Preventing machines from misunderstanding what we want.
Style Inference Labs, Inc is a consultancy and collaborative network of humanities scholars, social and computer scientists, policy experts, and business leaders. We are united by the belief that AI systems interpreting human behavior often rely on a shallow view of desire. Our approach, style inference, is a quantitative method adapted from psychoanalysis and the humanities. It introduces a missing interpretive layer to AI systems, offering a more logically sound and psychologically truthful account of human behavior. Moving beyond stereotypes in category based inference and focusing instead on how individuals personally relate to what they want, style inference makes machine predictions safer, truer, and more aligned with real human motives.
The twin experiment of the humanities
Though humanly impossible, imagine twins exposed to the same external world from the same perspective. If we could read their mind, will we learn that they relate to the world in the same way? Does everything in the world mean the same exact thing for one as the other?
This is not a trivial philosophical thought experiment. Many social and computer scientists already work with a stubborn answer to the puzzle. By reducing individuals to demographic categories like race, class, or gender or to standardized psychological models, they betray their commitment to an ultra-deterministic world. Our impossible twins, according to their logic, would not only look the same. Their feelings would be identical too.
The standard answer of the humanities to our thought experiment, though hard to quantify, makes more sense. Human beings are not animal reward mechanisms. We do not passively recognize and follow what is deemed valuable. We are ourselves creators of what we value and we create meaning - we don’t subscribe to it. Sounds human… but can we teach robots to see us in this psychologically deep way?
Two personal styles of desire
Personality is rooted in what philosophers have called the self and psychologists have called the ego. Critics have tended to see contemporary culture as increasingly egoistic, self-centred. What they missed is that we cannot escape ourselves. Even desire, meant to be an appreciation of something beyond us, implies a relationship with ourselves. We’re always trying to deal with ourselves: what we seek in the outside world is a way to do that.
Style inference identifies two personal styles of desire: projection and introjection. Think of something you like doing in your free time. Does x make you forget yourself? Do you consider it an escape from yourself in a bigger world? Then you’re likely desiring x projectively. Does y make you aware of yourself? Do you consider it a return to yourself in a distracting world? Then you’re likely desiring y introjectively. Two subtle and deeper ways we all desire… what can go wrong if we ignore them?
It’s about internal consistency
Existing algorithms pick up on fragments, not the whole picture. They detect external, observable changes but miss out on the common thread tying all our actions together. What does your switch from pop music to heavy metal have to do with your recent browsing of new shoes? Existing algorithms can’t tell: they only propose more shoes and more heavy metal.
Internal consistency is where style inference comes in handy. What is more consistent about ourselves than our relationship with ourselves, the interplay between be yourself and get over yourself? Playing new music or buying new shoes fulfill these deeper and more personal aspirations. More importantly, our styles are our real deal-breakers. If a giveaway campaign rewards users with selfies next to their favourite heavy metal artists, it probably shouldn’t target users for whom the whole point of listening to heavy metal is thinking about something other than themselves.
Style inference, a new name for an older science of desire
Before AI, there was no shortage of intelligence. The last century began with two revolutions: Einstein’s, who propelled us into the scientific and technological world we recognize, and Freud’s, who discovered a deeper psychological world in human beings. The legacies of these two geniuses diverged. Einstein became synonymous with high-level mathematics only machines can keep up with. Freud became associated with qualitative skills and methods of interpretation only used in the humanities. The divergence is partially the reason we are now troubled by machines actively interpreting our behaviour. The robotic successors of Einstein are trying to play Freud despite long being estranged from his descendants and their methods.
Machines misunderstand what we want because we ourselves can’t see what we want. The most important part of what we want is unconscious, beyond the observable realm. This misunderstanding is likely to grow as we increasingly use LLMs like Chat-GPT as deficient mental health assistants. But what if it is possible to reconcile Einstein and Freud, physics and psychology, science and the humanities, machines and us?
How style inference works
Figuring out whether people are projecting (giving meaning) or introjecting (receiving meaning) in a specific moment is a delicate task. Imagine twins asked why they both have Teslas:
Twin 1: Because it’s highly recommended & looks great…
Twin 2: Because I like how it feels when I’m driving one…
Absence and presence of self-reference respectively suggest projection and introjection. But with more input, probabilities can change:
Twin 1: …though I like that my brother has one too.
Twin 2: …though all its gadgets are pretty cool too.
Bear in mind it is always possible for people to relate to the same thing projectively and introjectively. Style inference is especially useful in cases of imbalance, when one style is more dominant than the other. These are the cases that make us most human. What’s something you ultimately grew to like for its own sake, without self-regard? Reversely, what’s something so connected to your sense of self? Survey questions like these are helpful for style inference, but we have developed a range of experimental and interactive assessment methods.
The humanities have spent centuries uncovering how people relate to themselves, to others, and to the world across cultures and contexts. Style inference translates these insights into probabilistic priors, which, when combined with behavioral data through our style-sensitive interpretive layer, can enhance the performance and safety of existing AI systems.
Ready for the paradigm shift?
We are actively seeking to implement style inference in a range of AI systems from LLMs designed to enhance mental health to recommendation systems. Whether you’re a researcher, company, or investor, we want to hear from you. Fill the form below or get in touch with us via contact [at] styleinferencelabs.com