Truth Perception Dataset (TPD) — anonymous data on how people tell true from false
The Truth Perception Dataset records how people judge statements as true or false: what they voted, how sure they were, and whether they got it right. It is the data missing to anyone working on misinformation, trust and model calibration — not stated opinions, but decisions actually made.
What it contains
Each row is one judgement: the statement (true or false), the category among 7, the language among 5 (it, en, de, es, fr), the vote, the stated confidence and the outcome.
From this you get the measures that actually matter: how calibrated a population is (says 80% and is right 80% of the time), which categories fool people most, how perception shifts across languages.
How it is anonymised
The dataset is anonymous: k-anonymity with k=5 (groups below the threshold are suppressed), differential-privacy noise on aggregates, keyed pseudonymisation with a per-export salt and synthetic replacement of residual values.
No re-identification key is provided, and the licence forbids attempting re-identification or linking the dataset to other sources to achieve it. Every export carries its own privacy scorecard in the response headers.
How to integrate
REST API with a key: CSV, JSONL, Parquet export, filters by time window, category and language, incremental resume with a cursor. Current version v1, with a declared deprecation policy and a minimum 180 days notice.
Keys are issued from the portal and governed: scopes (metadata only or export too), expiry, IP allowlist, rotation with a grace period. Hourly quotas per plan: TRIAL 5 requests/hour, STARTER 20 requests/hour, GROWTH 60 requests/hour, ENTERPRISE 1000 requests/hour.
Who uses it
Research groups on misinformation and the psychology of trust; teams training or evaluating language models who need human baselines on truth perception; products that measure the calibration of their own predictions.
Licence, privacy and service levels
The dataset licence and the portal privacy notice are readable without signing up. The dataset contains no personal data of the app's users; the notice covers the licensee organisation's own data.
Declared service levels: 99,5% monthly availability on export and metadata routes, first support response within two business days, incident notice within 24 hours.