AI can be a ramp. It can also be a chauffeur who grabs the wheel during a driving lesson.
Both make something difficult easier. Only one helps the person develop the skill they came to practice.
That difference matters more than the usual argument about whether AI is “good” or “bad” for us. The better question is what kind of help the tool provided—and what the person could still do after it left.
What Ten Minutes With AI Actually Changed
A Berkeley News headline makes a wonderfully sticky claim: use AI for ten minutes and your persistence starts to erode.
The research underneath it is interesting. The headline is also carrying more weight than the evidence can comfortably hold.
In a paper presented at COLM 2026, researchers ran three randomized experiments involving 1,222 recruited participants. Two experiments used fraction problems. The third used SAT-style reading comprehension.
Participants with AI performed better while they had access to it. Then the researchers removed the assistant without warning. On the remaining questions, the groups that had used AI performed worse than the control groups.
That is a real result. It is not proof that ten minutes with a chatbot causes lasting cognitive decay.
The effects also varied. The first fraction experiment found a moderate performance difference. A larger replication found a smaller one, and its overall difference in skipped questions did not reach statistical significance. The reading experiment found both lower solve rates and higher skip rates after the AI disappeared.
Researchers noticed another useful pattern. Participants who asked for direct solutions later performed worse and skipped more questions than those who asked for hints or clarification.
But this part was not randomized. People chose how to use the tool. The finding shows an association, not proof that requesting a direct answer caused the later difference.
In other words, “erosion” is doing excellent headline work. The study itself shows something narrower: after a short period of AI assistance on specific tasks, participants performed worse on several immediately subsequent questions without it.
That is less dramatic. It is also useful.
The Tool Can Remove a Barrier or Steal a Repetition
If I am learning fractions, an AI that supplies every answer is a very expensive answer key. It may get me through today’s worksheet. It does not give me the repetitions I need for tomorrow’s test.
If I am learning to troubleshoot a system, pasting commands I cannot explain may close one ticket while leaving me helpless when the next server starts making a different alarming noise.
That difficulty is part of the work. Remove all of it, and you remove the practice.
But plenty of difficulty teaches nothing.
Typing through pain does not make me a better writer. Reconstructing a half-finished argument after a caregiving interruption does not improve the argument. Fighting an inaccessible interface does not deepen anyone’s subject-matter expertise.
In Using AI Is Not the Same as Surrendering Judgment, I wrote about AI as an accommodation. It can transcribe speech, organize scattered notes, help someone resume interrupted work, or conserve energy for the decisions that actually require them.
An AI that helps put my existing thoughts into words removes a barrier. An AI that supplies the thoughts before I form them replaces the practice.
The tool may look identical from the outside. The job it is doing is not.
Ask What the Person Should Still Be Able to Do
Teams often ask whether a task can be automated before asking why a person is doing the task in the first place.
Are they trying to learn? Produce? Practice? Gain access? Demonstrate understanding? Make a decision someone will have to defend later?
The purpose changes what helpful assistance looks like.
When learning is the goal, do not hand over the finished reasoning before the person has made an attempt. Offer a hint. Point to the relevant documentation. Identify the step where the logic went sideways.
When access is the goal, remove the obstacle. Do not demand pain, fatigue, or interface gymnastics as proof that the work is authentically human.
When production is the goal, let AI handle useful scaffolding—but name the judgments that remain human. Someone still needs to verify the facts, resolve ambiguity, and accept responsibility for the result.
When the stakes are high, preserve the trail. Keep the sources, review authority, approval boundaries, and override decisions visible.
The question is not “Did AI make this easier?”
It is “What did AI remove—and was that the thing the person needed to practice?”
Give People Help Without Hijacking the Task
Most AI assistants have golden retriever energy. Ask for one tennis ball and they return with the entire sporting-goods aisle.
That enthusiasm works when the goal is fast completion. It works less well when the person needs to learn, think, or retain control.
A better system would offer graduated help. It could ask what the person already understands, provide a hint before a solution, surface documentation before generating an answer, or wait for an attempt before offering feedback.
Sometimes the best assistant should hold the flashlight instead of taking the engine apart.
Documentation and training teams have a role here. We need to separate content that supports immediate action from content that teaches durable understanding. We need clear prerequisites, decision points, common misconceptions, and escalation paths. We need source material reliable enough for the system to guide someone without inventing facts or smoothing uncertainty into nonsense.
A cheerful interface cannot rescue ungoverned knowledge. It can only hide the mess more politely.
Keep the Capability, Lose the Unnecessary Friction
The point is not to make work harder in the name of character building.
Pain is not pedagogy. Inaccessibility is not rigor. Busywork is not exercise for the soul.
But instant completion has its own cost when it leaves someone one outage, subscription change, or unavailable chatbot away from helplessness.
Good assistance should leave people with more access and their agency intact.
We should remove the barriers that keep people from using their capabilities. We should preserve the practice that creates those capabilities. And we should judge AI assistance by more than the answer it produces in the moment.
After the tool leaves, what can the person still understand, explain, decide, and do?