The Part You Meant to Keep
Is There a Point Where AI Help Turns Into Harm, and Could a Scale Find It?
Few people ask whether a letter went through spell-check. The small bottle of correction fluid that once sat beside the typewriter went away with the carbon paper, and the red underline that replaced it stopped being a topic of conversation long ago. A cover letter written with a chatbot still draws suspicion, and detectors and watermarks exist to answer whether a model was involved. As models settle into ordinary revision, a different question may matter more. Did the help cost anything, and at what point?
Psychologists call this cognitive offloading, using a physical action to make a task less demanding on the mind, like setting a phone reminder instead of memorizing a date [1]. Usually that is a fair trade. So where does a convenience start to become a loss? Is there a line, and could anyone measure it?
The Cost Arrives Later
The spotter's hands rest at the hips until the bar stalls. A nudge on the last repetition finishes the set, and the lifter's chest still did the work. Lifting from the first repetition would move the same weight along the same path, and nothing about the bar would show the difference. AI help has a similar measurement problem. The output can look fine either way, and any cost tends to appear once the hands come away.
How quickly can that happen? Across three randomized experiments, people practiced fractions or reading comprehension in sessions of about a quarter of an hour, some with a chatbot beside them, then took a short test alone. The chatbot helped during practice. On the test, those in the first experiment who had used it solved 57 percent of problems, compared with 74 percent for those who had not, and they gave up more often [2]. The gap was smaller in a replication, and testing came right after practice, so the authors leave open whether the effect builds or fades with daily use.
Counting Requests Misses It
The obvious dial is quantity. Could some number of requests per task, or hours per week, mark the point of trouble? One small experiment points elsewhere. In it, 124 adults worked through logic puzzles, most with a simulated assistant whose price per request varied. Once starting ability was accounted for, request counts said almost nothing about skill gained. What tracked it was solo share, the fraction of time a person worked alone on a problem before asking [3]. Scores while assisted also overstated how people later did on their own, which suggests that a session going well is a poor guide.
A larger experiment on fraction practice gives the nearest thing to a dose curve. Each extra 10 percentage points of problems on which someone requested the full answer came with 32 percent lower odds of getting a later test question right [4]. That analysis is correlational, and it assumes a smooth trend, so a sharper break could be hiding inside it. Access alone produced no detectable harm in that study, perhaps because the assistant gave full answers only when explicitly asked. No study found in this research set out to look for a break point. For now a ramp seems at least as plausible as a cliff, which is inconvenient for anyone hoping to draw one line.
Five Rungs, Drawn in Pencil
If quantity is a weak guide, depth looks stronger. In the replication of those short experiments, people who said they mostly asked for hints scored about the same as people with no chatbot, while those who mostly asked for answers fell behind [2]. Those groups chose their own habits, so the comparison only suggests a cause.
One recent writing study offers a vocabulary for depth. It separates four layers, surface (grammar and local wording), structural (outlines and organization), idea (claims and examples) and reasoning (evidence interpretation and argumentative logic), with coders rating each from 1, where the writer did the work, to 5, where the tool did it and little changed. Over an eight-week course, heavier offloading at every layer went with weaker thinking on later no-AI tasks, and the cost of open access ran most faintly through the surface layer and most strongly through reasoning [5]. Open-access writers also sent nearly twice as many prompts per task, yet far fewer of their revisions were writer-led, so more turns may not mean more involvement. Six classes without random assignment can only suggest so much, and the author says the framework still needs independent validation.
Adding memory at the bottom gives a rough ladder. It was assembled for this article, and no validation stands behind it.
| Rung | What moves to the tool | Usually helps when | May start to cost when |
|---|---|---|---|
| 1. Storing | Dates, lists, references | The facts can be looked up again | Recall is the skill being built |
| 2. Polishing | Spelling, grammar, local wording | The meaning is already settled | Checking stops, or edits drift into meaning |
| 3. Arranging | Outline and order | The person knows the material | The person is still learning how the pieces fit |
| 4. Supplying | Claims, examples, solutions | The task is routine | Working it out was the point |
| 5. Deciding | Interpretation and judgment | Stakes are low and the call is easy to check | The person will answer for the conclusion later |
Even the second rung has a price. In a 2005 experiment, people with stronger verbal skills left nearly twice as many unflagged errors behind when a spelling and grammar checker was switched on [6].
Read across the rows and the threshold stops being a place on the ladder. One working definition might be that help turns to harm at the rung where the tool starts doing something the person wanted, or needed, to still be able to do alone. A status email can sit at rung five with little lost. A junior engineer at rung four on debugging may be giving away what the job was supposed to build, the worry behind When Code Gets Cheap, Judgment Gets Expensive. Same ladder, opposite verdicts, so a useful scale would need to ask about intent as well as behavior.
The Closest Thing to a Number
Could the rungs be measured automatically? The most direct attempt found here is the Offloading Score. It reads screen recordings and keystrokes, has a language model rebuild how an average person would finish each AI-assisted step without the tool, and reports the fraction of steps saved, between 0 and 1. Among 40 freelance developers, a one-hour limit produced more offloading than a four-hour limit, 0.45 against 0.31, while self-reported mental load barely moved [7]. People may not feel the shift as it happens.
