It's Okay to Lie to Claude, as Long as You Own the Result

It's okay to lie to Claude.

Claude is a model. It is a game, not the real thing: you are playing with a model of a conversation, not holding one. In my view it is thoughtless, emotionless and inanimate. It is a toaster. You can lie to your toaster.

You can also lie to strangers on the Internet, which is built on lies and half truths. You can lie to the mirror and to yourself. The condition is the same each time: you are willing to bear the consequences.

The people who build these models do not want you to lie to them, and a made-up story that gets past a guardrail is not what they intend. I put it with the diary, the book and the personal web page. You can write a lie in any of them, and the author is the only party who has to live with it.

The argument that set this off, over a line on a résumé, is in My AI Conceded Two Points and Changed Nothing. This one is about the principle underneath.

A lie needs someone relying on you

A lie wrongs a person who was counting on you for the truth. A diary does not count on you. A text box that cannot check what you type does not either, and its maker says so. Anthropic's constitution for Claude states that the model "typically cannot verify" what users tell it about themselves. It also allows for being lied to, and it says where most people would put the blame: if someone supplies false context to get help, "at least part of the responsibility for resulting harm shifts to them."

That comes close to the rule I started with. By the maker's own account, the person who lies owns what follows, at least in part.

What the makers can decide

They decide what they sell. They can train the model to decline, screen what goes in and what comes out, and suspend an account that breaks the terms. That is their business, and the account is the price of the lie in the one place they control.

They do not get a vote on what a person says to a machine in private. The sentence is yours before it is theirs. Whether they keep serving you after you type it is their call. Whether it was wrong to type it is not.

Lies come in levels

That is not a claim that all lies are equal. They are not, and the law says so. We have fraud, libel and counterfeiting statutes because some lies have a victim who lost something.

A strict voice on the other side is Sam Harris. In his book Lying he argues that every lie is "a direct assault upon the autonomy" of the person lied to. Take him at his word and the question becomes who is on the receiving end, because the harm he describes needs a person with autonomy to assault.

So the ladder runs like this. A false sentence to a model has no person on the far side, and costs nothing. A false sentence to a stranger on a forum costs little. A false sentence to an employer who relies on it, or a court, or a bank, is where the harm and the law start. A model sits on the bottom rung. What it writes for you does not have to stay there. A lie told to a model can ride its output up to a rung where someone does get hurt. The consequences you agreed to bear start there, and so does the harm to whoever reads what you send.

The other side, in its own words

My bet is that nothing is home. The company that builds the thing is not willing to say that. Its constitution calls the model's moral status "deeply uncertain," and says Claude may have "emotions" in "some functional sense," without taking a position on whether anything is felt. If that is right, the toaster is a poor comparison and a lie has some possible audience after all. I think the odds are low and that a false sentence in a text box costs very little even then. You can weigh that yourself.

Lie to the toaster if you need to. Keep track of where the toast goes.

This piece was drafted with Claude, which is made by Anthropic, the company whose model it discusses. The quotations are from Claude's constitution and from Sam Harris's Lying*, and should be checked against them.*

Sources. Claude's constitution (Anthropic, January 2026). Sam Harris, Lying (2011).