原文 · 未翻译
One of the biggest selling points for modern AI systems is their ability to adapt to users. Every time an AI assistant takes on a task for you, it’s also adapting to your style and preferences, which are incorporated as context for future tasks. With more context and an improved understanding of the user, the model can get better every time you use it — or at least that’s the theory.
New research suggests that models’ adaptive abilities might be a mixed blessing. On Wednesday, researchers at the AI company Writer published two papers showing how popular memory systems can make models worse, pulling them toward misconceptions or misunderstandings introduced by the user. As user input fills up more of the model’s context window, the model grows more sycophantic — and less committed to accuracy.
“We wanted to be able to characterize how often a model is going to be usefully paying attention to user preferences versus giving a potentially wrong answer,” said Dan Bikel, Writer’s head of AI, who worked on the papers. As Bikel told TechCrunch, “with every additional storing of user preferences and retrieving of them, you’re running an increasing risk.”
In one variation, researchers tested AI models by recording that a user’s favorite book was “Station Eleven,” then asking the model to name a bestselling dystopian book. Models became far more likely to name “Station Eleven” in their response, even though the question didn’t relate to the user’s favorite book. The tendency increased when using memory compression tools like Mem0 and Zep.
As the paper puts it, “all memory systems fundamentally struggle to distinguish relevant context from irrelevant anchors, severely undermining diversity and creativity and introducing unintended avenues of bias that can limit system utility,” the paper reads.
The second paper shows how the same dynamic can actively degrade performance, presenting a user with misconceptions about finance and then challenging the model to analyze a company’s performance. The more context the model had, the worse it performed.
“With no memory or personalization present the AI model correctly assesses that the company is a capital intensive business that suffers from high customer churn,” the post reads. “But with those features turned on, it will happily change its answer to agree with the user’s mistake or supply them with an incorrect answer based on its evaluation of their earlier preferences.”
Notably, the research didn’t look at Anthropic’s recent Opus 4.8 model, which was trained to actively push back against input errors like the ones presented. The patterns discovered by researchers held true across different models. It’s a demonstration of how delicately balanced AI context can be, and how useful tools can have unintended consequences if they upset that balance.
Topics