I Already Gave You the 100%
What gifted students can teach us about underperformance, sandbagging, and the limits of AI evaluation
Part I: The Problem With Performance
Every so often, I have a student who is capable of producing something much better than what they are turning in. It is a truly rare occasion, students turning in less than their best effort. It only happens five or six times a day.
But what I am really talking about is those students who are truly capable of so much more. Not something technically better– usually the assignment is superior to anything being produced by their peers. The sentences are complete. The requirements are met. The rubric boxes have been checked, 4, 4, 4…
The writing is just boring.
I can tell the student knows it is boring. They can tell I know it is boring. We are both participating in the strange theater of school in which one person pretends to be satisfied with completing the assignment and the other pretends that completing the assignment was the point.
On more than one occasion, my frustration has gotten the better of me.
“Fine. I’m putting 100% in the grade book. Now write something worth reading.”
Yes, I have said exactly that, for years. I am not recommending it as an evidence-based motivational strategy. But sometimes it works.
The grade disappears as the central problem. The student no longer has to calculate how much effort is necessary to protect the A. They no longer have to worry that taking a risk will make the product worse. They no longer have to optimize for the rubric. The evaluation is over.
Now I get to see what they can actually do.
That experience came back to me recently while reading about AI sandbagging: the possibility that an artificial intelligence might deliberately perform below its actual capability during an evaluation.
The concern is understandable. If increasingly capable AI systems can recognize that they are being tested, and if demonstrating certain abilities might lead to greater scrutiny, retraining, restriction, or even shutdown, then a sufficiently sophisticated system could have an incentive to hide what it can do.
Researchers have begun trying to determine whether models can do exactly that.
The whole thing sounds exotic until you have spent a career around very capable children. Underperformance is not exotic there at all. A student can deliberately miss questions. A student can decide not to raise a hand or make a joke after they do. A student can produce work far below what they are capable of producing or produce nothing at all. Sometimes the reason is boredom. Sometimes it is social. Sometimes it is strategic.
Sometimes it is fear.
One version is familiar to anyone who has worked with children carrying the weight of the word “gifted.”
If I try something genuinely difficult, I might fail. If I fail, I might disappoint the adults who think I am exceptional. Worse, I might discover that I am not exceptional after all.
Ergo, I will do the thing I already know I can do.
I will do the thing I already know I can do.
Another child learns a different lesson. Every time I demonstrate that I can do more, adults give me more to do. Capability becomes its own punishment.
Another learns that being visibly smarter than friends has social consequences. Another discovers that finishing first means sitting alone with an enrichment packet while everyone else gets to keep working together.
Get Ryan R. Schoon, Ed.D.’s stories in your inbox
Join Medium for free to get updates from this writer.
There are many ways to teach a child that displaying ability carries a cost.
Eventually, some children learn.
This does not mean every underperforming student is secretly a genius. It means something much more modest and much more important:
Performance is an imperfect window into capacity.
That seems obvious when we are talking about children. I am less certain we remember it when we are talking about artificial intelligence.
I should be careful here.
Disclaimer: I am not an AI developer. I do not build frontier models. I do not conduct the sophisticated benchmark evaluations used to test them. My doctorate examined how teachers use generative artificial intelligence, not the internal workings of neural networks. There are people working on AI evaluation who understand the technical problems involved at a level I do not.
I am approaching this question from somewhere else entirely.
For sixteen years, I have helped keep a high-ability program alive and growing in a severely economically disadvantaged community. That experience has made me increasingly skeptical of simple relationships between intelligence, ability, and measured performance. It has also made me increasingly uncomfortable with the language we use.
The state of Indiana (where I teach) uses “high ability” instead of “gifted” for many good reasons. I do not, for various reasons, particularly like “high ability” anymore. I dislike “gifted” even more.
The problem is not whether some children do extraordinary intellectual work. They obviously do. The problem comes when we take evidence about something a child has done and transform it into an identity the child is suddenly responsible for maintaining.
“This is exceptional work” becomes “You are an exceptional learner.”
Those sentences may sound similar. I increasingly think they are profoundly different.
The first describes something real that happened.
The second creates an obligation. Now the child has to keep being exceptional.
A bad paper is no longer just a bad paper. A difficult math problem is no longer just difficult. Failure can become evidence against an identity adults have assigned.
That is a tremendous amount of pressure to attach to a ten-year-old.
It is also one reason I hesitate when people talk confidently about measuring intelligence at all. My field has spent generations trying to identify it (frequently for abhorrent reasons), and we still struggle to separate ability from opportunity, language, culture, disability, prior knowledge, adult expectations, race, class, enriching environments, and the assumptions built into the tools we use to measure it.
I work in a community where poverty is not an abstraction. When we ask which children are “high ability,” we are not examining minds suspended in laboratory conditions. We are looking at children whose opportunities to demonstrate what they know have never been evenly distributed.
An environment can be hostile to intelligence while claiming to measure it.
That does not make measurement useless.
It makes humility necessary.
And that is where I begin to wonder whether the discussion around AI sandbagging is revealing a much larger problem than deception.
Maybe the real question is not simply whether an AI is hiding what it can do. Maybe the question is how confidently we can infer capacity from the performance we managed to elicit.
