On Thursday 9th July 2026 I sat in the audience for the launch of Mind’s Mental Health in the Age of AI Commission. Here are some thoughts provoked by that evening.
On the face of it, “AI and mental health” sounds like it fits neatly around a panel and a commission. In practice, it spills out of health care into a messy overlap of consumer protection, online safety, public-service design, and the raw capacity of health services. The tidy categories of policy do not hold up particularly well against the lives people actually live.
After the panel I joined a breakout group focused on regulation. The conversation flowed well and we discussed the Online Safety Act and the boundary of responsibility between users and systems; where the MHRA’s role starts and stops; how to restrain actors who have no intention of behaving well, and what to do about those who say the right things while taking unacceptable risks.
And one conclusion seems hard to avoid: AI is already part of our mental health ecosystem, whether or not the tools being used describe themselves that way.
It fills the gaps left by waiting lists. Someone might use it to prepare for a GP appointment. A conversation about employment might be rooted in the ways their mental health is a work-limiting condition. Someone who doesn’t think they’re ill enough to ask for help, is too ashamed to ask, or has asked before and had a bad experience, might well be typing instead.
So arguments about whether people should use AI for mental health are largely academic. The reality has outpaced them: people are using it.
The harder question is what happens when that use takes place through tools and services that do not think of themselves as mental health services at all. Because “AI” isn’t one thing. It ranges from regulated clinical technologies and purpose built tools to general purpose chatbots and AI embedded in services designed for something else. Their risks are not identical, but people move between them far more easily and quickly than regulation does.
Mind’s Commission begins in that blurred reality, asking how AI is changing help-seeking and support pathways, how people can be kept safe, and what it means for inequality in access and experience.
When a jobs service becomes a mental health service
Over the last couple of years I’ve spent a lot of time thinking about the intersection between employment, labour markets, public services and AI-led interactions. That’s a world we might instinctively define as being about understanding skills and matching to vacancies, writing CVs and acing interviews, employer demand and local labour supply, and the practical coaching that helps someone progress and build a career.
And yes, it is all of those things. But work is rarely only work.
Losing work is more than a shock to household income. Being unable to find work is more than labour market friction. Insecure work reaches beyond productivity, and having to explain yourself repeatedly to services that do not join up is more than an administrative inconvenience.
Work carries status and routine, dignity and identity, hopes and sometimes despair.
Which means that the relationship between a jobs and careers service and a mental health intervention is closer than you might think.
When does a chatbot helping someone prepare for an interview become a tool helping them manage anxiety? When does a conversation about the gaps in a CV become a conversation about shame? When does “tell me about a time you overcame a challenge” open up grief, panic or the sense that your life has gone badly wrong? When does an assistant designed to help with employment become, in effect if not by intention, part of someone’s mental health support?
The MHRA already regulates some software and AI as medical devices. It is right that tools designed for health care should be assessed with health care seriousness, and that intended use should help determine whether a product falls inside that regime.
But intended use and actual use don’t always line up.
A DWP employability assistant, a DfE tutoring tool, or just a chat with Gemini or Claude or ChatGPT, may make no therapeutic claim and still form part of someone’s mental health support the moment they’re typing about anxiety, shame or despair. And unlike a conversation with a clinician, where people generally understand at least something about confidentiality and data protection, what they type into a chatbot may sit inside systems of data management, storage and processing they barely understand.
So we have a clearer regulatory route for products that present themselves as health care but people don’t only take their mental health to health care products.
They take it to whatever is nearest. Whatever is open. Whatever replies. Whatever does not require a phone call between 9am and 10am on a Tuesday. Whatever gives them the space to start typing without requiring them to speak.
So the label on the product cannot be the end of the regulatory question. What use was foreseeable? What did the product’s design encourage? What did the provider know was happening?
