Frequently asked questions

Yes. Independent teams at universities such as Harvard and Stanford have replicated social-science findings with language models. Their studies consistently show that AI agents can approximate human responses in controlled experiments.

• Higher fidelity: our replication accuracy improves the 50%-75% range reported in prior work to 93.2%.
• Stricter tests: we purposefully reproduce studies published after January 2023 to avoid overlap with the model's pre-training data.
• Operationalisation: the AITurk platform wraps the workflow into a service for researchers and practitioners.

No. All research materials are encrypted at rest and accessible only to the users and teams you authorise. Platform staff cannot inspect your projects without permission.

Intellectual property and authorship remain yours. We contractually commit not to reuse study assets outside the scope you approve.

Yes. Enable the "Custom" option in the demographic panel to define bespoke attributes for income, occupation, health status, overseas experience, and more.

Ensure that the traits you request are internally consistent (for example pairing the CEO role with an age bracket above 30). Consider cross-checking age, education and occupation together.

A small share of AI agents can fail to return valid responses, and the augmentation logic introduces minor noise when rebalancing the sample. The discrepancies are usually modest and do not affect headline conclusions.

Yes. \"Location\" refers to the country or region assigned to the simulated participant.

No. Empirically, 30 samples per cell already match the accuracy of larger crowdsourced studies (93.2% in our trials). We therefore recommend 30-per-group as the default setting.

Scenario experiments manipulate variables via vignettes that participants read before responding. Recall/writing designs ask participants to reflect on past experiences or compose text (for instance, answering prompts or describing imagined scenes).

Yes. Image upload is supported in instruction fields such as "Guidance" and "Condition description".

Not all of them. The platform focuses on tasks that could reasonably be executed on crowdsourcing sites like MTurk or Prolific, where structured instructions and textual responses are sufficient.

The sliders are subjective. Choose "Medium" for most studies. If you expect a strong relationship (for example job satisfaction vs. life satisfaction) pick "Large"; if you are purely exploring weak associations, pick "Small".

No. The AI respondents always follow instructions, so attention checks can introduce artefacts instead of improving quality.

No. It only affects authoring convenience. If you need to mirror a previous study, choose the presentation that matches the original.

Not necessary. The models interpret the text directly, so styling does not influence their responses.

No. Pagination is not needed because the models process the full prompt context at once.

Yes. You can provide materials in any language supported by the underlying model (English, Chinese, Spanish, Portuguese, Korean, and more). Be mindful of potential cultural differences when mixing languages.

We calibrate sampling temperature to keep variance low, enabling reliable pilot studies with small cell sizes (around 30 samples per group).

Edits are free before running training. Once the system generates data, retraining requires a new run because it triggers fresh model calls.

Yes. Use "Import study" on the workspace page and paste the shared link to copy it into your project list.