HumanRank guides make three kinds of statement, and we label which kind each one is. This page explains the difference, what a published test must include before a score appears, and how pages are dated and corrected.
Three levels of evidence
1. Documented fact. A statement that a vendor's own documentation supports: an upload limit, a supported language, a default training setting. Each documented fact links to the specific vendor page, and we record the date we checked it. Example: our chatbot guide cites OpenAI's help center for ChatGPT's 512MB file limit.
2. Editorial judgment. Our reading of what the documented facts mean for a task. "Start with Grammarly for everyday editing" is a judgment based on features, platforms and plan limits, not a measurement. We name the basis for every judgment so you can disagree with it.
3. Tested result. A finding from an experiment HumanRank ran, with the method and data published. As of 14 October 2026, HumanRank has not published any tested results. Every launch guide says so, and every score column reads "Not tested".
Independent research is a fourth input. Where a study by a research group, newsroom or standards body bears on a decision, we cite it with its own numbers, date and scope. Our AI search guide, for example, cites the Tow Center's 2025 test of eight AI search tools. We do not present another organization's results as our own. Two examples: the Tow Center's test of eight AI search tools, which found incorrect answers to more than 60% of queries, and METR's randomized trial of experienced developers, which found tasks took 19% longer with AI tools in early 2025. Both are dated, and both describe their own limits.
What a test must include before a score appears
A HumanRank score will be published only when the test behind it includes all of the following:
- A versioned task set that resembles real work, including difficult cases, published or described in enough detail to repeat.
- Exact product, model and plan versions, with the settings used.
- Run dates, because AI products change between releases.
- Retained failures. Every attempt is kept, not only the best one.
- Scoring rules written before the test, including who scored and whether scoring was blind.
- Denominators and uncertainty: how many tasks, how many runs, and how much the result could vary.
- Limitations: what the test does not show.
Scores from different tasks are not pooled into a single "best AI" number. A coding result says nothing about image quality.
Category criteria
Each category is judged on what matters for its work:
- Chatbots and search: whether answers are supported by the sources they cite, and how often errors appear without any warning.
- Coding tools: accepted, correct changes and the time spent reviewing them, not lines generated. See the coding tools guide.
- Writing tools: whether meaning, numbers, hedges and citations survive an edit, and how much repair the output needs.
- Image and video generators: adherence to a fixed brief, usable output at delivery size, and the rights and provenance attached to the file.
How guides are dated
Each guide shows the date its sources were last checked. That date moves only when we re-check sources or change what the page says. We do not refresh dates to make a page look current.
How corrections work
When a vendor changes a limit, plan or default, we update the affected sentence, re-check the surrounding claims and change the date. If we got something wrong, we correct it on the page. You can report an error through our contact page. Please include the page and the source that shows the change.
What does not affect rankings
Payment does not buy a position, a recommendation or a better score. Our disclosure page explains the commercial relationships that exist and how they are kept apart from editorial decisions. See all categories to start.