AI recommendation models fail basic tests; simpler methods work better

Researchers found that many new AI models for recommending products or content do not work as well as claimed. When they tried to reproduce the results, they found that simpler, older methods often performed better, especially when properly tuned.

This paper suggests that a significant portion of research in AI recommendations is not producing real progress. It appears many published results overstate the benefits of new AI techniques. This means companies and researchers might be investing time and money into AI approaches that are not actually superior to existing, simpler methods. The research calls for more rigorous testing and a change in how AI research is published.