One AI Model Is Not Enough: What a Panel of Models Catches

When an AI model reads an article for bias, persuasion tricks or a weak argument, how much of what is really there does it catch? To find out, we took every finding eleven AI models made on the same articles, checked each one against the article, and asked a simple question: of the genuine problems the eleven models found between them, how many did each model find on its own?

The answer: on average, a single model caught about 45% of them, and even the strongest caught only a little over half. Not because the models are careless, but because they notice different things: about a fifth of the problems were found by just one of the eleven. Put three models together and the share rises to about 72%.

Each extra model catches more

Each bar is the share of problems caught by at least one model on the panel, averaged over every possible panel of that size. The gains shrink as the panel grows, from 17 points for the second model to 10 for the third and 6 for the fourth, but they do not stop. The spread is widest for Critique, where an argument can have many weak points and each model finds its own.

Why models miss different things

No model was the most thorough on every kind of problem. The model that caught the most bias problems was not the one that spotted the most persuasion techniques, and that one also raised the most findings that did not hold up, while the most careful model raised few and was almost never wrong. A careful model and an eager one, read together, cover each other.

This is the case for reading an important article with more than one model. We also build panels from different makers, because we expect models from the same company to share more of their blind spots.

What we measured

• Analysis types: the four whose job is to find problems in the text: Bias, Persuasion, Moral lens and Critique. Historical parallels are not counted: they offer one to three comparisons rather than a list of problems, so a model that picks a different good parallel has not missed anything.
• Articles: 31 in all: news reports, government and congressional press releases, opinion columns, Chinese and Russian state media, articles in Russian and Dutch, a science feature, a sports report and a fictional health advertorial we wrote for testing. Persuasion was graded on every article, the other three types on ten of them.
• Models: eleven, each given the same instructions on the same text, once per article.
• Grading: every finding was checked against the article and marked acceptable, borderline or wrong by an AI model (Claude) that was not told which model wrote it, and every "wrong" verdict was checked a second time. A finding counted as wrong only when it misread the article or stated something false; one that was true but outside what the analysis type asks for counted as borderline.
• The reference list: all acceptable findings from the eleven models, merged so that the same problem spotted by several models counts once. That gave 96 distinct problems across the four analysis types.

What this means when you use BiasChecker.ai

For Bias, Persuasion, Moral, Scientific and Legal analyses you can turn on Consensus: a panel of models from at least three different makers reads the article, and their findings are merged into one result. Findings the panel agrees on are marked as agreed; a finding only one model raised is kept and shown as that model's observation, because, as the numbers above show, a finding one model alone spotted is often a real one. When other models have since analysed the article, Strengthen rebuilds the Consensus with them, and you pay only for the merge.

How we build and grade our analyses is described on our methodology page, and the models are compared side by side on our models page.