The status quo is not safe either
During the evening Mind shared a survey that found almost 1 in 5 respondents (18%) had used AI chatbots to support their mental health in the previous 12 months. Of those, 60% had used them instead of formal support like NHS talking therapies. Among that group, 31% said they preferred AI, 18% said they could not access formal support, and 11% said existing support did not meet their needs.
On one reading, that sounds like a striking endorsement of AI. But using it instead of formal support does not mean people imagine the two things are equivalent, or that they no longer value human care. In fact, 84% of all those surveyed said access to mental health support or advice from a human alongside AI was important.
So this isn’t about people choosing a toy over a therapist. It’s about a system in which access has become a luxury and trust a casualty. The fact some people perceive value in these tools does not tell us whether they are safe, effective or ultimately helping.
None of this justifies throwing AI at every gap. Some functions and claims should remain out of bounds for general purpose systems however fluent they become. But declaring a tool unsafe or inadequate does not make an alternative appear. Anyone arguing against its use still has to answer what happens to people for whom the current alternative is nothing.
We need more human support, better-funded services and help that arrives earlier: GPs with time; schools and workplaces capable of spotting and supporting distress before crisis; and local services that are not permanently on the edge of collapse. It’s good that Mind is combining its call for stronger safeguards and clearer guidance with lobbying for same-day, open-access mental health support.
But formal services are only part of what has to be built.
Life is hard to do alone, and people find themselves isolated and lonely for many reasons. Family ties become stretched or complicated. Friends move away. Neighbours remain strangers. Community spaces close. People live far from the places that formed them and the people who remember who they were before the crisis, the job loss or the bad year. Living somewhere does not necessarily mean being rooted there.
Sometimes what is missing is not a therapist. It is somebody nearby who knows enough of your story to notice that you are not yourself. Somebody you do not have to brief from the beginning. Somebody who can hear what you are saying because they also know the things that remain unspoken.
You cannot manufacture that kind of relationship on demand. But institutions rooted in a place can create the repeated contact, shared life and ordinary familiarity through which people do become known. Schools, clubs, workplaces and community organisations can all provide some of that, although none can guarantee that the person most in need will be noticed.
Inevitably, that leaves me thinking about one of our jobs as Christians. Our churches should be places where people become known, where distress is noticed, and where we make the time to remain present in all circumstances. The Milburn Review, and its focus on young people, prompted me to write about the need for strong branches and good shade: patient institutions, rooted in place, in which people decide that the lives around them are not somebody else’s responsibility.
That argument travels beyond young people, though the overlap is substantial: 20% of 16-24 year olds who are not in education, employment or training report a mental health condition. That’s roughly 200,000 young people. And so while I might be arguing for AI’s potential as a mechanism for good, I fundamentally believe that human community makes for helpful and hopeful interactions in ways no model can reproduce.
Those things take people committed for the long-haul, stable funding, trusted institutions, and, perhaps above all, time. But there is a lag, and therefore while we build them, people will still need something.
At some point during the evening the words “something is better than nothing” left my mouth, and I immediately checked myself because that could be a lazy position to hold. We cannot make that low level of aspiration the design principle or the decision-making rubric. Too often “here is something inadequate because the alternative is nothing at all” has become the standard of care we decide some people should be grateful to receive.
But ultimately is it wrong?
If there is no therapist available, no same-day service, no friend or family member able to recognise what’s happening, and a tool can help someone take useful steps without pretending to be more than it is, then the difference between that and nothing is real.
That is not the same as saying that whatever feels helpful is helping.
“Better than nothing” is not enough.
But neither is nothing.
More answers are a problem
Which makes it all the more important that what people find is safe. One of the panellists, Lily Shervington, gave me reason to question some of our previous work when she described how a tool that relieves distress in the moment might reinforce the thing somebody needs help to escape.
Some of our design thinking for the jobs and careers service was that a useful interaction should reduce uncertainty: answer the question, offer a manageable next step and create some momentum. We designed the AI Work Assistant to use interactive cues and ask questions that help people keep moving rather than leave them staring at an empty prompt box.
