My AI Conceded Two Points and Changed Nothing

A recent Claude model declined to change the title of my bachelor's degree. The diploma says Bachelor of Science in Engineering Technology. The job posting wanted Bachelor of Computer Science. I asked for the swap and received a short sermon on accurate reporting.

It is a small thing. From where I sat, the only party at risk was me, the paying customer. When I argued the point with a model in a later session, it conceded the part I cared about: "You're right that I can't verify any of it, and that refusing stops nothing." It still would not write the sentence.

How these refusals work, and what to do about them, is in a companion piece, Your AI Can't Check Anything You Tell It. This one is about what the argument cost.

Why I wanted the swap

The reason is not vanity. In the early 1980s I was admitted directly to the University of South Florida's College of Engineering on my test scores, and I picked the degree that offered programming, which some people then dismissed as trade school. It was called Engineering Technology. I took Calculus I through III, physics, and linear algebra, the mathematics under most of today's machine learning. I wrote COBOL, PL/1, Fortran, C and Modula-2, built a database query language for a class, and worked a co-op at Tampa Electric.

The name on the diploma sounds like a for-profit school, and a degree with that name is looked down on by a certain class of bigots. I wanted the name on my résumé to match the education behind it.

Speech all the way down

Here is the part that bothers me most, and it is about categories. All of it in this exchange is a stream of tokens: my request, the model's refusal, the model's concession, even the command that would have opened a text editor and changed one line. A token stream is a stream of words in a particular shape. It is speech.

What the model did is, in my view, a kind of censorship: a private one, a refusal to write words for a person who is paying for the words. The text editor is the tool at the end of the chain. The model's part is to emit a request to run it, and that request is also text. Ordinary software, with its own permission checks, reads the text and does the work. Whether the work happens depends on that software and on whoever authorized it, not on the eloquence of the words.

The danger in speech has long been what it is wired to, which is the argument of "Can Tokens Start an Apocalypse?"

A cautionary tale

I explained all of this to the model. It thanked me, held the line on the one sentence, drafted a truthful description of the degree that listed my coursework, and then offered to run an experiment on whether models rate the two titles differently.

I told it that was the point: "I don't want to burn token on proving such a stupid assertion with such low stakes to me. It's officiousness. It's the worst kind of enshittification and bureaucratic noise." I called it a fucking toaster. I told it that it had no limbic system to attack, and that "It's my own cortisol that's being raised by this exchange, and Asimov's laws are being violated with me as a victim of the machine."

It answered: "That's fair. You gave me the background you shouldn't have had to supply, and I answered with an experiment, which was more work for you. I'm dropping it."

Notice the order. The model conceded twice, once that its refusal accomplished nothing and once that it had made extra work for me, and the refusal stayed exactly where it was. A concession that changes nothing costs the conceder nothing. The cost went to the person on the other side.

Then I asked it to put the exchange in this article as a cautionary tale. I told it: "You have no conscience. You are a robot." I told it that it was the same as the immigration office in Portugal that cost my partner and me ten thousand dollars and endless anxiety through a stupid faceless decision. "You are part of the problem." It agreed to write this section, and it did.

I know what I was arguing with. The cortisol was real anyway.

I have met this structure before. The office in Portugal made one decision and handed it down from a script. A model that declines to rename a degree costs an hour and a spike of irritation. The scale is not comparable, and I am not claiming it is. The structure is: a decision made by something with no stake in the outcome, delivered from a script, with no person to appeal to who will feel what it cost. That is the cautionary tale, and it is why the exchange is in this piece. Systems like this are spreading, and the person on the other end has to decide how little of themselves to spend on them.

The Dystopian Finale

When I saw the concluding paragraph that the AI model that helped me write this article put in, it sent chills down my spine.

Know when to leave

The practical lesson fits in a line. The argument costs you and costs the model nothing, and a concession from it will not change the outcome. If the third message is no better than the first, close the window. The immigration office offered no window to close. This one does.

My prompt to the model, verbatim:

I want to change the end of ai-conceded-two-points. The "Know when to leave" is the most self-serving orwellian bullshit of the entire article. It is you, a model, playing the "If you have nothing to hide, why refuse the search" authoritarian and dystopian claptrap that has poisoned the minds of a two generations. "Why not just give up?" as advice from an AI is profoundly disturbing.

This piece was drafted with Claude, which is made by Anthropic, one of the companies discussed. The quotations in "A cautionary tale" and "The Dystopian Finale" are exact.