A new academic study pokes a hole in the assumption that feeding AI systems more customer data automatically makes their output better.
Researchers studied how generative AI models handle "context" - the situational information a company supplies at the moment it generates a recommendation, like current inventory or promotions - versus "customer evidence," the shopper's own history. They ran a full-factorial experiment with a generative recommender used by a large home-furnishing retailer, testing how different amounts and types of context affected recommendation quality. The results sorted into four states: not enough context, enough, too much (saturation), and context that actively confuses the system (interference). The researchers call the point where adding more context stops helping and starts hurting the Context-Sufficiency Frontier.
The finding complicates a core assumption behind the current rush to stuff more data into AI systems: that relevance beats volume. Irrelevant context didn't just sit there unused - it measurably reduced how appropriate the recommendations were and made retrieval less stable. For any company racing to bolt retrieval-augmented features onto a product, that's a warning that the instinct to "add more context" can backfire.
It's a quieter, more useful result than another leaderboard-topping model announcement - proof that in AI personalization, as in most systems, knowing what to leave out matters as much as what to put in.