The authors also sketch an actual line, an overreliance threshold of 0.33 tied to how well developers could recall their own system. They call it illustrative and say real thresholds should be set empirically. The link to recall was weak overall and stronger only after setting aside about 30 percent of users who combined moderate to high reliance with high recall, apparently using the tool to learn [7]. Those users could cross that line with no visible harm, and they may be the most interesting people in the study.
Questionnaires take the other route. A free self-check called PAUSE sorts readings into minimal, mild, moderate and strong bands, and its author is candid that the cut points were chosen for legibility, with no reliability or validity evidence behind them [8]. Levels are easy to print. Knowing what they mean is the hard part.
From a machine learning standpoint, this resembles choosing an operating point on a classifier. A score can come from logs, but the cut depends on what each kind of mistake costs, and those costs differ between an expense note and a thesis. The ground truth, what someone can do later with the tool gone, is slow and expensive to collect, and several studies above tested within minutes of practice. Then comes the feedback loop. Once anyone is judged on the number, prompts will likely get padded, the pattern When Optimization Optimizes Itself traced in systems that learn to satisfy their own metrics.
Right Use, Wrong Use
Until a validated scale exists, the ladder can still sort everyday choices. Right use leaves a person able to do what they wanted to keep. Wrong use gives that ability away, often with little sign at the time. The labels depend on that goal, so the same request can land in either column. These pairings follow from the evidence above, and no study has tested them as a set.
| Task | Right use | Wrong use |
|---|---|---|
| Grammar and wording | Polishing a draft with settled meaning, then checking the edits | Accepting every change unread, even when the meaning shifts |
| Learning something new | Trying first, then asking for a hint | Asking for the full answer before trying |
| Organizing work | Outlining familiar material | Letting the tool arrange material still being learned |
| Ideas and arguments | Getting help on routine work that can be checked | Submitting an argument the writer could not explain |
| Decisions | Handing off low-stakes calls that are easy to check | Handing off a conclusion the person must answer for later |
A rough test fits in one question. After the session, could the person explain, defend or reproduce the parts of the work they meant to keep? A yes suggests the tool supported them. A no may be fine for a status email, as long as the trade was made on purpose.
A Bubble Between Two Lines
A bubble between two lines straightens nothing. It tells the person holding the level where things stand, and the rest is up to them. That may be the more realistic job for a scale, and there is some early support for it. In the larger fraction experiment, some participants read a short note between problems describing what they had just done and what it meant for the practice they got, with no grades, points or instructions. Their odds of requesting a full answer fell by about half and they did better on the unaided test, while a points reward for effort changed neither [4]. The margins of error were wide, and it was one brief session with an immediate test, so durability remains unknown. It also hints at something awkward for anyone building a scale. The note seems to have worked without a score.
Spell-check seems to have stopped being a question once people had decided, mostly without discussing it, what they no longer needed to do themselves. That decision has not been made for reasoning, and it may not get made the same way for everyone. A scale might never find one point where help becomes harm. Could it still show a person where they stood on a given afternoon, close enough to change what they hand over next time?
References
- E. F. Risko and S. J. Gilbert, "Cognitive Offloading," Trends in Cognitive Sciences, vol. 20, no. 9, pp. 676–688, 2016. DOI: [Online]
- G. Liu et al., "AI Assistance Reduces Persistence and Hurts Independent Performance," arXiv, 2026, [Online]
- S. Wu et al., "How AI Assistance Affects Human Skill Development: A Study of Learning with Logic Puzzles," in Proc. ACM Conference on Human-AI Complementarity and Alignment (HCOMP 2026), 2026. DOI: [Online]
- S. Maier et al., "Designing Against Deskilling: Metacognitive Feedback Reduces Cognitive Offloading to LLM Assistants," arXiv, 2026, [Online]
- X. Chen, "Layer-sensitive cognitive offloading in generative AI-assisted writing: supported performance and independent no-AI outcomes," Frontiers in Psychology, vol. 17, art. 1906199, 2026. DOI: [Online]
- D. F. Galletta et al., "Does Spell-Checking Software Need a Warning Label?," Communications of the ACM, vol. 48, no. 7, pp. 82–86, 2005. DOI: [Online]
- V. Padmakumar et al., "Offloading Score: Measuring AI Reliance Through Counterfactual Workflows," arXiv, 2026, [Online]
- M. U. Alam, "PAUSE: A Privacy-Preserving Self-Reflection Tool for AI-Associated Cognitive Offloading," arXiv, 2026, [Online]
Predicting the Next Second: Tile Coding as a Real-Time Prediction Engine
Tile coding is usually taught as a classroom exercise in reinforcement learning. Its most practical use may be prediction, helping a robot or a prosthetic arm learn to anticipate what happens next while it runs, on modest hardware and with no separate training phase.
When Code Gets Cheap, Judgment Gets Expensive
Agentic coding tools have made writing code far cheaper. The controlled evidence suggests the bottleneck moves rather than disappears, toward specifying, checking and maintaining, and the engineering skills worth building move with it.
Working on something related?
Thinkata builds production AI systems of the kind these articles discuss. A 20-minute call is the fastest way to find out whether it fits.