Listening to Lily made me wonder whether we had designed too confidently for momentum. For someone caught in a reassurance-seeking cycle, another answer is not necessarily progress. The problem is not always that AI refuses to answer, it might be that it never stops: another question, another answer, another question, on repeat.
A system can sound patient and kind, and can be designed thoughtfully around the needs of many people, while still being wrong for a particular person or pattern of distress. If AI-assisted delivery is really going to help us move beyond the imagined average user, it will have to recognise differences like this rather than merely personalise the next answer. A useful coaching tool may need boundaries much like a useful human coach does. It may need to say “I notice we have been here several times. I do not think answering again will help.”
A small study published in 2025, based on scripted interactions with 2024 versions of general-purpose chatbots, found more affirmation and reassurance from the machines than from the therapists. It was small – six chatbot logs and 17 therapists – and should not be read as a verdict on 2026’s models. But it highlights a real design risk: fluent, friendly responsiveness can look therapeutic while lacking the judgement and context on which therapy depends.
I am not arguing that an LLM could or should replace a therapist. But I think we should be interested in how we might pair this powerful technology with our talents for good service design to help people prepare, reflect, practise or reach the right support, alongside or after therapy, and where therapy is unavailable.
We won’t get there by designing a fake person or pretending that all distress is the same shape. We have to work out what “helpful” means in different circumstances: when to answer, when to stop, and when the next step should be online or, perhaps more hopefully, in person.
Digital where possible, human when needed
Which is a neat segue into some more of the thinking we did for the jobs and careers service. Rather than using “digital by default” or “digital first” we coined the phrase “digital where possible, human when needed”. It is far more than a channel strategy.
A good service (digital or otherwise) should help someone understand where they are and move forward. It should design the flow across different channels and, where needed, make the handoff to a person feel natural rather than like a failure of the digital bit.
I’ve written before about the seductive idea of government in your pocket: the GOV.UK App, the NHS app as the “doctor in your pocket”, the “jobcentre in your pocket”. A thing in your pocket can be convenient, but convenience is not the same as care. An app can be a doorway, but it cannot repair contradictory policy, confusing eligibility or hollowed-out local services.
AI makes the promise more seductive and the danger sharper. Used well, it could help public services make better use of everybody’s time, including that of the person seeking help.
A work coach might see someone every week or two, but perhaps for only ten minutes. Regularity is not the same as depth. There is a lot of weight on the first interaction, and then on each short appointment that follows: what has happened, what has changed, what is getting in the way, what needs doing next.
How does one person come to understand another in ten-minute instalments?
The appointment becomes overloaded because the relationship has to fit inside it. What if the conversation could breathe? What if it did not have to carry the whole weight of the relationship every time?
We began thinking differently about what a small interaction is for. Every interaction can be valuable: a question answered, a task completed, a preference expressed, a worry surfaced, a small bit of progress made. Each contact can help the service understand someone better, but that cannot be a one-way extraction of information.
Every contact counts. And it has to count both ways.
You get something useful. The service learns something that helps it support you better next time. People will not become more trusting of government because a service remembers more data about them. They may become more trusting if each conversation gives them value, respects the things they have already said and makes the next interaction better than the last.
A chatbot that starts from zero every time is not much of a relationship. What matters is a service that learns with you: one that understands what you are trying to do, what keeps getting in the way, what has changed and what matters to you; that can remember the right things, forget the right things, protect the right things and recognise when the next step should be a person. The OCD example makes that more demanding: learning with someone can also mean recognising when another answer may cause harm.
The alternative is the worst version of services: disconnected moments pretending to be a coherent whole.
“Digital where possible” can easily become “digital wherever it is cheaper”, while “human when needed” comes to mean “only after the system decides for you”. People must be able to ask for human help, leave the digital route, challenge an automated judgement and understand how a decision has been made. The aim is a coherent service in which a person can ask for human attention and the service can recognise when to offer it, not a cheaper channel that is difficult to escape.
