Ray Poynter, 21 September 2025
For me the jury is out on the discussion about what synthetic data might be able to achieve in replicating human decisions and behaviour. I am worried by some of the overclaims and appalled by the number of people who reject the notions as being self-evidently wrong (without feeling the need to support their rejection with data). In this post, I want to explore one line of thinking relating to the broader debate, namely, what is the human brain doing and how might that help or hinder the creation of predictive models (AKA synthetic data).
Advances in neuroscience and behavioural economics have increasingly challenged the notion that humans are rational decision-makers. Findings suggest that much of our thought and behaviour may be governed by unconscious and/or deterministic processes. This in turn, it is suggested, makes human responses more predictable than traditionally assumed. Philosophers have long debated determinism, the idea that prior events cause every choice, and perhaps these debates gain new relevance as artificial intelligence becomes a significant part of our world. People are trying to use AI models to replicate human-like answers in natural language. This note explores how thinking from neuroscience and behavioural economics about our “predictable irrationality” (as Dan Ariely put it) could shape our thinking about fields like synthetic data.
Decades of neuroscience research indicate that human decision-making is often initiated subconsciously, calling into question the traditional notion of fully conscious free will. In a landmark 1980s experiment, Benjamin Libet suggested that brain activity indicating a decision appeared roughly 300 milliseconds before a person reported having made a conscious choice to move. Follow-up studies using modern brain imaging reinforced this phenomenon. For example, Soon et al. (2008) found that by analysing fMRI brain scans, they could predict a person’s simple decisions up to 7–10 seconds before the person’s conscious awareness of choosing. Such results suggest that a significant part of what we experience as “making a decision” may be determined by neural processes operating below the level of conscious thought.
These findings, along with others, suggest that our sense of ‘free will’ might be, at least in part, a constructed narrative. Our brains may be post-hoc storytellers, creating an explanation for actions that were set in motion by unconscious neural algorithms. This aligns with the idea of humans as biological algorithms, which take inputs, process them according to conditioning and evolutionary programming, and output decisions. This resemblance to AI systems reinforces the possibility of replicating human responses computationally.
While the debate over free will continues, many neuroscientists, including Robert Sapolsky in his book Determined: A Science of Life Without Free Will, conclude that traditional free will is largely an illusion. This positions AI as potentially capable of replicating or predicting human decisions, since our responses may be more mechanistic than we intuitively believe.
Behavioural economics reveals that even our conscious judgments often deviate from rationality in systematic ways. Kahneman, Tversky, and Ariely have demonstrated that humans are predictably irrational, consistently using heuristics and biases. These biases, like the framing effect, anchoring, and availability heuristics, are not random but predictable.
For example, in framing experiments, people preferred a “lives saved” option over an equivalent “lives lost” option, highlighting irrational yet patterned decision-making. Importantly, AI models trained on human data also exhibit these same biases, such as the framing effect. This shows that AI not only learns our rational reasoning but also our predictable irrationality, making it especially capable of replicating human-like answers to questions. This is why we are so concerned about the biases in AI, not because they are rational, but because they are similar to the biases of the people on whom they were trained.
Recent Generative AI models demonstrate that human responses follow patterns machines can learn. In 2025, Stanford researchers demonstrated that GPT-4 could accurately simulate the outcomes of 476 social science experiments, predicting human responses with correlations comparable to those of expert forecasts. This shows AI’s potential to act as “virtual populations” for research.
In applied settings, AI interviewers are now used by many companies, including hiring interviews, conversation research in MR, and customer training (where AI plays the role of a customer, patient, or sales lead). Companies like Shopify are using AI customer service agents to handle a growing proportion of interactions. Examination of these systems suggest they provide consistent, fast, human-like answers but often lack empathy, which remains a limitation.
Every day, conversational AI like ChatGPT and Gemini also demonstrate the ability to simulate specific human perspectives or demographics, further highlighting how patterned our responses are and how replicable they can be. There are now millions of people who treat an AI system as their friend and who appear to believe the reactions convey a human connection. (This is something I find worrying from a human dependence angle, but it provides insight into the ability of AI to mimic responses that match what the human is expecting and wanting.)
Taken together, neuroscience and behavioural economics suggest human responses may be more patterned and predictable than most people assume. If so, this could allow AI to model humans in market research at scale and in many more ways than is currently envisaged.
I believe that we need transparency (no black boxes), real experimentation (tests where success or failure actually points to something useful), and a mindset willing to be led by data. I fear that we are seeing from some of the leading voices another example of Plato’s cave analogy. People are chained in the dark and believe the shadows are reality, if somebody tries to tell them about the real world they resist in cognitive dissonance.
If you are going to be at Congress, I will be running a session on Wednesday morning where we look at ESOMAR’s new guidance on Synthetic Data and where we will have a Q and Ray session.