The consequences of relying on AI for accurate news

Overview

A new open-access MIT Media Lab study found that people who leaned on AI chatbots to fact-check news grew worse at spotting misinformation on their own once the AI was removed. Across four weeks, 67 participants were 21 percent more accurate while assisted, but their unassisted accuracy on fresh news items fell 15 percentage points by week four. Researchers call this the AI dependency paradox, and they compare it to how GPS has dulled our natural sense of direction. About a quarter of participants thought they were improving even as their real performance dropped.

Key Takeaways

AI chatbots make people better at detecting fake news in the moment but weaker at detecting it later on their own, the study found.
Participants improved 21 percent in accuracy while using an AI helper, then declined 15 percentage points without it by week four.
Roughly one quarter of participants believed they were getting better at spotting fake news even as their actual scores fell, a Dunning-Kruger style overconfidence.
The researchers labeled about one-fifth of participants Dependency Developers, who shifted from thinking for themselves to passively accepting whatever the AI said.
AI that asks Socratic questions taught people more durable skills than AI that simply handed over direct answers, the team reported.
Pew Research data cited in the study shows one in five U.S. teens regularly use large language models for news, and one in four young adults have used them for news at least once.

Stats & Key Facts

#67 participants were tracked over four weeks while evaluating news headline-image pairs.
#Accuracy rose 21 percent when participants were assisted by an AI chatbot during a session.
#Unassisted accuracy fell 15 percentage points by week four compared with the start of the study.
#About 25 percent of participants reported feeling they were improving even as their performance declined.
#Roughly 20 percent of participants were classified as Dependency Developers who grew passive toward AI guidance.
#Pew Research found one in five U.S. teens regularly use large language models for news, and one in four young adults have done so at least once.

How the MIT Media Lab Tested AI and Fake News Detection

The experiment followed real people over a month to see what happens to their skills when an AI helper disappears.

Researchers at the MIT Media Lab tracked 67 people over four weeks. In each session, participants looked at pairs of news headlines and images and decided whether the news was real or fake, sometimes with an AI chatbot to help and sometimes on their own.

The design let the team measure two different things. First, how well people did while the AI was present. Second, and more telling, how well they did on brand-new news items after the AI support was taken away. That gap is where the trouble showed up.

The Gain That Came With a Hidden Cost

AI help worked in the short term, but it left people weaker once it was gone.

›Participants were 21 percent more accurate at flagging fake news while an AI chatbot assisted them during a session.
›By week four, their unassisted accuracy on new items dropped 15 percentage points compared with where they started.
›The short-term gain confirms earlier MIT Sloan School of Management work showing AI can reduce belief in false information.
›The long-term decline is the new and worrying finding: the help did not build a lasting skill.

The AI Dependency Paradox and Cognitive Offloading

The pattern fits a problem researchers have watched in other fields for decades.

The team calls this the AI dependency paradox. The more you rely on a tool to do the thinking, the less practice your own brain gets, so the underlying skill fades. Scientists call this deskilling or cognitive offloading.

The comparisons are everyday ones. Calculators weakened mental math. GPS dulled our natural sense of direction. The study points to a 2025 finding that doctors who used AI grew worse at spotting cancer on their own. Fact-checking news appears to follow the same path.

Overconfidence and the Dependency Developers

Many people felt sharper while quietly getting worse.

›About 25 percent of participants reported feeling they were improving even as their measured accuracy fell, a Dunning-Kruger style gap between confidence and reality.
›Around 20 percent were labeled Dependency Developers, drifting from active judgment to passively accepting the AI's verdict.
›One participant noted the chatbots reminded them to check multiple sources but never taught them how to read the context of an image itself.
›This blind spot matters because doctored or out-of-context images are a common form of visual misinformation.

Coach Versus Crutch: Why How the AI Talks Matters

Not all AI help produced the same result.

The researchers drew a clear line between AI that coaches and AI that acts as a crutch. Co-lead author Valdemar Danry, a PhD student in media arts and sciences, said AI that simply tells people the answer tends to breed reliance, while AI that asks questions in a Socratic style pushes people to think and actually learn.

Co-lead author Anku Rani added a reminder about what these systems are. In her words, people get excited about magical large language models but forget they are statistical models predicting the next token in a sequence of words, with real limits on what they reliably produce. The paper, titled around dialogues with AI and lasting discernment skills, was presented at the 2026 CHI Conference on Human Factors in Computing Systems.

Why This Matters for Everyday News Readers

The findings land at a moment when AI news use is climbing fast.

Citing Pew Research, the study notes one in five U.S. teens regularly use large language models for news, and one in four young adults have used them for news at least once. That growing reliance is exactly what the experiment tested.

The team warns that AI models are especially error-prone during emotionally charged breaking news, such as conflicts or attacks, when accurate information is scarce and training data may be unreliable. The practical takeaway for readers is to treat AI as a thinking partner that prompts you to check sources yourself, not as an oracle that replaces your own judgment.

What the Researchers Plan Next

The work points toward education and broader testing.

›Test the effect with geographically diverse and lower-resource communities rather than a single group.
›Explore culturally adaptive AI interactions tailored to different users.
›Help educators build curricula that use AI responsibly and grow AI literacy in schools.
›Encourage habits where people verify claims independently instead of accepting AI answers at face value.

Frequently Asked Questions

What did the MIT study actually find about AI and fake news?

It found that AI chatbots helped people detect fake news better in the moment, raising accuracy 21 percent, but their own unassisted accuracy fell 15 percentage points after a month once the AI was removed.

How many people were in the study and how long did it run?

The MIT Media Lab study tracked 67 participants over four weeks as they evaluated pairs of news headlines and images.

What is the AI dependency paradox?

It is the pattern where relying on AI to do a task improves results short term but erodes your own underlying skill over time, similar to how GPS has weakened people's natural sense of direction.

Did people realize their skills were slipping?

Mostly no. About a quarter of participants reported feeling they were improving even as their measured accuracy declined, an overconfidence gap the researchers tied to the Dunning-Kruger effect.

Is there a safer way to use AI for checking news?

The researchers found AI that asks Socratic questions helped people learn, while AI that handed over direct answers fostered dependence. Using AI as a prompt to verify sources yourself, rather than as a final answer, preserves your own judgment.

The study suggests AI works best as a coach that sharpens your own judgment rather than a crutch that replaces it. For news in particular, the safest habit is to let AI prompt you to check sources yourself instead of accepting its verdict outright.