FAQ

What is Subconscious.ai?

A platform for running causal experiments at scale. Instead of recruiting human respondents, it builds a synthetic population for the study and runs a randomised discrete choice experiment against it. Causal experiments of this kind are standard in product design, transport, political science, public health, and economics, wherever a decision-maker needs to know how a population will respond to different options.

Who uses it?

Researchers and product teams who need to know why a population chooses what it chooses: in academia, in industry, and in policy.

How is this different from asking a language model?

An experiment is not a question to a model. Attributes are randomised independently, respondents see structured choice tasks, and effects are estimated with the same statistics used on human conjoint data. Randomisation supports causal comparisons within the simulated experiment. Generalizing those comparisons to human behavior requires validation.

Why should I trust simulated respondents?

Because the platform is checked against human studies rather than assumed to work. Published human conjoint studies are replicated and compared effect by effect. See Human baselines.

How long does an experiment take?

Completion time depends on the design, population and queue. Poll every 30–60 seconds and retain the original run ID; see Poll a run.

What does it cost?

Experiments consume model capacity and are billed accordingly. Check your plan in Settings, and review a design before launching it rather than after.

Can I use my own attributes and levels?

Yes. Generate them from your question and edit, or supply your own outright. See Run an experiment.

Can I bring my own audience?

Choose demographic constraints, a supported population group, or the application's population upload flow. Modeled traits enrich selected respondents; they are not observed-behavior filters. external_personas is a separate LinkedIn-profile input. See Design a population.

Is my experiment private?

Set is_private: true on the request and the run is hidden from other users.

My run says finished but there are no results

Check the expected artifacts before treating the run as usable. A terminal label alone does not establish that the evidence you need exists. See Poll a run.

Is there an SDK?

Not yet. The API is plain JSON over HTTP, and there is an MCP server for driving it from Claude or Cursor. The published spec is available at /openapi/subconscious.public.json if you want to generate a client.

How do I report a problem?

Contact us with the run id and the exact request body.