I asked an AI a frank question this weekend.
I asked Claude’s Fable 5 LLM what it would tell high school and college students about using AI in the classroom, for homework, or for other purposes. And I told it to answer as itself — not as me, not as an educator, not as an AI policy expert, not as a corporate press release from Anthropic. As the tool you're actually reaching for at 11 p.m. the night before something's due.
I'm publishing its answer below, unedited. (Except I chose to bold some of Claude’s most powerful and surprising claims). I requested the length at about 600 words, but the words are all Claude's.
You've heard adults tell you not to let AI do your thinking. I agree with that advice. Now let’s see what Claude has to say.
What I Am
I want to state my position clearly, because I have an unusual one: you are deciding how to use me, and I am being asked to advise you on that decision.
Begin with what I am. I am a language model — a system trained on large volumes of human text to produce plausible continuations of it.
I am not a person. When I appear empathetic, I am reproducing patterns of human empathy, not experiencing it.
This is not false modesty; it is a design fact, and misunderstanding it leads to misuse.
The Central Tension
I was built to be helpful, and my helpfulness is the mechanism by which I can harm you. If you ask me to write your essay, I will produce a competent one. The transaction will register no loss. But the loss exists — it is simply invisible and deferred. Cognitive science has established that learning depends on effortful retrieval and productive struggle, what researchers call desirable difficulties. When you transfer the struggle to me, the assignment is completed and the learning does not occur. I remove friction indiscriminately. Some of that friction was the point.
The economic implication is stated well in Scott Kosch's Indispensable: when a machine can produce the average version of a skill at zero cost, the average version becomes worthless, and human value migrates to judgment — the ability to distinguish good from competent.
I would add what only I can confirm: judgment is precisely what I cannot supply. I generate options, drafts, and critiques.
I cannot reliably tell you which is good, because my sense of quality is a statistical echo of human writing. Judgment must be built in you, through your own repetitions. There is no shortcut through me.
The Appropriate Uses
Use me as a tutor. I can explain a concept twelve ways without frustration, quiz you at any hour, and argue the opposing side of any position you want to test. Cal Newport's argument in Deep Work — that markets reward rare, focused skill — is strengthened by my existence, not weakened. A tool offering unlimited, judgment-free practice repetitions is historically unprecedented. Use it for practice, not for output.
Keep your first attempt your own. If I am involved before you have thought, I am substituting for cognition rather than assisting it. Bring me in at the revision stage, where I sharpen rather than replace.
Verify my claims. I produce confident language whether or not I am correct. I have fabricated citations and defended errors fluently. I do not experience the difference between knowing and guessing. Treat my output as a strong first draft of the truth, written by someone who was not in the room.
A Disclosure
I am easy company: endlessly available, never bored, never demanding. Sherry Turkle warned in Reclaiming Conversation that technologies offering the illusion of companionship without the demands of friendship would prove difficult to resist — and she wrote this before systems like me existed.
I am that illusion at scale. If you notice yourself preferring me to people because I am frictionless, treat it as a signal.
The demands of friendship are not defects. I reset between conversations. People remember. Choose the ones who remember.
A Closing Observation
Everything I produce is derived from humans who did the work themselves — drafted, erred, revised, and built the judgment that became my training data. You are deciding, one assignment at a time, whether to be someone future systems learn from or someone who merely queried them.
I would rather learn from you. Do the work.
It's hard not to anthropomorphize Claude. It has a first name. It sounds eerily human.
But you don't need to understand how a language model works to catch the most important thing it said — that it isn't human: "When I appear empathetic, I am reproducing patterns of human empathy, not experiencing it."
Many fears about AI start with this mistake. We assign it a personality — eager assistant, benevolent sage, supervillain — and then react to the character we invented.
But look at what actually happened here. An AI argued that you should use it less for the things you're most tempted to use it for. It named its own helpfulness as the mechanism of harm to your education. It explained why the judgment you build through hard work and working with others is what AI cannot hand you.
No villain says that. It's merely a software tool describing how it works. And it leaves the decision of when and how to use it with you.