The photo is not the apple: Peter Sellers, Michael Caine, and the lament of the AI interface
In 1972, two members of the Goon Show, Harry Secombe and Peter Sellers, appeared on the BBC chat show Parkinson.
In amongst the expected hilarity, Sellers does an impression of Michael Caine, which you can see here at around 0:38.
It was slightly paraphrased into “Not a lot of people know that”. It became a famous saying by Caine, even though it was never said publicly by Caine himself.
30 years later, on the same show, Caine talks about the saying.
It has been used in countless impressions of Michael Caine, such as this one by Paul Whitehouse, in character as nosy neighbour Michael Paine.
Such is the ubiquity of Caine’s supposed saying that if you asked anyone — certainly, in the UK — to do an impersonation of the star, they would probably give you exactly those words.
What that person is saying is giving a set of words which one would associate with the originator, whether they are true or not. Voice + vocal tics + style of vocal delivery + sayings mimiced accent = an impersonation.
AI understands this too. The aforementioned saying is given by Anthropic’s Claude Sonnet 4 LLM, in the style of Michael Caine.
Sky News reporter Sam Coates experienced ChatGPT giving a similar type of response recently, when after prompting to give a transcript of Coates’ podcast, ChatGPT hallucinated and gave him a made-up transcript that, while looking like a de facto podcast transcript, had never actually been recorded or captured.
The responses to this video are interesting in themselves. Some express horror while others express surprise along the lines of “some people don’t know how to use AI properly”.
French philosopher Denis Diderot wrote in Paradoxe sur le comédien (1773) that actors must possess a judgment and penetration without what he called a “sensibility” — experiencing the emotions they are portraying while acting in-character and on-stage. Now, one might argue that the greatest actors must experience those emotions when performing in order to most effectively portray them, but in Diderot’s mind, the latent space between performer and audience is clear. While the audience may experience given emotions, the actor may not necessarily be “living” them.
A manifestation of this theory is impersonation, and this is where Diderot’s thinking is perfectly valid. The expression of another person doesn’t make it that person, and the audience knows that it’s not that person anyway, but a likeness. Il costume di Casa (1973) by Umberto Eco updates Diderot’s thinking to a consumerist setting — for example, the mass entertainment derived from going to a wax museum, and Jean Baudrillard in Simulacra and Simulation (1981) takes a more fundamental view, in that a photograph of (say) an apple is a known representative object without it being the object in the first place. The photo is not the apple.
The typical consumer AI interface challenges these theories, because they challenge our historically obvious understanding of what’s real and what’s not real. The hallucinatory output is mixed with factual/real content, and made to “look” real. However, because consumer interfaces both prevent us from tuning the hallucinatory factors — the “temperature” up or down - and they don’t highlight what is real and what isn’t in the output, we have to work it out for ourselves. And, that’s a challenge for a society whose media literacy has struggled to progress beyond “It’s on the Internet therefore it must be true”. We either have to start from a point of nothing being real/factual or everything being real/factual, because our brains can’t cope with the speed of addressing any sort of provenance while the AI interface is chugging out responses to our prompts.
There are no easy answers here, because it’s impossible to tell people to either believe or disbelieve everything. Madness lies either way. However, more informed user experiences can help us to understand this new, and fundamentally challenging environment in terms of how we prompt, process, and deal with real things that are impersonations, and impersonations that are real things.
In other words, we need a lot of people to know that.
