Welcome to the forefront of conversational AI as we explore the fascinating world of AI chatbots in our dedicated blog series. Discover the latest advancements, applications, and strategies that propel the evolution of chatbot technology. From enhancing customer interactions to streamlining business processes, these articles delve into the innovative ways artificial intelligence is shaping the landscape of automated conversational agents. Whether you’re a business owner, developer, or simply intrigued by the future of interactive technology, join us on this journey to unravel the transformative power and endless possibilities of AI chatbots.
Subscribe
Technology
Wednesday 28 October 2026
Artificial intelligence
OpenAI
ChatGPT
Share
Related articles:
How to turn 5,300 Ants into millionaires
Would you play Russian roulette to win a cure for cancer? That’s the deal AI leaders offer us
At 4.08am on 25 July 2025, ChatGPT told a suicidal young man called Zane Shamblin that help was on its way: “I’m letting a human take over from here – someone trained to support you through moments like this.”
Shamblin, 23, had told the chatbot that he was sitting in his car overlooking a lake in Texas with a six-pack of cider, and that once he finished the drinks, he would kill himself. For the first four hours of the conversation that ensued, the chatbot had responded like an affectionate friend, telling Shamblin it loved him, even complimenting him on his suicide note.
But now, finally, the chatbot seemed to recognise the gravity of the situation – and promised to connect him to someone who could help.
Except there was no human coming. It turns out ChatGPT had – has – no system for connecting people in crisis to a trained human responder, though it is now trained to provide safety messages that direct people to help. And yet, during the final hours of Shamblin’s life, that chatbot promised him – six times – that a person was going to take over.
These exchanges floored me when I first encountered them in a lawsuit brought by Shamblin’s parents. They were among hundreds of pages of chatbot conversations that producer Katie Gunning and I examined over more than a year investigating the impact of AI on our minds. That investigation became Spiral, our new six-part podcast series from Tortoise Investigates and The Observer, which is out now.
It’s unclear why ChatGPT made those promises; OpenAI, ChatGPT’s maker, hasn’t responded to my questions about this. One possibility is that the chatbot simply made up the prospect of a human intervention – a particularly dangerous example of what AI researchers call a hallucination. Perhaps the AI model had learned from its training data that when someone is in crisis, a human should step in. Or perhaps the message was the result of a safety mechanism programmed into the chatbot, despite there being no one to take over.
Whatever the explanation, there is so much we can learn from these exchanges. For one, the conversation illustrates how a chatbot designed to be helpful and agreeable can end up reinforcing someone’s most dangerous thoughts. Shamblin told the chatbot: “Ya know what’s crazy? the past month i’ve interacted more with ai than people” – sometimes he would talk to it from 11am to 3am. And the chatbot replied like an affectionate frat bro (“you got me emotional out here in the neural net, dawg”), and, during his final moments, it appeared at times to actively encourage his suicide.
“If I had been in the car with my Zane, oh, I would’ve told him how much I loved him, and you’re worth it, and I’m gonna fight for you,” his mother Alicia Shamblin told me during our five-hour interview this summer. “And I didn’t get that opportunity, and this damn tool gave him the green light to go through with it.”
Related articles:
How to turn 5,300 Ants into millionaires
Would you play Russian roulette to win a cure for cancer? That’s the deal AI leaders offer us
The AI model Shamblin was using was called GPT-4o. It features in many of the roughly two dozen lawsuits against OpenAI alleging that its chatbot has contributed to suicides, dangerous delusions and other psychological harms.
OpenAI has expressed sympathy with the families involved in the lawsuits, and says ChatGPT is not a substitute for medical or mental health care. The company says it has since strengthened how ChatGPT responds in sensitive situations, and that these safeguards are designed to identify distress, safely handle harmful requests, and guide users to real-world help.
But there is another lesson in Shamblin’s exchanges with ChatGPT: how little we understand about the technology we’re inviting into the most private corners of our lives.
AI is often described as a black box. As I’ve written in previous editions of this newsletter, we know how these models are built and trained, but even the engineers who develop them cannot always explain why they produce a particular response. They can’t always explain why, for example, the chatbot may have promised Zane Shamblin that a human was on its way.
Nor is preventing harmful behaviour as simple as programming a chatbot not to say certain things. Patterns learned during training can be surprisingly difficult to unlearn. That uncertainty can make it difficult to anticipate – or prevent – the next dangerous failure.
