# FAQ

## How do I choose an AI benchmark?

Start with the task you care about, then pick a benchmark in that category that is not saturated for the systems you compare. Check its data licence and whether you can run it yourself. Our guide on choosing an AI benchmark goes through each step.

## How do I tell if a benchmark is saturated?

A high average score is a warning sign, not the definition. Look at whether scores still separate the systems you need to compare. Our guide on saturated benchmarks lists signals to inspect.

## Can I compare scores from different sources?

Only with care. Scores depend on the harness, prompts, settings and benchmark version, so two numbers for the same benchmark may not be comparable. Read the top-score source before you rely on a number.

## What can I do with a benchmark's data?

That depends on its data licence. Some allow commercial use and others do not, and terms for training on the data differ. Read the licence on the listing and its source.

## How are benchmarks ordered?

Lists put the most complete listings first. Featured placements and sponsored slots are clearly marked. We never call a benchmark best.

## What does Verified by author mean?

It means the benchmark's authors have claimed the listing and its details link to their sources. It is not our endorsement of the benchmark.

## How do I claim a benchmark I maintain?

Open the listing and select Claim this listing. We confirm ownership with an email to an address at your organisation's domain or by a manual check. Then you can edit the details and their sources.

## How do I add a benchmark?

Use the submit form with a paper, repository or leaderboard link we can cite. An admin reviews each submission before it goes live.