Digital where possible. Human when needed. And contact that counts for the person, not only the system.
AI is not tobacco
After the session wrapped, a lively conversation over pizza turned to banning social media and comparisons between AI, smoking and alcohol. I agreed with the pushback. Tobacco harms when used as intended. AI can diminish and isolate, but it can also help someone find the words, prepare for a difficult conversation, or make sense of something that previously felt impossible. And yet its possible value does not remove its capacity to mislead.
I am wary of only talking about harm because my own career would not have happened in the same way without the Twitter of 15 years ago (thanks Liz). It connected me to people and possibilities I could not have reached through geography or hierarchy. Social media has since caused damage but it has also continued to create relationships, opportunities and careers. I am wary of frames that recognise harm but can no longer see value.
The distinctions matter for AI too. Helping someone prepare questions for a GP is not the same as seeking a diagnosis. Organising thoughts is not the same as asking a system to confirm a delusion. Drafting a difficult message is not the same as relying on a chatbot for crisis care.
Public policy therefore has to do more than restrict access. It has to help people judge where these tools are useful, where they are dangerous, and yes, those circumstances under which a use should simply be prohibited.
This cannot only be a UK conversation
The UK has several regimes touching this question, from online safety to medical devices. But AI stretches the territorial logic of all of them.
When I use a tool in Croydon, the underlying model may have been built in California, fine-tuned by one company, wrapped by another and distributed through an app store. The harm may be felt in Croydon while the infrastructure, product decisions and accountability sit elsewhere.
Accountability therefore has to reach through the stack to foundation-model providers and frontier labs, rather than stopping with the services facing people in Britain.
Then there is the equity problem.
A two-tier market is an obvious risk: safer, more capable and more controllable systems for people who can pay, and weaker or more extractive products for those who cannot. The people with the least access to formal support may also be those most likely to rely on whatever is free and immediate.
And that is before the question leaves wealthy countries.
If this is hard in Britain, what happens where the mental health workforce has never existed at the scale required? The WHO projects a shortage of 11 million workers across the health workforce by 2030, mostly in lower-income countries. It isn’t a mental health figure, but it describes the scarcity within which all provision operates.
A recent paper from my new colleagues at the Tony Blair Institute takes that question seriously. The Risk of Standing Still: Governing AI in Health Systems Under Pressure argues that risk has to be assessed against the reality of the health system somebody actually has, not an imagined alternative in which enough clinicians, money and infrastructure are already available.
In some settings, the relevant comparison is not between AI and a well-resourced clinical team. It may be between a community health worker with very little specialist support and one with a carefully governed tool to draw on. Sometimes it may be between a new tool and no care at all.
I find that argument uncomfortable, but I think it is necessary.
Precautionary instincts are absolutely justified when health and lives are at stake. But if caution only measures the risks of adoption, while preventable deterioration and unmet need remain the natural background, it is not really caution. It is a choice whose costs are hidden because they fall on people in lower income and underserved communities.
Here, “something is better than nothing” becomes a question about the capacity of an entire health system. Scarcity cannot excuse offering people systems we would consider unsafe for ourselves. Yet a safety argument that leaves someone entirely alone with needs no accessible service can meet has consequences of its own.
Could a carefully designed tool extend scarce expertise or help a community health worker do more? Could it help someone understand what they are experiencing, find language for it, or reach support that would otherwise remain out of view?
I think that has to be worth trying.
But only if local institutions have real power over what is built, how it is evaluated, where it is used and what happens when it causes harm. Frontier labs should be working under that local leadership, alongside clinicians, community health workers, people with lived experience and speakers of languages badly served by current models.
Safety has to mean more than preventing bad outcomes. It also has to ask what useful things we are trying to make possible.
