AI tools for detecting fake news may not be as effective as you think.

Major tech companies like Meta, Google, and X are investing heavily in AI tools to detect fake news on the internet. This sounds reassuring, but according to a new study from the Université de Montréal, these systems still have many limitations hidden behind seemingly impressive accuracy figures.

 

Doctoral student Dorsaf Sallami analyzed numerous AI systems for detecting fake news and found that they don't actually fact-check the information. Instead, these systems only calculate probabilities based on the training data they are provided.

In other words, their operation is not like a journalist verifying information, but rather like a mirror reflecting the data they have been trained on. This means the system can reproduce biases and blind spots that exist in the training data.

According to Sallami, a system that achieves 95% accuracy in a laboratory setting can still fail when applied in real-world situations, and the gap between these two environments is a significant concern.

The issue of bias is rarely discussed.

Beyond the issue of accuracy, the research also found that many AI-powered fake news detection systems carry underlying biases that users are largely unaware of.

Some models tend to label women as a greater source of misinformation. Others are biased when evaluating non-Western sources, or inadvertently replicate pre-existing political biases in training data.

 

Furthermore, a deeper problem exists concerning how these systems are trained. Much of the label data is obtained from fact-checking organizations, but many of these organizations lack transparency, and some even operate as for-profit businesses.

This means the data foundation that AI systems rely on is not really robust.

Furthermore, the emergence of content creation tools like ChatGPT has made creating fake content easier than ever. Therefore, fake news detection systems trained just months ago can quickly become obsolete.

To address this problem, Sallami proposed a new tool called Aletheia. This is a browser extension that helps explain why content might be suspicious, rather than simply stating whether it is true or false.

In tests, Aletheia achieved a reliability rate of approximately 85%, surpassing many existing fake news detection tools. The biggest difference of this tool lies in its design philosophy.

Instead of delivering a 'verdict' and requiring users to believe it, Aletheia presents its entire analytical process. The tool gathers evidence from online sources, presents it in easily understandable language, and lets users make their own final decision.

Additionally, Aletheia provides a bulletin board that updates users on recent fact-checking activities, along with a community forum where users can share and discuss their findings.

The message this research conveys is quite clear: AI should support human judgment, not completely replace it.