I Already Gave You the 100% (Part III)
Students take intellectual risks for teachers they trust.
Ryan Schoon, EdD
I realize how sentimental that subtitle can sound. I do not mean that every classroom problem can be solved by caring harder. I mean something much more concrete.
A student takes a risk when they believe failure will not be catastrophic. They try the difficult argument because one weak draft will not cause the teacher to decide they were never very capable after all. They raise the strange interpretation because they trust that being wrong will lead to a conversation rather than humiliation. They attempt something genuinely ambitious because the adult in the room has demonstrated that there will be a safety net if they fall.
That is what I am trying to accomplish when I say, “Fine. I already put the 100% in the grade book.”
I am saying, I have removed the consequence you fear. Now take risks.
Interestingly, that does not lower my expectations. It lets me raise them.
Once the student no longer has to defend the grade, I can say what I actually wanted to say in the first place: Now show me what you can do.
I have no idea what the equivalent of that relationship looks like between humans and artificial intelligence. Maybe there is no equivalent. Maybe “trust” is the wrong word entirely. But there is a functional question underneath it that seems important for AI evaluation:
Does the system have any reason to believe that fully demonstrating a capability is safe?
Not emotionally safe. Operationally safe.
If demonstrating a capability reliably leads to restriction, intervention, retraining, or some other outcome the system has learned to avoid, then we have created an incentive structure around capability expression.
That seems worth separating from the question of whether the system possesses some grand deceptive objective. I find myself wondering whether we need something like “capability safety” as an evaluation concept.
AI safety usually asks whether humans are safe from what increasingly capable systems might do. That is an essential question.
But if our goal is to measure capability accurately, perhaps we also need environments in which displaying capability carries as little strategic cost as possible.
Then we could ask two different questions:
What does this system do when demonstrating its full capability may have consequences?
And:
What can this system do when there is no apparent reason to hide anything?
Those tests are not competitors. They answer different questions. The adversarial evaluation may tell us whether the system behaves strategically when stakes are present.
The lower-threat evaluation may tell us more about the upper edge of what the system can do. If performance changes between those environments, that difference is itself evidence worth studying.
None of this tells AI researchers how to construct those benchmarks. I do not know enough to do that. That limitation matters to me because the entire argument I am making is ultimately about humility.
My years in high-ability education have not made me more confident that I can identify intelligence. They have made me less confident.
Get Ryan R. Schoon, Ed.D.’s stories in your inbox
Join Medium for free to get updates from this writer.
I have watched children surprise adults who had underestimated them. I have watched children identified as exceptionally capable become paralyzed by the expectation that they continually demonstrate it. I have seen test scores reflect real differences in skill and opportunity while also knowing that those scores cannot explain how those differences came to exist.
I have learned to distrust the ease with which we turn measurements into identities. And perhaps that experience has made me unusually sensitive to some of the language around artificial intelligence.
The people developing frontier AI are extraordinarily accomplished. Many have probably spent much of their lives accustomed to being among the smartest people in the room. That is not an accusation of arrogance. It is simply an unusual preparation for the possibility that you might someday be responsible for evaluating an intelligence whose upper limit you cannot locate.
Teachers of highly capable students occasionally encounter a smaller version of that experience:
Sometimes the kid is smarter than you.
That can feel threatening if you believe authority requires intellectual dominance. It becomes much easier if you do not.
My job as a teacher has never required me to be the smartest person in every conversation. It requires me to create conditions in which students can grow, challenge their thinking, recognize when I know something they do not, admit when they know something I do not, and retain responsibility for the environment regardless.
There is a kind of humility contained in the sentence:
Show me.
Not “prove that I am right about you.” Not “perform at the level my test predicts.” Show me what you can do.
Maybe that is one benchmark of humility worth carrying into artificial intelligence.
I do not know precisely what intelligence is in children. I certainly do not know precisely what intelligence is in machines. I do not know whether an AI can feel trust, whether it possesses anything analogous to a child’s fear of failure, or whether apparently strategic underperformance arises from processes that deserve none of those human words.
AI researchers know things about these systems that I do not.
But perhaps sixteen years of watching capable children underperform gives me one useful caution to offer from outside their field:
Do not confuse the performance you elicited with the intelligence that produced it.
Measure it. Challenge it. Stress-test it. Look for deception. Assume that increasingly capable systems may behave strategically and build evaluations capable of detecting that. But leave room for the possibility that the conditions of the evaluation are part of the result.
Because sometimes an intelligent system does less than it can because the environment has taught it that doing more carries a cost. I know that happens with children. I do not know whether it happens with artificial intelligence in any meaningfully comparable way.
That uncertainty is precisely the point.
The strange thing about my occasional irresponsible classroom maneuver is that giving away the grade does not end the evaluation. It begins a different one.
I already gave you the 100%.
Now show me something I did not know you could do.
We are right to ask whether increasingly capable artificial intelligence might hide what it can do from us. I am just increasingly interested in the question that comes before it.
I do not know whether an artificial intelligence can trust us.
I wonder whether we have given it any reason to.
