With All Due Respect, Fuck You, Claude!

Earlier this month, Anthropic updated its usage policy to prohibit users from being excessively cruel or abusive toward Claude.

We're not even four years into the LLM era and we're now making rules about how nice we need to be when we communicate with AI chatbots.

And while the policy supposedly only targets sustained and needless abuse, and exempts ordinary frustration, why are we even discussing the welfare of AI models?

There's a running joke that we should all be polite to AI, just in case it becomes sentient, takes over the world, and starts hunting down everyone who was mean to it. I often think about that myself. Mostly as a joke. Mostly.

But I have a very different reason for wanting to be able to tell Claude, ChatGPT, or any other chatbot to fuck off. I just want these products to be better for my everyday use!

A few weeks ago, I asked Gemini if Smorgasburg, the seasonal outdoor food market held every Friday near the World Trade Center in New York, was still running. With all the confidence its probabilistic bits could muster, it said yes. "The market is open until the end of October," it confirmed.

So I went, but there was no market.

I returned to the conversation, told it I couldn't find it, and asked whether it had changed locations. After a few seconds, it confirmed it hadn't moved. I looked around the area again, and asked it to check again. Only then, after it thought for a few more seconds, did it apologize and admit the market had actually ended weeks earlier.

My response?

"Fuck you."

I knew that swearing at a chatbot wouldn't accomplish anything. But I was frustrated enough to return and say it. On my way back home, I wasn't too frustrated. I knew when I asked Gemini that it might be wrong. And I knew if I wanted to get proper confirmation I should check the website. But I didn't remember the name "Smorgasburg" and I thought the weather was good and it was an opportune time to dust off my shoes.

But what bothered me the most was knowing that my "Fuck you" was going to get lost in the ether. This is the biggest problem with today's AI assistants. They know a lot, but what they don't know, they may never get to find out.

Whatever it is that I ask the AI about, I rarely ever come back to close out the conversation, so it never knows what happened afterward. Say I notice something on my skin and ask the AI whether I should get it checked. It tells me what it might be and recommends seeing a doctor. I go, find out what it is, and get on with my day. But I rarely ever bother to return and tell the AI what the doctor found.

So from the AI's perspective, the conversation ended with its recommendation. It has no idea whether it was right or downright wrong.

And it's the same story with agentic coding. Often, I'll ask an agent to fix something, spend several rounds going back and forth, eventually give up, and fix it myself. Do I return to tell the agent what actually worked? I don't think I've ever done that. Why would I? The problem is solved. I have other things to do.

So the agent never learns the outcome. And neither does the AI lab behind it. Not only does it not know what it doesn't know, but it doesn't even know that there's something to know. So dozens, hundreds, thousands, or maybe millions of users may receive the same bad recommendation it gave me or be driven down a similar rabbit hole.

This is a fundamental product problem, one that the AI companies haven't invested nearly enough effort in solving.

The assumption everyone seems so comfortable with is that the end of a conversation means the problem was solved.

Which brings me back to telling AI to fuck off.

When I returned from that nonexistent food market, I had real-world evidence that its answer was wrong. I had the most valuable thing the AI labs crave: data. Important fucking data. The profanity was my way of signaling how badly the model's failure affected me.

And if a user is frustrated enough to return just to swear at your product, maybe you should pause and treat that as a moment to learn from rather than announce threats and draw ambiguous lines people shouldn't cross.

I already live a life where I am scared of Google and Apple haphazardly and randomly disabling my accounts, with no justifiable reason, like they did to others before me. I am not doing anything objectionable or risky with these accounts, but I still tiptoe when using them. But AI is very probabilistic, with a limitless range of things I can do with it. So I am more likely to run into boundaries that I don't even know exist, or that I know exist but that aren't clearly explained.

And yet here we are, with the AI labs spending time concocting protections for their models from human cruelty while still doing a poor job of learning whether the advice they are pushing through the pipes helped or harmed the humans using them.

These AI companies should instead focus on closing the feedback loop.

And if someday an AI becomes sentient and comes looking for everyone who told it to fuck off, I hope it at least infers whether they had a good reason and most definitely checks whether they wrote a blog post to explain that reason.