Can AI predict human behaviour?
In a job interview, a candidate pauses before answering a question about conflict. She could be recalling a difficult conversation or choosing her words in a second language. Either way, a recruiter would like to know how she handles disagreements at work. New York startup millionways is betting that the language people use contains clues to what they will do next. Where conventional language models are trained to generate language, millionways is using it as evidence about the person speaking.
On November 12, millionways introduced Thorsten-4, a foundational 'Large Psychology Model' that predicts how humans think, decide and act.
Businesses can buy access to Thorsten-4 through a platform that builds profiles from material such as transcripts and documents and simulates how someone might respond in different scenarios. They can also integrate the model into software they already use through an application programming interface (API). A customer could feed in interview transcripts, emails or chats with an AI agent and use the resulting assessments in an existing workflow. The company offers API access under a separate written agreement and lists self-hosting as another option.
The model grew out of a research programme that co-founder and chief executive Martin Cordsmeier began through a nonprofit foundation in Germany, according to the company. He went on to build millionways in New York with chief technology officer Max Weidemann. The company says it has raised $6 million in US financing. Its case for Thorsten-4 rests on the data gathered over years of research, which it says involved more than 7,000 participants and over 50 psychologists and behavioural scientists. The launch release says psychologists annotated more than 35 million words.
What language can tell us about behaviour
There is a scientific basis for looking for psychological clues in language. Linguistic Inquiry and Word Count (LIWC), developed by psychologist James Pennebaker and colleagues, counts words in categories such as pronouns and emotional language so researchers can test whether their use varies with psychological measures. In a meta-analysis covering more than 85,000 people, those categories explained about 5% of the variation in self-reported Big Five personality traits on average. A separate review of 44 deception studies found small overall effects that varied with the setting and the speaker’s involvement. Both findings show how easily a measurable association across a group can be overinterpreted in an individual case.
Following people over time could help distinguish a recurring pattern from a response to one particular situation. It still matters where the model learnt the pattern. An executive’s cautious phrasing on an earnings call could be meaningless in a different setting, for example. Even when a pattern carries across populations, it cannot by itself establish why someone acted. New information or a change of mind can also change what happens next.
millionways says Thorsten-4 assesses 50 traits, covering areas such as motivation, decision-making and responses to stress, and can show the conversational evidence behind an assessment. Its research page says Thorsten does not infer intent, even though the launch says Thorsten-4 predicts it. The company reports 90% predictive accuracy and says it checked its outputs against established psychological assessments and participants’ subsequent behaviour.
How businesses use Thorsten-4
What counts as a good prediction depends on the job the customer gives the model. In quantitative finance, millionways’ Behavioral Alpha product draws on earnings calls and other corporate communications to generate signals for trading strategies. The company says an unnamed hedge fund began trading a dedicated $50 million portfolio in the second quarter of 2026. The 37% return cited through mid-September came from testing before trading costs. The launch disclosed no live return for the fund.
Executive search firms look for a different kind of signal in interviews about decisions candidates have made under pressure. millionways says several firms compared the model’s readings with their own assessments and that it is building a tailored tool with one of them. To find out whether the model predicts performance in the role, researchers would need to follow candidates after they are hired.
A sales team could use the model to pick up on a buyer’s concerns about implementation, while a customer service team could flag a complaint for human follow-up if it looks likely to escalate. millionways says pairing Thorsten-4 with GPT improved assessments of customer intent and preferences in enterprise pilots, without giving the size of the improvement. Developers who build AI support agents could use the same API to look at the exchange from the agent’s side, flagging chats where it keeps replying without resolving the problem.
Who gets to use those inferences?
A customer who agrees to a call being recorded for quality assurance may have little reason to expect the company to infer motives from it. That change of purpose raises questions about mental privacy and consent, including what the original research participants agreed to when they shared their experiences. The youngest interviewee, according to the launch, was eight.
The stakes are particularly visible in hiring, where a score can affect who makes the shortlist. A University of Washington-led study found that AI models evaluating speech sometimes scored identical answers differently depending on the speaker’s accent. A separate study found that adapting interview questions improved ratings of autistic candidates’ answers. These studies examined other systems or interview practices, and illustrate how an assessment can reflect the conditions under which someone speaks.
In sales and financial services, the concern extends to how a business might use information the other party cannot see or challenge. An inferred concern could prompt a clearer explanation, or it could be used to press a customer towards a decision. In a controlled study of short debates, GPT-4 with access to basic demographic information was more persuasive than human opponents in about 64% of pairings where one side was more persuasive. The researchers tested demographic personalisation in debates, leaving open how psychological profiling would affect a commercial conversation.
Where the EU draws the line
The EU AI Act already bans using biometric data (such as facial expressions or tone of voice) to read people's emotions at work or in school, except for medical or safety reasons. It also outlaws certain forms of biometric profiling and AI designed to manipulate or exploit people in harmful ways. However, European Commission guidelines make an exception: reading emotions from written text alone is allowed. For a tool like Thorsten-4, this draws a sharp legal boundary. Analysing the words in an interview transcript is regulated very differently from analysing the sound of a candidate's voice.
Separately, the law classifies AI used to screen job applicants or check credit scores as high-risk. Any system in this category (known as Annex III) that profiles individuals automatically faces strict rules, including mandatory transparency and human oversight, starting on 2 December 2027. For a recruiter using behavioural scoring, human oversight means more than simply rubber-stamping an AI's decision. They must understand what evidence produced the score and have the genuine power to ignore it.
In practice, language can only predict how someone will act if that connection has been tested in the real-world setting where the tool is used. To prove Thorsten-4 works, its creators would need to define what behaviour they are predicting upfront, test the model on brand-new people, and openly show where it gets things wrong. In hiring, the real test does not happen during the interview. It happens months later, when the new hire actually faces the kind of workplace disagreement they were asked to describe.
Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.