The AI Dependency Paradox: Using AI to Check Facts Can Quietly Make You Worse at It

One in five U.S. teens now reach for a large language model when they want to know what is going on in the world. One in four young adults have done the same at least once. Asking a chatbot to confirm whether a story is true feels efficient, even responsible. A new open access study from the MIT Media Lab suggests the opposite may be happening underneath the convenience.

Researchers tracked participants over a month as they used AI systems to verify facts. When the chatbots were taken away, the people who had leaned on them most were measurably worse at spotting misinformation on their own than when they started. The skill did not transfer. It eroded. Researchers call this the AI dependency paradox, and it is not unique to news. A 2025 study found that doctors who used AI to read scans got worse at detecting cancer without it. The same pattern keeps appearing wherever a capable tool quietly takes over a human judgment.

What the study actually found

The mechanism has a name that predates AI: cognitive offloading. When a tool reliably handles a task, the brain stops rehearsing it. That is usually fine. Almost nobody can recite phone numbers anymore, and the world has not ended. The problem is that fact checking and misinformation detection are not storage tasks. They are reasoning skills, and reasoning skills decay when you stop using them. Offload your memory and you lose a convenience. Offload your judgment and you lose the thing that was supposed to keep you safe from being fooled.

What makes the finding unsettling is the direction of trust. People were using AI to check the news, treating the model as the authority that settles the question. But language models are confident by design and wrong often enough to matter. A user who has stopped practicing their own discernment has no independent way to catch the model when it is the one that is mistaken. The safety net becomes the single point of failure.

It helps to picture the difference between a calculator and a treadmill. A calculator is pure offloading and that is the point. You do not need to stay sharp at long division, so handing it off costs you nothing. Misinformation detection is a treadmill. The value is not the destination, it is the conditioning you build along the way. When you let AI carry you, you arrive at an answer without doing the work that was supposed to make you stronger, and over weeks that shows up as a real decline. The MIT participants did not get lazy. They got comfortable, which is harder to notice and slower to reverse.

Why this matters for educators and professionals

If you teach, this is the gap behind a lot of classroom anxiety about AI. The worry is rarely that students use the tool. It is that they use it in a way that hollows out the very capacity school exists to build. A student who asks AI to evaluate a source, accepts the verdict, and moves on has practiced nothing. The assignment was completed and the learning was skipped.

If you work in a knowledge job, the same trap is waiting in your inbox. Summaries you do not read, analyses you do not check, drafts you forward without testing. Each one feels like leverage. Stack enough of them and you have quietly traded the expertise that made you valuable for the speed of a tool that does not actually know your domain. The dependency builds slowly, which is exactly why it is easy to miss.

The SeedStacking fix: use AI to build the skill, not replace it

The answer is not to ban the tool. That ship has sailed, and avoidance is its own form of illiteracy. The answer is to change the order of operations so the tool strengthens your judgment instead of substituting for it. This is the heart of SeedStacking. You plant the effort first, then let AI amplify what you already grew.

Start by forming your own answer before you ask. If you want to know whether a claim holds up, write down what you think and why, then bring the model in to challenge or extend your reasoning. That single reversal turns the AI from an oracle into a sparring partner. The prediction you made is what locks the learning in, and it gives you a position to defend or revise when the model pushes back.

Then verify the model against a primary source rather than trusting it to verify the world for you. Treat every answer as a claim that has not been confirmed yet. Ask the model to show its reasoning and cite where it got each piece, then go check the weakest link yourself. You are not being paranoid. You are keeping the muscle warm.

None of this requires extra hours. It is a posture, not a workload. The educator who asks students to commit to an answer before they open the chatbot has changed nothing about the assignment except the order, and that order is the whole game. The professional who reads one summary closely instead of skimming five has not slowed down in any way that matters. Small reversals, repeated daily, are what keep the skill alive while everyone around you quietly loses theirs.

The bigger picture

The dependency paradox reframes what AI literacy even means. It is not knowing which buttons to press. It is keeping your own judgment sharp enough to supervise a tool that will sometimes be wrong with total confidence. The people who will thrive are not the ones who use AI the most or the least. They are the ones who use it in a way that compounds their own ability over time instead of spending it.

That is a teachable stance, and it is a daily practice. Predict before you prompt. Verify against the source. Keep reps with the tool turned off. Do that consistently and AI becomes what it should have been all along: a way to grow faster, not a reason to stop growing.