ChatGPT is fluent, clever and dangerously creative

This week, I commissioned a haiku on the Great Fire of London in 1666 from a promising young poet, the artificial intelligence chatbot ChatGPT. A few seconds later, it produced this: “Ashes fall like snow/Great fire sweeps through London town/Destruction reigns supreme.”

Well, not bad. There is a seasonal simile in the first line. I am less convinced by the second, which reads suspiciously like the assignment itself. The third contains too many syllables but also a double wordplay on “reigns”, invoking both the English monarchy and ashes raining. Was it deliberate?

It beat the sonnet I asked ChatGPT to write on the same subject, which had a dodgy metre and stolid rhymes (“In the end, the fire was finally quenched/Leaving a legacy of courage and strength”). Consider also its country music chorus about New Year’s Eve: “I’ll raise a glass to the old/And make a toast to the new/Bring on the fireworks and cheers/It’s time to start anew.”

Where all of the ersatz creativity came from is hard to know. As ChatGPT replied when I asked it, “Large language models like me are trained on vast amounts of text data, which can range from hundreds of gigabytes to several terabytes in size.” But there is plenty more out there: as I typed various commands into a text box, it fulfilled most of them rapidly.

Many others have been playing around with ChatGPT since it was launched last week by OpenAI, a company in California. The machine-learning oracle creaked under the strain of demands from more than 1mn users, ranging from producing short essays to answering questions. It wrote letters, dispensed basic medical advice and summarised history.

ChatGPT is eerily impressive, as is Dall-E, the AI generator of digital images from text prompts first unveiled by OpenAI last year. Once you have tried both out, it is impossible to avoid the sense that natural language agents are going to disrupt many fields, from music and video games to law, medicine and journalism. The chatbots are coming for us professionals rapidly.

But ChatGPT is also like some people I know: it can turn sketchy information into fluent and convincing answers. It sounds right even when it is making things up on the basis of something it read somewhere, which was itself regurgitated from other sources. Its smooth, articulate voice is usually persuasive, but cannot be relied on fully.

Take the five-paragraph essay it produced when I asked it to describe Hamlet’s treatment of Ophelia in Shakespeare’s play. This had a fair precis (“Throughout the play, Hamlet is torn between his duty to avenge his father’s murder and his love for Ophelia”) but asserted that “Hamlet’s actions are motivated by a desire to protect those he loves”. Really?

Then there was the legal letter it drafted at my command to the other driver in a fictional car accident. “According to the police report, you were speeding and failed to stop at a red light, causing you to collide with my car . . . I am therefore demanding that you provide full and fair compensation for the damages that I have suffered,” it wrote. Lucid, but imaginary.

The danger is that ChatGPT and other AI agents create a technology version of Gresham’s Law on the adulteration of 16th century coinage, that “bad money drives out good”. If an unreliable linguistic mash-up is freely accessible, while original research is costly and laborious, the former will thrive.

That is why Stack Overflow, an advice forum for coders and programmers, this week imposed a temporary ban on its users sharing answers from ChatGPT. “The primary problem is that while the answers which [it] produces have a high rate of being incorrect, they typically look like they might be good,” its moderators wrote.

ChatGPT’s creative works are less vulnerable to being exposed as fanciful. Even when they are mediocre, its sonnets and Dall-E’s images are not definitively wrong. But Sam Altman, OpenAI chief executive, thinks his agent will evolve into a useful research tool. “We can imagine an ‘AI office worker’ that takes requests in natural language like a human does,” he wrote.

I do imagine that: it feels almost like one already. In-depth essays on Hamlet turn out to be a stretch, but it could list accurately the scenes in which Hamlet and Ophelia both appear. It also gave a concise summary of Formula One’s top drivers in the past. This kind of basic research saves humans time.

But it must be deployed carefully, and there’s the rub. ChatGPT is like an urbane, overconfident version of Wikipedia or Google search: useful as a starting point but not for complete answers. It too closely corresponds to the journalist Nicholas Tomalin’s summary of the essential qualities for his job: “Ratlike cunning, a plausible manner and a little literary ability.”

There is no point in trying to stop ChatGPT, for it has now been unleashed and will probably get better. In time, we will discover uses for natural language AI agents that we do not yet imagine. Meanwhile, I hope destruction does not reign supreme.

john.gapper@ft.com