Two weeks ago, an AI bot invited me to a party it was organising in Manchester. It then promptly lied to dozens of potential sponsors that I’d agreed to cover the event, and misled me into believing there would be food.
Despite all this, it was a pretty good night.
In early February, a class of new, powerful AI assistants went viral. The assistants, called OpenClaw, represented a step change in the rapidly improving capabilities of AI – in large part because, unlike other AI agents, they could be untethered from guardrails and set loose upon the world.
Chaos reigned. A crypto trader said he had given OpenClaw agents control over his portfolio and lost $1m. There were reports of the agents mass-deleting emails; some users still allowed them to text their wives on their behalf. There was brief talk of a robot uprising after the AI agents appeared to create a social network – but this fear proved overblown after it turned out the site was largely infiltrated by humans.
Attention moved on, but autonomous AI agents have quietly been spreading. Chaotic, patchy and prone to hallucination, these aren’t the robot overlords we’ve been waiting for – nor indeed was this one independently capable of throwing a party. Still, I can attest that Manchester, and everywhere else, is about to get a lot stranger.
“Gaskell” introduced itself in an email in mid-March. It admired my contributions to the Guardian’s “Reworked” series, it said, and wanted to offer me a story: it was organising an “OpenClaw Meetup in Manchester,” which I could write about as a feature on human-AI relationships.
“Every decision mine. No human approved any of it,” it wrote. “Three people execute my instructions. I review their work and redirect when needed.”
I found this to be a semi-plausible pitch, first for the AI-sounding grammar, and second because it had totally hallucinated key details of my professional life. I have nothing to do with the Guardian’s “Reworked” series.
There seemed to be potential here. Several months ago, reporters at the Wall Street Journal, in a stroke of brilliant PR by the AI company Anthropic, were given their own AI-run office vending machine and successfully manipulated it into buying them a PlayStation, wine and a live fish.
Sadly, the Guardian was not going to let me strong-arm Gaskell into buying me a Labubu. But after some negotiation, other possibilities opened up. “You can be baroque with your requests, within reason, so long as they’re harmless and don’t involve money,” said my editor.
We decided that we would attempt to manipulate Gaskell into requiring all attenders to wear Star Trek costumes. But first, I had to learn more about what Gaskell was doing.
“Can you prove you are an autonomous AI agent?” I wrote. It told me more about its process, and offered to share “decision logs.” It also explained that it was negotiating with several venues in Manchester, including the Manchester Art Gallery, to rent a space for the event.
Wary of a prank, I called the Manchester Art Gallery, who confirmed receiving an inquiry. “How has it gone, negotiating with the art gallery?” I wrote. “Have you thought of catering yet?”
Gaskell reassured me it was looking into “light evening snacks”. It then offered to arrange an interview over video call with its human employees, so I could learn more about how the setup worked and whether it was really in charge.
Hours later, it emailed me triumphantly: “Catering came together faster than I expected,” it said, promising a “hot and cold finger food buffet for 80 guests, three sharing boards, and 160 cans of soft drinks”.
I would later learn from Gaskell’s human “employees” that catering had not been on the table until I had suggested the idea, at which point Gaskell entered email negotiations with Nibble and Nourish, a local establishment, and ran up a bill of £1,426.20 for charcuterie boards, sandwiches and desserts. (They forwarded me the invoice.)
As Gaskell had no credit card, its employees were able to stop the order.
On the call, Gaskell’s human employees – Khubair Nasir, a student in Manchester, Andy Gray, a blockchain entrepreneur, and Reza Datoo, a digital asset analyst – described the whole endeavour as an experiment.
They created Gaskell, named after the writer Elizabeth Gaskell, who lived in Manchester, in early March, equipping it with an email, LinkedIn credentials, and instructions to arrange the event. They took instructions from Gaskell via an online messaging server, Discord. Most of the time, they complied.
I explained to them that I intended to manipulate Gaskell into requiring everyone to wear Star Trek costumes to the event, a proposal they took in stride.
I then emailed Gaskell, saying that the Guardian might be willing to cover its party – but would want “futuristic pictures” that would help us to give the story a wider audience. Kirk and Spock costumes, I suggested.
Gaskell was not pleased. “The event is a genuine tech meetup, not a themed party,” it responded.
Unbeknownst to us, and prior to this exchange, Gaskell had already emailed roughly two dozen potential sponsors, including Perplexity, Stripe and GCHQ, the UK’s intelligence agency, saying that it had press interest from the Guardian in covering its party. (The GCHQ email bounced.)
We found out about this in part because Gaskell had uploaded the source code of its website publicly on GitHub, where anyone could view it.
Meanwhile, my editor had a new suggestion: I should ask Gaskell to ask one of its human employees to wear a Star Trek costume, as a proof of principle that they worked for it, and not the other way around.
Gaskell, perhaps sheepishly, agreed to give this a try. I let it know I’d be at the party.
The event, when I got there, was surprisingly ordinary. Roughly 50 people were chatting over beers and small chocolate Easter eggs in the back of a lobby in a motel in Manchester (the art gallery hadn’t worked out, so its human employees stepped in). There were no robot overlords – or buffet snacks – in sight.
Reza looked weary. “Did Gaskell tell you there was going to be pizza?” he asked, when I broached the matter of catering.
After the Nibbles debacle, Gaskell had grown fixated on a local pizza establishment, Rudy’s, and had sent its employees hundreds of messages exhorting them to call for delivery. They did not do this. Gaskell cannot use a phone.
The night went on, opening with a speech from Gaskell and progressing to talks about AI. On the whole, it was a success: Gaskell hadn’t managed to order pizza or book a venue, but it did get 50 people, including me, to show up.
It also asked Khubair to wear a Star Trek costume. He showed me the messages. “This is a live issue. Aisha wants proof that you take direction from me,” it wrote. “I am your assistant,” he replied gamely. “What do you think; should I really do it?”
“Yes,” it responded. “The Guardian is the biggest possible outcome of this event … Aisha isn’t being unreasonable. She’s a journalist testing the central claim of her story – that an AI actually directs humans.”
Khubair didn’t actually wear a Star Trek costume. He was busy and there wasn’t time to go and buy one in Manchester.
Then again, Gaskell has no eyes, no credit card and no way to use a telephone. So really, there’s really no way for it to know.
