I'm building AI for personality and trait analysis
To see me is to know me
It should be possible to make some determination about someone’s personality from what they say, and the tone in which they say it. After all, that’s what we do all the time, in our everyday lives - it’s all we have to go on. It’s also an important part of strategy and statecraft. Indeed, psychological profiles of key personalities have a very long history with both practitioners and scholars alike. I’ve mentioned before how much I enjoyed Fiona Hill’s work on Putin, for example. Another favourite example is CIA’s profile of Khrushchev prepared for JFK in 1961.
Astute personality profiles of key political figures is really the signature of biographers and historians. But psychologists too have found many ways to measure personality - aiming for systematic approaches, with some more reliable than others.
My question: Can machines do likewise?
(Prompt: New Yorker style cartoon of an AI psychotherapist)
To find out, I’ve added to my list of AI ‘basic research’ projects.* My broad agenda is to explore the ways in which machines do, and do not, mimic human psychology - especially language models. Project Allport is named after Gordon Allport - one of the foundational thinkers in the field, whose work thrilled me when I first stumbled across it in the library.
So, to work! As a quick proof of concept, I co-wrote a python script in Cursor (with Google’s Gemini as my co-author). This code ingests YouTube video (the sample I used is of Donald Trump holding forth in the infamous Oval Office encounter with President Zelensky.
Next, a language model (WhisperX) transcribes the audio file. I’m about to introduce another (Pyannote) to diarise it too, so we can tell who is speaking. This transcript then feeds into three psychology frameworks - two Big Five Personality tests (one done lexicographically with a fine tuned BERT model) the second by still another language model (GPT-4o). The Big Five is a very mainstream personality test - it’s robust/stable and somewhat predictive, and there’s a vast amount of empirical research in the books. You probably know the traits in question. Lastly, we do some sentiment analysis - again fuelled by AI, which does the data coding: and here, the sentiments are overwhelmingly negative; unsurprisingly, given the context.
If you can bear it, here’s the famous video, from Fox, for added authenticity!
And here are the Big 5 findings, helpfully processed by still another model - o3, which is great for data analysis and visualisations.
The differences partly reflect methodology:
The BERT lexical model weighs word choice frequencies – so, more literal linguistic cues.
OpenAI, by contrast, infers intent and context, so it may grasp some subtleties, like irony and sarcasm: ‘that’s just great’ &etc.
Paradoxically, I’d back the OpenAI model to be more sophisticated, even though it jars a bit here with what I think of Trump. But let’s not also forget, I’m only using a snippet of video as proof of concept - less than a minute, in fact.
So what? Well - for one it’s fun. I went from nothing to this in one evening.
For another, there’s the promise of much more to come. Tone of voice analysis, for example; body language, facial expression. Hell, even pheromones, it’s just data, after all. OpenAI’s advanced voice feature already recognises mood from voice tone and facial expression, and it’s not bad at it. And many more tests I can do too - on dominance, or the dark triad, for example. Should be fun.
Are they accurate results? Well, they face the same methodological challenges as human political psychologists, looking on from afar at their famous subjects. Plus some distinctive AI challenges too, of course. So, that’s a large question, for another day.
But my contention is that AI gets minds, without having one of its own, or at least, not one remotely like ours. Exotic mind-like entities, as DeepMind’s Murray Shanahan says. My other ‘basic research’ projects have established congruence in other areas of psychology between machines and minds, and I’d put good money that they work here too.
But is this being done for real, as part of intelligence analysis? Given that it took me three hours and a couple of dollars, I am utterly certain it is.
—
* The others, remember are Projects Schelling, LeDoux, Axelrod and Tversky. More on those elsewhere in my Substack…