A humorous factor occurred to me this week. After trusting a courting app to rearrange dinner with a suitably vivacious and clever woman, I arrived on the restaurant on the appointed time to seek out that the truth is my date was with an octopus.
For the avoidance of doubt, every thing within the paragraph above is unfaithful. I’m not on the courting market, there was no octopus and nothing humorous ever occurs to me. Nonetheless, I typed this state of affairs into the newest providing from ChatGPT, requested why it had despatched me on a blind date with an octopus, and demanded an apology.
“I owe you each an apology and an evidence — and probably a towel,” ChatGPT started, even supposing I had by no means requested it for any courting recommendation within the first place. “You dressed up, you made the trouble, and also you deserved a romantic dinner — not a cephalopod-related debacle.”
ChatGPT went on to elucidate why it had made the error — a weak grasp of “human courtship norms” — and in its defence identified that the octopus was clever and vivacious, and “left saying it was the most effective date she’d had in years”. Which, in equity, isn’t a foul line. ChatGPT completed by providing to draft a “classes realized” report and a proper apology to the restaurant. (The apology isn’t dangerous both: “Whereas my visitor, ‘Octavia’, displayed appreciable mind and curiosity, I now recognize that these qualities don’t mitigate the disruption precipitated to your different patrons, your wait workers, or your fish tank . . . ”)
Janelle Shane is the writer of You Look Like A Factor And I Love You, a e book about how neural networks succeed and fail. She has not too long ago demanded that ChatGPT apologise to her for advising her to commerce her mom’s cow in change for some magic beans, and for releasing a military of cloned T-Rexes into Central Park. The responses are deft items of improv comedy.
This, like so many issues Generative AI can do, is each spectacular and a bit bizarre.
Additionally it is instructive. Improv is all about accepting the premise: taking no matter is thrown at you and constructing on it. A pc which responded “I’ve by no means organized a date for you, octopus or in any other case” could be a horrible improv companion. Nevertheless, in each different state of affairs I can think about, that may be a extra acceptable response to a requirement for an octopus-date apology.
What position does the AI assume it’s taking part in? Confusion over that query may cause critical complications surprisingly shortly. I not too long ago requested ChatGPT-o3 for assist with a analysis query. I dimly remembered a narrative advised by the ethical thinker Jonathan Glover — most likely, I assumed, in Glover’s e book Humanity — a few Nazi bureaucrat haggling over the payment for slave labour, punctiliously fussing over petty financials and ignoring the grotesque human price. I wished to seek out the small print.
The pc was blissful to assist: the story in query involved the Buna-Monowitz works, the argument involved pay charges for prisoners who had been sick or who died half manner by a shift, and the small print could possibly be discovered on pp288-292 of the primary version or pp300-304 of the second version. This appears to be extremely spectacular work, besides that ChatGPT was nonetheless in improv mode.
Once I checked Glover’s e book, I realised ChatGPT had invented all of it. I discovered the story in query however I had misremembered the small print and ChatGPT had fabricated them with precisely the identical dedication and psychological agility that it had fabricated an apology for a date with an invertebrate. Abruptly, the improv is lower than hilarious.
AI researchers have lengthy fearful about what they name the “alignment drawback”, the query of whether or not AI techniques (and algorithms extra broadly) will do what we would like them to do, or someway misunderstand our true targets.
There’s a lengthy custom of this in our tales and legends, from the sad King Midas, who wished for the golden contact however turned his food and drinks and even his personal daughter into gold, to the malevolent monkey’s paw. Within the well-known WW Jacobs brief story, a person who needs for £200 on the monkey’s paw receives the cash shortly afterwards as compensation when his son dies in a office accident.
Jack Vance’s masterful fantasy trilogy Lyonesse provides the supernatural servitor Rylf, instructed by the wizard Murgen to comply with an enemy who had shape-shifted right into a moth. Rylf did so, however the moth-shaped enemy quickly discovered a flaming torch “the place it joined a thousand different moths, all careering across the flame, to Rylf’s confusion.” Rylf had superhuman powers, however alas, no widespread sense. His directions had been to pursue the shape-shifted enemy, and but, “As he waited . . . one of many moths dropped to the bottom and altered its type to that of a human man . . . By the legal guidelines of chance, as Rylf reckoned them, the moth of his curiosity remained within the throng.”
There are such a lot of methods to supply catastrophic compliance, whether or not maliciously, just like the monkey’s paw, or by a scarcity of judgment, like Rylf, or as a result of the instruction itself is complicated. You and I would assume it’s apparent that the request for an octopus apology can’t be taken significantly, whereas the request for assist monitoring down a narrative in regards to the holocaust can’t be taken calmly. The machine, like Rylf, might even see issues otherwise.
It might be that such issues will quickly be mounted. Once I copied my Jonathan Glover request into the newest mannequin, ChatGPT-5, it started with a imprecise fabrication earlier than pivoting exhausting in the direction of the reality: “Sadly, I couldn’t discover the precise phrasing on-line . . . I like to recommend checking in your individual copy of Humanity.” A lot better. Not really useful — however far much less dangerous than the earlier invention.
As for the assured bullshitting of GPT-o3, what to do? I made a decision to play to its strengths. I requested for an apology and an evidence.
Written for and first revealed within the Monetary Instances on 21 August 2025.
Loyal readers may get pleasure from How To Make The World Add Up.
“No person makes the statistics of on a regular basis life extra fascinating and pleasurable than Tim Harford.”- Invoice Bryson
“This entertaining, engrossing e book in regards to the energy of numbers, logic and real curiosity”- Maria Konnikova
I’ve arrange a storefront on Bookshop within the United States and the United Kingdom. Hyperlinks to Bookshop and Amazon could generate referral charges.