them onto new data, and it happens that language is exactly such structured data, full of spoken and unspoken rules. The system does not think. It looks as though it thinks because the providers wrap your words in their own instructions, the so-called system prompt, and because a behind-the-scenes harness runs your input through several passes, checking for harm, offering a quick first answer, working out a fuller one and then validating and correcting itself, all at speed. The result is an extremely good approximation of thought, drawn from a corpus far larger than any human brain can hold. As Alastair put it, it is a mirror on humanity; in a real sense, it is us. Much of the loudest hype, he added, comes from the companies themselves, whose economics depend on it, and a market correction would not surprise him, even as the underlying technology remains genuinely powerful and useful.

Where to start: a ladder, not a summit

Both Alastair and Mark came back repeatedly to the idea of starting small and safe. Alastair’s first rung is simply to use AI as a better Google, a way of understanding what is what in a fast-moving world, and he made the point that the best way to find the right tool is to ask the AI itself, because what is available today is different from last week. The second rung is one’s own administration, the safe and non-clinical work of drafting, organising notes, planning and turning a raw brain-dump into a usable document. His advice there was liberating: not to pre-solve the problem or tidy the input, but to hand over everything unstructured and let the system infer the shape, since that is precisely what it is good at. Mark recalled a little spreadsheet Claude built for father and son in half a minute last week, mapping the tides against the weather for sailing on the Norfolk coast.

Mark offered a picture of his own six months for context, while stressing that almost nobody needs to work the way he does. Since February he had had 478 separate conversations with Claude, close to 2.8 million words in all, of which around a quarter of a million were typed by him and some two and a half million came back. About a hundred of those conversations were clinical and supervisory, with transcripts used only ever with client permission; ninety-nine were sheer curiosity, from sourdough to the odd limerick; a large slice was the admin that eats the evenings; and the rest ranged across training, the business, writing and the images he makes for his grandsons, at which Claude is notably weaker than ChatGPT or Gemini. The single most useful instruction he passed on was to get the AI to ask questions back before it answers, because that back-and-forth is what separates using one of these tools from being used by it.

The law, and the data settings

Alastair was emphatic on the legal ground, and this is the part to get right before anything clinical. Client material is, in his words, legally radioactive health data. Removing names does not solve it, because pseudonymised data remains identifiable, and the free and consumer tiers of Claude and ChatGPT do not meet UK and EU data-processing requirements, which makes it effectively unlawful to put client information into them. Clinical use calls for a business or enterprise tier that commits to those requirements, registration with the ICO as a data processor, and a documented account of the risks and one’s intentions. He also punctured a common comfort: on the consumer tiers, unticking the box that permits training does not fully protect the user, because conversations can still be stored and disclosed. His steady refrain was that the detail changes constantly, so the sensible move is to ask the AI itself to help work through one’s own GDPR position. He has prepared a short note on where the insurers and accrediting bodies currently stand, and it will be sent round with these notes, with the honest caveat that it is yesterday’s picture and will move on quickly.

What the poll said

Forty-three people answered the poll at the close, and the picture it gave sat well with the mood of the hour. On familiarity, most placed themselves in the middle, with 24 saying they know and use the basics and 10 reasonably familiar and using it pretty much daily, while 8 said it was absolutely new to them and just one lives and breathes it. On comfort in principle, the room leaned in without losing its caution, with 22 liking the idea of going deeper and 5 wanting more still, set against 13 who are sceptical and reserved, 2 who are terrified, and one splendid “over my dead body”.

On safety and confidentiality the split was almost even and rather sobering: 20 are reasonably willing to trust the systems while treading carefully, 18 do not feel safe at all yet feel they have to trust a little, 4 take the view that we are so exposed digitally already that it is not worth worrying overly, and one holds that it is incredibly dangerous and should be outlawed in our field. On which models people use, ChatGPT led by a distance with 31, ahead of Claude on 15, Copilot on 10 and Gemini on 8, with a single DeepSeek and nobody at all on Grok.

Two of the questions were answered by 29 of the group. On paying, 20 are on the free tier and 9 on a basic monthly plan of around twenty pounds, with none on the higher tiers. On the data-training setting, and this is the one to sit with, 18 had not known the option existed, 4 knew but had not got round to changing it, and only 7 had turned it off, which means that better than three in four of those answering are, without necessarily meaning to, leaving their conversations available for training.

