AI fake news detectors are not as good as you think

Tech giants like Meta, Google, and X are investing heavily in AI tools designed to detect fake news. It sounds reassuring, but according to a new study from the Université de Montréal, these tools have some serious drawbacks hiding behind impressive-sounding accuracy numbers.

Doctoral researcher Dorsaf Sallami examined AI fake news detection systems and found that they don’t actually fact-check anything. They calculate probabilities based on their training data. Think of it less like a journalist verifying a story and more like a mirror reflecting whatever it is shown, including the same biases and blind spots.

According to Dorsaf Sallami, a system that scores 95% accuracy in a lab setting can still fail in the real world, and that gap is a serious problem.

The bias problem nobody is talking about

Beyond accuracy, Sallami found that many of these systems carry embedded biases that largely go unnoticed. Some models are more likely to flag women as sources of misinformation. Others are biased against non-Western sources or reproduce political prejudices.

There’s also a deeper issue with how these systems are trained. They rely on labels from fact-checking organizations, many of which lack transparency and some of which are for-profit businesses. The entire system is built on a shaky foundation.

Add to that the rise of tools like ChatGPT that make fake content easier to produce than ever, and detection systems trained even a few months ago can quickly become obsolete.

A better approach

Sallami’s solution is Aletheia, a browser extension that explains why content might be suspect rather than just saying whether it is true or false. In tests, it achieved 85% reliability, outperforming many existing tools. What makes it different is its philosophy. Instead of handing you a verdict and expecting you to trust it, Aletheia shows its work. 

It pulls evidence from available online sources, presents it in plain language, and lets users make the final decision. It even includes a live feed of recent fact checks and a community forum where users can share and discuss findings. The takeaway is simple: AI should assist your judgment, not replace it.