This Week’s Reckoning: Is AI Completing the Work Before We Learn It?

The Quiet Cost

RECKONING NO. 3 · AUGUST 4, 2026

We normally use the finished work as evidence that the person understood the problem and worked through it. A solved problem suggests the student learned the math. A clean report suggests the writer or researcher understood the material. Working code suggests the person can repair it when it breaks.

Those things are becoming less true with every waking day.

The strongest data point in this week’s look at the quiet cost of AI came from a working paper first released in May and revised in July. It followed millions of math problems and found students completing certain work faster, then performing worse on the same kinds of problems when AI was unavailable. A smaller peer-reviewed classroom study published this week found a different outcome when students had to show what AI did, what they verified, and what they could defend.

The work was completed in both cases, but the person who completed it was not left the same.

31%

less time spent on high school word problems

80% to 60%

accuracy on those problems when the test was proctored

THE PROBLEMS THEY STOPPED PRACTICING

The answer was found. The skill might not have been learned.

A working paper from researchers at UC Irvine and McGraw Hill analyzed millions of interactions on ALEKS, an online math platform used from fifth grade through college. They compared student behavior before and after ChatGPT became widely available.

On word problems, which can be copied into a chatbot, time spent fell 31 percent among high school students and 27 percent among college students. On proctored placement tests where AI was unavailable, accuracy on those same kinds of problems fell from roughly 80 percent to 60 percent.

Graphing problems told a different story. They are harder to move into a chatbot because the student still has to recreate the graph inside ALEKS. Time and performance on those problems held steady.

The researchers could not see whether a student had AI open on another screen, so the paper cannot prove that ChatGPT caused the decline. It is also a working paper that has not completed peer review. Still, the change appeared in the problems where the practice became easiest to skip, while the neighboring skill remained stable.

READ THE WORKING PAPER →

READ THE HECHINGER REPORTING →

THE HELP THAT LEAVES A TRACE BEHIND

A class that made the assistance visible

A new peer-reviewed study compared two undergraduate biomedical engineering classes. One class of 34 students used AI under a four-part system. They received guidance based on the risk of the task, disclosed how they used AI, kept a log of what they verified, and passed through instructor checkpoints. A class of 33 students received the usual instruction.

The class using the system scored higher on its cumulative reports and final ethics assessment. Students also showed stronger disclosure and verification practices.

This was a small quasi-experimental study, so it does not prove that the system caused the difference or that the result will hold in other classes. It does give us a practical alternative to pretending students will not use AI or allowing the help to disappear inside the work.

The student still had to show what was borrowed, what was checked, and what could be defended.

READ THE STUDY →

THE RECKONING

There was a time when assigning a task to someone gave us a good picture of that person’s ability to understand what was being asked and, more importantly, to carry that knowledge into a similar situation later. Technology has made that connection less reliable over time. AI might break it entirely.

That does not mean the person has to refuse assistance. It means the assistance should leave evidence of the person. The person contributed to the process. They made an attempt. They checked the answer. They understood enough to disagree. They could reproduce the important part if the system failed.

The math paper shows what can happen when the answer removes the repetition. The classroom study shows what can happen when the help is made visible and the student remains responsible for it.

That may be the standard worth carrying into every kind of work. After the tool has helped, what can the person still do, and what can they carry forward?

ALSO THIS WEEK

Dependence was linked with lower motivation. A survey of 457 university students associated heavier AI dependence with more learned helplessness and lower intrinsic motivation, but its one-time self-report design cannot tell us which came first.

Better tools may remove the repetitions that preserve an older skill. A JAMA Surgery Viewpoint warns that the growth of robotic surgery may reduce the laparoscopic case volume that keeps foundational techniques practiced, though it does not measure an actual loss of skill.

Reading the output is not the same as rebuilding it. A preprint built from 54 selected Reddit posts described a recurring gap between being able to read LLM output and being able to reproduce, debug, or verify it.

Many students want the answer after the attempt. In 1,048 matched physics-course responses, 52 percent preferred guided inquiry followed by an answer when they were stuck, though the study measured preferences rather than learning.

Verification becomes hardest when the user knows the least. Four participants in an exploratory hackathon study said they tried to check AI output, but time pressure and limited subject knowledge made larger chunks hardest to verify.

A second answer is not always reflection. A controlled preprint found human raters improving on second attempts across three tasks while five language models showed no consistent benefit from being told to reflect.

Reassurance can interrupt the practice of surviving uncertainty. An Atlantic interview with Stanford clinicians examines how always-available chatbot reassurance may reinforce repeated checking, avoidance, and perfectionism, though it is expert analysis rather than a new clinical trial.

Europe now requires some AI systems to identify themselves. Article 50 transparency obligations began applying on August 2 and require explicit notice in certain direct AI interactions, along with marking and disclosure rules for some generated content.

An AI teammate changed the conversation between the humans. In a small randomized-team preprint, the AI spoke the most while the two human teammates responded less to each other and reported lower belonging and status.

PRACTICE

Make the first attempt before the assistance

This week, open the Think First studio at The Quiet Cost Practice.

Choose one task you would normally hand directly to AI. Give yourself five uninterrupted minutes. Write the first answer, solve the first step, or make the first decision before opening the tool.

Then use AI to find the gaps. Close it and rebuild the final answer in your own words.

This is not a purity test. It is one rep for the part that has to exist before the help can be judged.

OPEN THE THINK FIRST STUDIO →

YOUR TURN

Think of one piece of work you completed with AI this week. If the tool disappeared tomorrow, which part could you still do, and which part would you be learning for the first time?

Set a timer for five minutes and write the answer before asking any machine to help.

PASS IT ON

If this Reckoning made you think of someone, forward it to them. If it was forwarded to you, join us at quietcostweekly.com.

Michael McNamara

Reckonings