Artificial Societies releases Rothko human behavior prediction model

Article

Artificial Societies releases Rothko: enabling AI human decision prediction with only one-quarter of Qwen’s survey error

Dongcha Beating AI breaking news, London-based AI startup Artificial Societies has released Rothko, a human behavior prediction model. Companies can first have AI simulate thousands of consumers, industry professionals, or policymakers to test how new products, advertisements, and major decisions may be received before deciding what to do in practice.

Even when ordinary large models role-play people of different ages and occupations, they tend to produce similar answers. Rothko modifies the architecture of a pretrained model so that it directly predicts people’s choices. It continues training on anonymized social data, transaction records, travel trajectories, expert interviews, and survey questionnaires to reproduce the actual preferences and disagreements of different groups as closely as possible.

Using data from 2796 real respondents, the company simulated and generated 68280 survey responses. Compared with 9 general-purpose large models, Rothko’s distribution of answer proportions was closest to that of real people, with an average error of 8 percentage points. The best-performing general-purpose model,Qwen3.8 had an error of 31 percentage points. After the wording of the questions was changed, Rothko’s prediction error for the magnitude of response changes was only 1.4 percentage points, compared with 6.2 percentage points for the second-place model.

Original link https://m.theblockbeats.info/flash/373436

Key points

London AI London startup Artificial Societies has released the human behavior prediction model Rothko.
Rothko can simulate consumers, industry professionals, or policymakers to test the potential responses to new products, advertisements, or major decisions.
Rothko continues training on anonymized social data, transaction records, travel trajectories, expert interviews, and survey questionnaires.
The company used 2796 real respondents’ data to simulate and generate 68280 survey responses.
In a comparison with 9 general-purpose large models, Rothko’s average error was 8 percentage points, while Qwen3.8 had an error of 31 percentage points.
After the wording of the questions was changed, Rothko’s prediction error for the magnitude of response changes was 1.4 percentage points, compared with 6.2 percentage points for the second-place model.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Rothko does not merely generate seemingly different answers; it attempts to predict how different groups of people will choose. Companies can first use it to simulate survey or decision responses and then use the results to plan actual testing.

Why it matters to readers

If the relevant test results can be independently reproduced, models of this kind could help companies conduct market research, product testing, and decision simulations.

It demonstrates AI an application that shifts from generating text toward predicting human choices, but the performance data in the article still requires further validation.

Risks and unknowns

The performance data in the article comes from information released by Artificial Societies; no independent verification results were provided.
The authorization, privacy protection, and representativeness of the training data were not explained in the evidence.
It remains unclear whether the model can maintain its advantage with other populations, questions, or real-world decisions.

Artificial Societies releases Rothko human behavior prediction model

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center