Say vs do: how much of human behavior lives in what people actually did?

Most AI that simulates people today is a language model asked to play a role. Language models learned from text, and text is mostly what people say: posts, reviews, essays, answers. That is a lot of human behavior, and it carries real signal. But it is the side of behavior people choose to show.

The other side, what people actually do, lives somewhere else. It lives in purchase histories, viewing logs, clicks, returns and cancellations, and it trains a different family of models: the recommender systems behind every feed and store. Those two families have mostly been built apart. Whether they belong together is one of the most important open questions in modeling people.

The gap between saying and doing

Market researchers have a name for the problem: the say-do gap. The best-known example comes from sustainable shopping. In surveys, around 65% of consumers said they wanted to buy from purpose-driven brands that advocate sustainability; only about 26% actually did, according to research summarized in Harvard Business Review in 2019.

A model trained only on what people say inherits that gap. It learns how people describe themselves, which is often generous, and it has never seen the record that would correct it.

Behavior has its own scaling laws

Behavioral data turns out to be learnable at enormous scale. Meta's generative recommender work trained models with up to 1.5 trillion parameters on users' action sequences and found that quality improved as a power law with compute, the same kind of scaling that made language models work. The paper's title made the point directly: actions speak louder than words.

Behavioral data also travels. Centaur, from Binz and colleagues, fine-tuned a language model on Psych-101: more than 10 million trial-by-trial choices from over 60,000 people across 160 psychology experiments. It captured the behavior of held-out participants better than existing cognitive models, and generalized to new cover stories, changed task structures and entirely new domains.

Can one model learn both?

The natural next step is a single model that learns from what people say and what they do. The worry is collapse: training on interaction logs might wash out what the model knew about language and the world, or the reverse.

Early evidence is encouraging. Kuaishou's OpenOneRec, released at the end of 2025, co-trained foundation models on general knowledge and 96 million interactions from 160,000 users. It reported recommendation ability that scaled predictably while mitigating catastrophic forgetting of general knowledge, and a 26.8% improvement in recall on Amazon benchmarks. That shows the two can live in one model. It does not yet show that each makes the other better.

1. 01SayInterviews, surveys, reviews: what people tell you and why
2. 02DoPurchases, views, churn: what people actually did
3. 03One modelTrained on both, per person
4. 04TestDoes each improve, or only survive?

Why it matters outside the lab

Inside a company the two kinds of data already exist, in different departments. Research teams hold interviews and surveys. Product and growth teams hold behavior. If a model trained on both predicts customers better than either, the payoff is a single model of the customer that every team can ask, grounded in what customers said and what they did.

There is a smaller version of the same idea in digital twins. The Stanford study of 1,052 people found twins built from two-hour interviews beat twins built from demographics by a wide margin. Interviews ask about what people did, not only what they think. Grounding in behavior, even reported behavior, already helps.

What we would want to see

• The same people, with both interviews and behavioral records.
• Three models: trained on what they said, on what they did, and on both.
• Each tested on held-out answers and held-out actions.
• A clear verdict: does combining them improve both, improve one, or only avoid making either worse?

This is the question behind what we call a foundation model of human behavior. We think the answer is that both together beat either alone, and we would rather show it than say it.