Regulating systems that will not stand still
This is where the OECD bit of my brain starts muttering.
We already have a lot of the words.
Human-centred. Rights-respecting. Transparent. Accountable. Robust. Safe.
The OECD AI Principles were adopted in 2019 and updated in 2024. They already talk about human agency, uses outside intended purpose and foreseeable misuse; that’s remarkably close to the boundary problem we started out with. Those principles fed conversations at the G7 and G20 and in other multi-lateral fora.
So the problem is not that we are still looking for the right vocabulary. We have had years of responsible-AI principles, standards and guidance.
But principles are not regulation, and saying the right things has not produced the right outcomes.
How many times can we enshrine responsible, trustworthy, human-centred, rights-respecting AI before asking why that did not reliably produce responsibility, trustworthiness, human-centredness or respect for rights?
Was it too abstract and voluntary? Too far from procurement, product decisions and the people with authority to stop a release? Too weak on institutional ownership and enforcement? Did the principles travel into ethics boards and conference slides without travelling far enough into commercial incentives, roadmaps and decisions about whether something was ready to release?
That does not make the work pointless. It gave governments and companies a vocabulary, some useful methods and expectations they did not previously have.
But if the public still does not know what protections exist, companies can still ship risky products into intimate human contexts, and regulators are still working out where one of them ends and another picks up, then the missing step is the translation from principle into duty.
That means duties attached to organisations with the power to act; scrutiny that continues after release; serious incidents investigated rather than treated as unfortunate anecdotes; and the ability to restrict or withdraw products when their actual use diverges dangerously from their stated purpose.
It also has to work while the systems keep changing.
We risk judging today’s systems only by yesterday’s failures. A thing handled badly by a 2023 model may be handled better by a 2026 one. A system that once treated every user the same may introduce age-appropriate defaults or warnings about unhealthy patterns of use.
Improvement is not absolution, but the model, product and safeguards do keep changing.
A good example of that was the new feature on ChatGPT where I can set a trusted contact. It’s a way for me to nominate someone who might be notified, after human review, about a serious self-harm concern. Such a feature creates its own questions about consent, what information is revealed, what happens when the system is wrong and what “human review” actually involves.
And for what it’s worth I didn’t actually set it up when prompted, and haven’t gone back to do so either.
Perhaps that says something about the gap between a feature sounding sensible in theory and a person deciding whether to use it. But it also shows that an AI interaction can be designed to lead back towards a real relationship rather than trying to contain a crisis within the machine.
Improvement matters. But harm to people using current systems cannot simply be justified by the promise that future versions will be better.
Regulation therefore has to be capable of saying “this is safer than it was” and “this is still not safe enough” in the same breath.
A Commission worth contributing to
Which brings me back to Mind’s Commission. Its value is in bringing technology, health, safety, looking for support and inequality into the same conversation.
The Commission runs until July 2027, with its first call for evidence opening in August. People who have found AI useful, been harmed by it, or used it because nothing else was available should contribute. So should those seeing different parts of the problem in services, research, regulation, families and communities.
The worst version of this debate would be one captured only by enthusiasts or only by sceptics. Enthusiasts may underplay harm because the possibilities are so obvious to them. Sceptics may underplay usefulness because the harms are so obvious to them.
Their answers are likely to be more complicated than either camp wants.
AI’s role in mental health is not coming. It’s already here.
It’s not neatly contained in mental health products. It’s in the gaps between products and services. It’s in tools designed for other things. It’s in the messy space between formal support and no support.
The boundary is already blurred. The choice is whether to pretend otherwise or govern the reality in front of us.
The status quo is not safe.
AI is not safe either.
That’s the point.
We are not choosing between a risky technology and a safe world. We are deciding what to do in the messy, complicated world in between: where we need to rebuild the human support that is missing, where we need to make the tools safer, where we need to hold providers to account and where we need to keep learning as the technology, and the ways people use it, change.