AI That Spots Anxiety Online May Just Be Spotting the Word

A new study asks an uncomfortable question about AI that scans social media for signs of anxiety: is it actually reading people's emotions, or just spotting the word "anxiety"? Researchers compared six AI setups on a set of Reddit posts — renting a top commercial model (a "frontier" model, meaning the biggest, most expensive AI available), custom-training a smaller one in-house, and using a plain keyword-counting program — all tested under identical conditions.

Then they found the trap. In that Reddit collection, 69.3% of posts labeled as showing anxiety contained the word "anxiety" or a variant — roughly twice the rate of posts labeled otherwise. That means an AI can look accurate without understanding anything, simply by hunting for a keyword. So the team re-ran every test with those words deleted and measured how far each model's score dropped.

The drops ranged from 8.6 to 25.4 points. Surprisingly, bigger didn't mean better: the most keyword-reliant system was a custom-trained model, even more reliant than a simple word-counting program, while the top commercial model was the least reliant. And a small in-house model (110 million parameters — small enough to run on ordinary servers) scored 0.831 versus the commercial model's 0.846, essentially matching it without sending anyone's data to an outside company.

What this means for you: published accuracy claims for anxiety-detection AI are inflated, especially when the model was custom-trained. If a platform, employer, or health app claims its AI spots mental-health struggles, that number is probably too good to be true. The upside is genuinely encouraging — a cheap, private, locally run model can nearly match the expensive ones, so smaller clinics and apps don't have to hand people's most sensitive posts to a big tech vendor.