In week one, people using an AI assistant to identify fake news were 21% more accurate than the control group. By week four, when the AI was taken away, their accuracy didn’t rise — it fell, ending up 15 percentage points below where they started before the experiment. Stranger still, about one in five participants “felt they had gotten better.”

This comes from a study published on August 25, 2026 by Pattie Maes’s team at the MIT Media Lab (covered the same day by MIT Technology Review; paper at arXiv:2510.01537). They named it the “AI dependency paradox.”

1. What the study found

The experiment used paired-materials evaluation: for four consecutive weeks, participants were exposed to pairs of news headlines with accompanying images (real/fake combinations), judged each one’s authenticity, and the system recorded accuracy, response time, and subjective confidence with and without AI assistance.

• Week one (with AI): identification accuracy was 21% higher than the control group;
• Week four (without AI): accuracy was 15 percentage points below the pre-experiment baseline;
• Subjective experience: about one in five participants “felt they had gotten stronger” — a stark contrast with their objective decline.

Note one boundary: that 15-point degradation refers specifically to the “week-four test phase without AI assistance” — it does not mean performance worsened while AI was available. What it really implies is this: when AI suddenly becomes unavailable (system failure, policy restrictions), users face a clear capability gap in recovering independent judgment.

2. Why more help makes you worse

1. The nature of LLMs gets overlooked. Co-first author and PhD student Anku Rani points out that users are often dazzled by the “magic” of large models while forgetting their essence — an LLM is a model that predicts the next token in a sequence based on statistical patterns; it lacks genuine understanding and reasoning. Hand your judgment over to something that doesn’t understand, and you stop practicing.

2. Telling-style vs. questioning-style determines whether you practice at all. This is the core of the mechanism:

• “Telling-style” AI: gives answers directly — instantly accurate, but breeds dependency and erodes independent judgment;
• “Questioning-style” AI: uses Socratic counter-questions to guide your thinking — slower at first, but pushes you to actively discriminate, improving you over the long term.

Co-first author and PhD student Valdemar Danry added a key line: “Real learning happens when users personally engage in the discrimination. When AI does everything for them, users lose the opportunity to practice and to make mistakes.”

3. A trade-off. There’s a trade-off between immediate accuracy and long-term learning: the more the tool helps, the less you practice; the less you practice, the worse you become once the tool is removed — and that closes the loop of the paradox.

3. What to compare it against

This study is not an isolated case; it echoes several recent empirical findings:

• After doctors over-relied on AI imaging diagnostic systems, their missed-diagnosis rate rose significantly when working without the systems;
• Socratic questioning interfaces in programming assistants have also been shown to improve users’ problem-decomposition ability.

But this study is the first to quantify the long-term erosion of cognitive ability from dependency effects in the high-stakes social scenario of “fake news identification.”

4. What it means for you

• The information-literacy strong: proactively ask AI to interact in a “questioning” mode (counter-questions like “What cues are you using to judge authenticity?”), and build habits of independent verification.
• High-frequency decision / time-sensitive scenarios (e.g., breaking-news public-opinion assessment): short-term reliance on efficient AI is reasonable, but schedule an “AI-free review” session to guard against capability decay.
• Media and educational institutions: use Socratic-style AI as training material, guiding readers/students to internalize the signatures of misinformation through interaction (exaggerated phrasing, unclear sourcing, traces of image compositing).

This study does not recommend any specific product — its conclusions focus on interaction-design principles, not on any particular model. When choosing tools, pay attention to whether they support guided follow-up questioning rather than only delivering definitive answers.

5. Judgment

AI can be a bridge pier for cognitive extension, but it can also be a hotbed of mental laziness. Future human-AI collaboration design must distinguish between two modes — “doing it for you” and “assisting your learning” — and build cognitive retention into the interface layer, so that speed and depth no longer have to be either-or.