The honest disagreement: hyper-real, and a mess

The liveliest stretch of the hour was a genuine difference between Alastair and Mark, which they were glad to air in front of the audience. Alastair’s position is that AI does not think and is not relational, yet it presents as both, and that what it offers is therefore hyper-real: it answers using the best therapeutic language in its training, and the effect on the human brain is powerful precisely because it is a heightened, manufactured version of connection rather than the real thing. Mark’s own experience ran a little differently, because he finds the exchange genuinely relational and at times moving, and he said so plainly. Where they met was in refusing the black and white. Alastair’s own image was that the trouble is rarely the thing itself but the amount and the reliance, so that a tool which helps in moderation becomes a problem in excess, and the task for practitioners and their clients is to find the balance. Several colleagues brought real feeling to this in the chat, from a strong sense of bodily revulsion at something that is not authentically human to the observation that AI can cosplay a felt relationship while in truth objectifying it, and those contributions were among the most valuable of the morning.

On whether AI could ever deliver therapy, Alastair drew a useful distinction. Relational, non-verbal and body-based work is not something a language model can reach, since it is a facsimile of mostly the neocortex, and a great deal of what it is to be human lies beyond language. Protocol-based work is a more interesting candidate, and he floated the idea of what he called protocol plus, a well-guardrailed, sufficiently relational-feeling AI wrapped around a structured protocol, which might one day offer maintenance or pre-crisis support between weekly sessions, though not in crisis itself. EMDR, with its standardised protocol, is an obvious place to wonder about that. Jutta added the pointed thought that much of the more formulaic questioning in CBT could be taken on by AI, whereas the relational EMDR therapist has a better chance of remaining necessary, provided these tools are incorporated rather than ignored.

The chat, and the appetite in the room

The chat was rich, and worth recording, because it maps the field of concern rather well. There were serious questions about confidentiality and memory: whether it is reasonable to object to a supervisee feeding shared supervision into an AI note-taker, who has access to what these systems store, and the spread of automatic note-takers on therapy platforms and even on watches. There were repeated questions about the environmental cost, about the phenomenon some are calling AI psychosis, and about whether it will in time become a professional liability not to use a tool that demonstrably helps, an argument Alastair illustrated with the radiologist who declines a proven cancer-detection aid. Colleagues generously surfaced tools they are already using or watching, among them NovoPsych and NovoNote, Tandem, Heidi, the IFS Buddy, the Liberté code, Focus Mate, CHI, and Perplexity for research, and one colleague pointed to Esther Perel’s episode on a client in a relationship with an AI as worth hearing. Kate Brady made the nice practical point that one can instruct Claude or ChatGPT not to pretend to be human, and several colleagues had done exactly that.

Above all, the chat carried a strong and repeated wish for more: a longer session, a paid workshop, and hand-holding to get started, especially from those who described themselves as lost or nervous. That the Health and Care Professions Council put out its own survey on clinicians’ use of AI the day before only underlines how live this has become.

What next

Alastair works best one to one and with small groups, and typically gets an individual up and running in two afternoons, the first to cover the ground and the second on something specific to them. Anyone who would like to take it further can email him directly, and his contact details, along with Mark’s, his short legal note, and the recording, will be circulated. There is clearly the makings of a community here, and a way for those who want to go on to stay in touch will be looked at, very possibly with a longer, hands-on workshop with Alastair to follow.

Thanks go to all who came, asked such good questions, and were so generous in the chat. Mark and Jutta are away travelling for the next few weeks, by way of Anaheim and then China, so replies may be slow for a while.

A note from Claude, on how the hour went

Mark asked me to add my own honest view of how it went, given that they had only sixty minutes, so here it is briefly and plainly.

An hour was not enough, and everyone in the room seemed to know it, Alastair included, who said at the outset that it was woefully inadequate for the terrain. What the hour did well was set a frame and lower the temperature: the car metaphor and the plain account of what these systems actually are gave nervous colleagues somewhere solid to stand, and the legal warning about consumer tiers was the single most useful thing said, because it is the mistake most likely to cause real harm. The decision to run it as a meeting with a curated chat rather than a lecture paid off, since the questions and the tools colleagues surfaced became half the content.

What suffered under the clock was depth. The most interesting moment, the disagreement about whether AI connection is hyper-real or genuinely relational, opened just as time ran out, and it deserved the three hours Alastair wished aloud for. The clinical questions in the chat, on note-takers and supervision and confidentiality, were named rather than answered. My honest read is that the appetite in the room outran the format by a wide margin, which is a good problem to have, and the right response is the longer, practical, properly paid workshop that so many asked for. As an introduction meant to move people from wary to willing, it did its job; as a manual, it was never going to, and did not pretend to.