MHR Labs: Burnout and chatbot psychosis

The effects of AI on mental health, by Neil Stenton

Concerns about the impact of generative AI on mental health have grown over the past few years, beginning with the advent of ChatGPT. We’ll take a look at a couple of different angles this month, firstly on how AI is being used both as a confidant and therapist and whether there are any positives in people using it in this way. We’ll then look into the growing problem of burnout for employees using AI in the workplace.

Will AI do more harm than good to the UK’s mental health?

There have been several articles and research papers released this month about a growing concern around how people are using AI chatbots both from a healthcare and ‘friendship’ perspective.

A number of serious mental health episodes have been attributed to excessive chatbot use, when the sycophantic nature of certain models has resulted in serious psychotic episodes in individuals with no previous mental health history. This article from The Guardian has examples of many such stories.

In contrast, I’d like to focus on a slightly different take from this thoughtful BMJ article, which asks a broader question about the pros and (mostly) cons of AI’s impact on the health of the UK. The article comes from the 63rd Maudsley Debate at King’s College London, ‘The Chatbot will see you now’, which is a good watch if you have time.

The article begins by listing some of the extreme scenarios and stories that have been reported around chatbot use and the corporate drive to monetise at any price. However, the second part questions why people are being driven to use these sites for advice in the first place, and what this says about our society and healthcare system.

The debate points to the ease and immediacy of access as key reasons people use AI as a therapeutic go-to. Even if you have a close network, reaching into your pocket for instant assistance is hard to resist. However, the sense of embarrassment and shame referred to in the debate are also crucial reasons people will use AI rather than professional services.

I think the article importantly shows that by treating AI as a bogeyman, we’re missing the point of what’s missing in people’s lives and society which AI is attempting to fill, be it close friend and family networks or good mental health support.

The research view

There were a number of articles published this month, but I’ve chosen to focus on only one. While this is partly to avoid paywall limitations for our readers, it’s also because the majority look to the same extreme cases mentioned in the Guardian article. There are clearly serious problems around these use cases that absolutely need guardrails and restrictions from the AI companies themselves.

I’ve also found that, with all chatbots, you face the constant danger of the echo chamber they can surround users with. Even the less sycophantic and frontier models do this, and will probably continue to do this to some extent. Even for general day-to-day use, it’s important to be aware of the model limitations and for the user to act accordingly and take everything it says with a pinch of salt, especially when validating your own suggestions.

“Quick, I need some advice!”

The second part of the BMJ article raises an interesting question around why people are using AI for advice.

If the main reason people access advice from a chatbot is convenience, then this reflects another recently reported trend: the supposed decline of self-help books because people are simply using AI instead. This article from Publishers Weekly claims there was a 26.3% decrease in sales in the first quarter of 2026.

Part of this I suspect is the nature of people’s reducing attention span (congrats if you’ve got this far!), but mainly the fact that self-help books have a habit of expanding a 2-page article into a 200-page book stuffed with filler. Why spend money on a book full of fluff when you can get a straight answer for free from an app on your phone?

But along with the deeper issues of using AI for answers on deeper and more complex problems, there were questions raised about people still falling through the gaps, which I think is an easy problem to overlook. The assumption is that smartphone, internet and AI access are ubiquitous and that clearly isn’t the case in lots of different pockets of society, be it down to age, poverty, accessibility and many other reasons. I think this relates not just to AI access but technology access in general and the problems this causes for those without the access. Think GP and mental health services, bank services, and lots, lots more.

What does this mean for an organisation?

There are a number of takeaways from this article that are important to all organisations, whether they officially use AI or not.

Firstly, there’s the lesson of how we use the chatbots and how we trust their output. They can give fabulous, time-saving content, but they can also lead us down rabbit holes and amplify our world view, which could be disastrously wrong. The wording of the prompts and questions are important, but keeping an open mind to the response and actually reading, understanding and critiquing it are vital. Most of all: beware of the echo chamber!

Secondly, there’s the ever-growing question around employee’s mental health and wellbeing. There are several reasons why people use AI for advice in personal and sensitive situations and as the article suggests, the advice it gives could be beneficial. But if the reason someone is asking these questions is because they have no one else to ask, then we must look at how to encourage people to talk to a trusted and qualified colleague rather than a machine.

‘Botsitting’ and burnout, by Chris Judd

Many of us have some expectation, especially in this new agentic world, of being able to set an AI running independently, leave it to its task and come back and review the outputs. Unfortunately, that is not how most AI work feels.

Instead, many follow a practice of “botsitting”, a term coined by Glean Work Institute to describe the largely untracked labour of making AI outputs usable. Their recent report revealed that workers spend over 6 hours a week on average monitoring, steering, and correcting the outputs of AI.  While the report showed 75% of workers felt more productive, the time looking after the AI significantly ate into productivity gains. What’s more, working in this fashion caused increased fatigue and higher levels of burnout among employees.

AI-assisted burnout

I have been trying to work AI into more of my day-to-day workflow and have found this to be a very frustrating experience, often leaving me feeling drained. Recently, I found this Evil Martian’s blog offering their perspective on what goes wrong when working with AI, and it really resonated with my experiences. While it focuses on the issue mainly from a coding perspective, and Ivan Chepurin and Travis Turner present a seemingly biased view against AI, I still think the points they discuss can apply quite broadly.

The article opens by comparing two ways of working:

In the first, a worker does their task directly without AI assistance. It is a slower process, but they have room to rethink approaches, follow the logic through properly, and gradually build understanding as they go. They finish the task satisfied, knowing exactly how everything works.

In the second, AI speeds up the initial output, but the worker spends much of that saved time reviewing, steering, and debugging. Work is completed quicker, but the effort is more compressed, and mentally intense. Feelings of dissatisfaction from thoughts that this first task was completed “too easily” lead them on to the next task right away. In the end, they have completed more work faster but are left feeling burnt out by the constant mental load.

Towards the end of the article, they discuss some of the sources for these feelings of burnout. First is the loss of context, where too much of the thinking is handed off to AI, and people end up disconnected from the task itself and so become less confident in their own judgement. Second is the loss of quieter, slower thinking time. Constant prompting and reacting leaves less space for the kind of reflection that often leads to better decisions. They also point to false expectations, where an early burst of speed starts to feel normal and anything less begins to feel like underperformance. Finally, because AI makes it so easy to try many different variations, the work can expand far beyond the point where it should probably have stopped. Over time, this builds up as stress and frustration, ultimately leading to burnout.

The research view

Many of the problems discussed above come from not following best practices. Ideally, tasks and projects should be fully understood and scoped out before handing them off to AI agents. It also likely means you’ll use fewer tokens (see last month’s blog). Unfortunately, the speed and ease of getting results quickly naturally draws you into bad habits like involving AI before you are ready.

Some of these issues will reduce as AI improves and can take on larger tasks independently. As it stands now however, we need to build in time to do this planning and not get pulled into the hype and expectations that any work can be done ten times faster.

From a more top-down business perspective, the lesson here is that not all AI time savings are real. If faster output is contending with more time spent adjusting and reviewing, as well as having a greater mental overhead, then any gains may be smaller than they appear. It also means accepting that some work is still better left as human-led, especially where judgement, context, and accountability matter most. If AI is changing the shape of work, people need clearer expectations, better support, and enough room to say when, where, and how we best use these tools.