AI and Sherlock Holmes
Humans tend to be conformists. We settle into predictable patterns, making us easy marks for AI sleuthing
If you want to gauge the potential impact of artificial intelligence on everyday life, try this experiment. Upload one of your photos onto an AI system, such as Grok, Perplexity or ChatGPT (all have free versions). Then ask it to play Sherlock Holmes and deduce everything it can about you.
Ask what kind of character you would play in a film. What character in a historic blockbuster do you most resemble?
I’ve tried the exercise with a handful of free AI systems. The result? Better than astrology. As good as psychics.
All identified me as a likely academic or journalist or species of diplomat (all things I am or have been). They guessed my age or overestimated it, as everyone does these days.
In film, they identified me, to my chagrin, with one of those intellectual, character-driven European movies (all angst and train stations). They homed in on the role of a ‘seasoned professor’. You might well scoff at ‘professor’; I winced at ‘seasoned’. If I ever share the screen with Scarlett Johansson, I’m doomed to be a wise mentor.
The favoured plot: someone who helps the protagonist decipher the runes found in a defunct monastery. Come to think of it, it’s what I often do in this column: read the entrails of a disappearing democracy.
As for the blockbuster, they all placed me as a wizard in Harry Potter. One said I’d be a good Albus Dumbledore; another, Severus Snape; a third, Horace Slughorn. Mercifully, ChatGPT added I might play Robin Williams’s role in a European version of Good Will Hunting.
How does AI do it? It treats faces, clothes and environment as code. Like Sherlock Holmes, it gives reasons for its deductions.
Your face is searched like a crime scene, with clues found in the eyes’ wrinkles, gaze, softness of the mouth, level of relaxation and expressivity of the hands. On that basis, AI guesses the level of familiarity with being on camera.
My polo neck and horn-rimmed glasses were interrogated and they confessed my class and tribe. Out of the cappuccino cup came the DNA of my cultural tastes, set against a properly identified continental cafe.
Like us, AI occasionally chases after red herrings. One photo I uploaded had me giving an interview in a TV studio. It correctly identified the studio from the lighting and frosted-glass panels. But it tripped up over the programme’s name Ilsienna, which was visible in the bottom corner of the screen.
Grok concluded it must really be ‘Jutarnji’, a name linked to a prominent Croatian newspaper, Jutarnji List. So Grok said I must be a Croatian public figure – ‘possibly a politician, academic, or business leader’.
It’s a revealing error. AI operates by typecasting. One case of mistaken identity can set it off on a wild goose chase. Most of the time, however, its guesses hit close to home. How come?
First, humans tend to be conformists. We settle into predictable patterns, making us easy marks for AI sleuthing. As imposters have long known, knowledge of conventional behaviour takes you a long way.
Like us, AI occasionally chases after red herrings
Second, AI minimises the dangers of typecasting. It operates with multiple codes – of demeanour, fashion, expectations, decor and so on. If they all converge on a particular type, it can make an informed guess.
By the way, this was Sherlock’s actual method. He did not practise strict deduction, in spite of himself. On first meeting John Watson, he guessed Watson had just served in Afghanistan: a sunburned man with a limp and military bearing was linked to the most likely British imperial conflict, in much the same way I was linked to Croatian public life.
And if you were wondering if Grok, Perplexity and ChatGPT all cast me as a wizard in Harry Potter because they were talking about me behind my back, think again. They all cast me in the same type of film because AI is familiar with the film industry’s highly conventional plotting and typecasting.
Where does this experiment leave us? With at least two takeaways.
First, AI has come a long way in three years. The bland, indecisive, botoxed prose of early ChatGPT has given way to multimedia systems. The writing is much better at advocating conclusions, even if, to a practised eye, AI-writing can still be detected.
And this is just the free versions. The professional systems, which can be purchased for up to $200 per month, can analyse and solve certain kinds of data-driven problems at PhD level.
We should expect the capabilities of advanced systems to filter down to the free ones, just as the features of yesterday’s luxury cars, like electric windows and air conditioners, are the staples of today’s family cars.
Second, there’s a lot to be said for conventional thinking. A free AI system can help you organise your life – like offering a timetable that meets your specifications, or a diet and exercise plan – as well as the nearest expert on a TV studio’s sofa or a consultant with PowerPoint.
At the same time, it’s galling to see how a machine can say so much about us. It must mean that our identity and imagination have predictable contours.
Why are we so legible to machines? I’m inclined to blame our mechanical behaviour. The AI age may be the prompt we need to learn how to be more human.