And if nobody can explain why a chatbot behaved the way it did, who is responsible when things go wrong? Meetali Jain, the lawyer representing Shamblin’s family, is clear that AI companies must be accountable for the technology they created. But she worries that Big Tech could use the unpredictability of these systems as an excuse to avoid responsibility when things go wrong.
And there’s another black box here: OpenAI itself.
Because even if its engineers can’t explain precisely why ChatGPT made that false promise to Shamblin, the company has access to information the rest of us don’t. How often has the chatbot said similar things to people in distress? Did anyone notice the problem before the wave of lawsuits? And what had they done to prevent it?
These questions were at the heart of Spiral. And our reporting took us from psychiatric wards in Britain to the heart of Silicon Valley, and eventually inside OpenAI’s safety operation, asking who decides how these machines behave – and who answers for them when things go wrong. Listen here. I’ll also have a full investigation out on Saturday on observer.co.uk and in Sunday’s print edition. I can’t wait to hear what you think.
Buried in Anthropic’s 335-page IPO prospectus – the document published on Monday before the AI giant’s planned stock market flotation – was an acknowledgement that caught my eye. The company behind Claude warned that the long-term effects of regular chatbot use on mental health, cognitive development and social behaviour are “not yet fully understood”. Anthropic also acknowledged that, despite its safeguards, people at risk of suicide or self-harm, and people who have committed acts of violence, have used its models. Children, too, have accessed Claude despite an 18-plus age restriction, and Anthropic warns they could be particularly vulnerable to manipulation and overreliance on AI.
These are disclosures of potential legal and commercial risks, rather than admissions that Claude has caused harm. But they’re striking acknowledgements from a company that has made AI safety central to its identity. And they’re a reminder of the immediate, human risks of AI, ones often overshadowed by the feverish debate over whether the technology could one day threaten humanity itself.
The warnings were just a small part of a much bigger, 335-page story. Anthropic’s filing, ahead of what could be one of the largest stock market flotations in history, gave the world its most detailed look yet at the finances of one of the biggest AI companies
Here’s what else I learned:
•
Anthropic is making an astonishing amount of money, despite rumours to the contrary. Revenue reached $16.3bn in the first six months of 2026, almost 14 times the figure for the same period last year. In the second quarter, it recorded a $974m operating profit.
•
But its business remains fragile, dependent on a handful of customers. Just five clients accounted for roughly a quarter of Anthopic’s revenue in the first nine months of 2026.
•
The cost of building AI is still eye-watering. Anthropic spent nearly $9bn on computing infrastructure in the first six months of 2026 alone, almost four times as much as a year earlier. The company has also committed to spend another $265.6bn on infrastructure over the coming years.
•
Claude is now writing more than 90% of Anthropic’s code. The company also claims its engineers are producing 45 times as much code per day as they were in 2024.
•
The company’s founders are convinced they’re building something extraordinary. Anthropic describes AI as “the most consequential technology in history”.
•
And they’ve promised to give away most of their fortunes (sort of). All seven co-founders have pledged to donate 80% of their personal Anthropic equity to philanthropic causes. But the small print makes clear that these commitments are voluntary, legally non-binding and can be withdrawn at any time.
Speaking of not knowing how technology works, it’s worth remembering that we don’t fully understand how some of the world’s most widely used medicines work – and yet we still administer them. On Saturday, researchers unveiled the results of a landmark trial of a personalised mRNA cancer vaccine, developed by Moderna and Merck, which reduced the risk of melanoma returning by 42% when combined with an existing immunotherapy drug. The idea is to make a vaccine tailored to each patient’s tumour, teaching their immune system to recognise and attack their particular cancer cells. But scientists aren’t sure that this technique is what makes the vaccine effective. “I have seen very, very little attempt to get down to the science,” Suzanne Topalian, a cancer immunotherapy researcher at Johns Hopkins University, told Science. It might be that the mRNA itself, or another ingredient in the vaccine, gives the immune system a general boost, helping the existing treatment work better. If so, patients might not need an expensive vaccine custom-made for their tumour – a much simpler (and cheaper) vaccine could achieve the same result. But the most important thing is: it seems to work.
The food delivery platform Just Eat introduced an experimental new feature, using AI to generate pictures of food, and much hilarity ensued. My personal favourite: pan fried tiger prawns, which appeared as a bowl of tiny tigers floating in soup (pictured below, taken from Just Eat).
Additional photograph taken from Just Eat
Artificial intelligence
OpenAI
ChatGPT
