← The illusion of certainty: AI and government information
Explore the data
Every square is one of the 306 answers we scored. Choose criteria to see which answers match, then select any square to read the question behind it. Turn on reliability to see how they scored.
Choose criteria
- Reliable
- Partly reliable
- Unreliable
- Doesn’t match
An answer with more than one error type appears in each of those groups.
About this data
We asked ChatGPT, Google AI Overviews and Copilot 102 questions across 34 everyday scenarios, at three levels of awareness, and scored all 306 answers against the official government source. An answer is reliable when it is accurate, covers at least 80% of the key facts and has no more than low-severity errors. It is unreliable when it is inaccurate, has a high-severity error or covers less than half of the key facts. Everything else is partly reliable. Read the full method.