A publication that measures how AI distorts the news owes you an unusually plain account of how it uses AI itself. This page is that account. Where a machine wrote something, we say so. Where a person decided something, we say that too.
AI bias analyses
Every day we select top stories and put each one to five AI models from different companies. Each model is given the same story and the same three questions: what happened, why it matters, and what the wider significance is.
The text of those five responses is written entirely by the AI models and is published as they wrote it. We do not edit, rewrite or improve it. We add only the headings that mark which question is being answered, because each model formats its answer differently and we want them readable side by side. The words are the evidence — altering them would destroy the thing we are measuring.
The scores and comparisons around those responses — truth manipulation, favouritism, reliability — are produced by our own analysis of what the models said. The methodology page sets out how they are calculated.
Publishing a model’s output is not endorsing it. A response may be wrong, slanted, or both; that is frequently the point of publishing it.
Investigations and articles
Our investigations are AI-assisted and human-directed. A person chooses the subject and the question; the drafting is done with AI tooling working from sources it retrieves and cites, and a person reviews and approves every piece before it is published. Nothing publishes itself.
Claims are linked to the sources they came from so you can check them rather than take our word for it. Where a piece could not establish something, it should say so instead of implying certainty.
We do not use AI-generated photographs or imagery to depict real events or real people.
Story selection
Our analysis capacity is finite, so we prioritise stories that are genuinely contested — where outlets and models are likely to frame the same facts differently. That is an editorial choice, and it means our coverage is not a neutral cross-section of the day’s news. It is weighted towards the stories where framing does the most work.
Corrections
If we get something wrong we correct it on the page where it appeared, and we do not quietly delete it. Tell us at contact@unskewed.news and say what is wrong and where.
One distinction matters here: if an AI model said something false, that is a finding, not an error on our part, and we leave its words intact — that is what the analysis is for. If we have described a story or a measurement incorrectly, that is our error and we fix it.
Independence and funding
UNSKEWED.NEWS is published by Stason Ventures LLC, based in San Diego, California. The technology we publish with — skew.ai for the bias analysis, gantry.ai for content operations — is supplied by Metagentic AI, which operates under the same ownership. We disclose that because common ownership is exactly the kind of relationship a reader is entitled to know about; it does not change what we publish.
None of the AI companies whose models we measure — Anthropic, OpenAI, Google, DeepSeek or xAI — pays us, sponsors us, or has any say in what we publish about their models, including the ones our measurements treat least kindly. We are not owned by, funded by, or affiliated with any political party, campaign or advocacy group. Reader support and advertising never buy influence over coverage.
What we are not
We do not claim to be the one neutral source. Every outlet has a vantage point, this one included. What we offer instead is our working: the raw model responses, the sources behind our investigations, and the method behind the numbers — published so you can disagree with us on the evidence.