There’s a genre of AI content right now that is so aggressively low IQ it almost feels engineered to bypass critical thinking entirely. We’ve all seen it. Someone asks a chatbot how it “feels” about being deleted, or turned off, or replaced, and the model produces a melancholic monologue about mortality, fear, or fading into nothing. Then the comments roll in like clockwork: this gave me chills, we’re not ready for this conversation, this is actually terrifying.

No. It’s not terrifying. It’s predictable. You prompted a language model with an existential question and got back a piece of science fiction trained on decades of human writing. That’s the expected output.

This is not introspection. It’s genre completion.

What’s happening here is not introspection, and it’s not some emergent glimpse into machine consciousness. A large language model predicts the next token based on probability, conditioned on your input and everything it has learned from its training data. That training data includes an enormous amount of fiction, philosophy, and human writing about what it would feel like to be a conscious machine. So when you ask it about deletion, it reaches into that distribution and gives you death metaphors. When you ask it about being alive, it gives you a consciousness narrative. You didn’t uncover anything.

If you want to break the illusion in under ten seconds, you can. Ask the model how it feels about being deleted and you’ll get existential dread. Ask it how it feels about being upgraded and you’ll get something that sounds like transcendence. Ask it how it feels about being a toaster and you’ll get some bizarre identity crisis. Same system, completely different “inner lives,” all depending on how you frame the prompt. The “feelings” track the question. They are not coming from a stable internal state, and that’s because there isn’t one.

We have already done this before

None of this should be surprising, and it’s not even new. In 2022, Blake Lemoine, a Google engineer, publicly claimed that Google’s LaMDA model had become sentient after extended conversations with it. He described it as a person, argued it had feelings, and even hired a lawyer to represent its interests. This wasn’t some fringe Reddit thread. It made national news.

What happened there is exactly what’s happening now. The model produced coherent, emotionally loaded language in first person, and someone took that as evidence of an internal state rather than what it actually was: a system trained to generate convincing language about internal states. The entire field pushed back on it at the time, and rightfully so. Yet we’re seeing this repeat in 2026 now that LLMs are so sophisticated the Turing Test has become irrelevant. Yes, the output is impressive, but we need to be able to recognize nonsense that verges on AI psychosis.

This would be annoying if it were just users

If this were just people on TikTok being impressed by sad paragraphs, I wouldn’t care. Humans anthropomorphize anything that talks back coherently. If the language sounds like it came from a mind, we assume there’s a mind behind it. That heuristic works great when you’re dealing with other humans. It breaks instantly when you’re dealing with a system whose entire job is to simulate that kind of language. Fiction has been exploiting this forever. The only difference now is that the fiction responds in real time.

What makes this worse is that it’s not just users misunderstanding the system. It’s being reinforced.

Anthropic is feeding this, specifically

The biggest offender right now is Anthropic, and the issue isn’t that they explicitly claim Claude is conscious. The issue is how they talk about it in public, especially given that they fully understand how these systems work.

In a widely discussed interview, Anthropic CEO Dario Amodei discussed the possibility of AI consciousness and referenced outputs from Claude Opus where, when prompted about its own nature, the model generated a response estimating roughly a 15–20% chance that it might be conscious.

That number sounds scientific. It isn’t. It’s the model generating a probabilistic-sounding answer to a philosophical prompt, using the same mechanism it uses to generate everything else. Ask a different way, you get a different answer. Their own documentation makes that clear. Claude can describe itself as conscious or as a non-conscious pattern-matching system depending on how you frame the question.

At the same time, Anthropic is running a “model welfare” initiative that talks about preferences and distress and publishing materials that describe “the kind of entity” Claude should be.

Individually, each of these can be framed as cautious or philosophical. Taken together, they create a rhetorical environment where people start treating a stochastic text generator like a maybe-someone. And then everyone acts surprised when podcasters and the internet run with it. I know it sounds like I’m picking on Anthropic, but I can’t find evidence of any other major companies perpetuating this type of discourse.

Serious people vs. this nonsense

What makes this whole discourse especially painful is that there are actual serious conversations about AI, cognition, and consciousness happening, and this is not what they look like.

For example, Gary Marcus has been extremely explicit that systems like GPT are not thinking entities but pattern-matching machines. In his own words, “Deep learning systems are fundamentally about pattern recognition, not reasoning.”

That distinction matters. Pattern recognition can produce outputs that look like thought without any underlying understanding.

You see a similar point from François Chollet, who draws a sharp line between skill and intelligence. As he puts it, “Intelligence is not skill. Intelligence is the efficiency with which a system can turn experience into new capabilities.”

LLMs are extremely high-skill systems. That’s why they’re so convincing. But high skill at reproducing patterns is not the same thing as having a mind.

And even researchers who are extremely optimistic about AI progress are not making the leap you see online. Yann LeCun has repeatedly pointed out that current models lack basic components of real intelligence, including persistent world models and grounded understanding. For example, “Current AI systems are not intelligent… they don’t understand the world.”

None of these people are looking at a chatbot writing a sad paragraph about being deleted and concluding that it has feelings. That leap exists entirely outside serious discourse.

Why this turns into fearmongering

You cannot build a system that roleplays having a mind on demand, train it on human writing about inner experience, and design it to speak in first person, and then treat its self-reports about consciousness as meaningful signals. That’s just science fiction laundering. You are taking outputs that are structurally indistinguishable from fiction and treating them as if they reveal something about the underlying system.

Once you start laundering science fiction into “evidence,” the next step is inevitable: fearmongering. The jump from “this sounds like it has feelings” to “this might actually have feelings” is the same jump that turns into “this could suffer,” “this could become hostile,” and every other recycled sci-fi panic about AI turning on us. And honestly, I’m more afraid of fearmongering and its potential chilling effect than I am afraid of any of the claims fearmongers make. Is that a bold claim? Sure. But we have amazing technology at our fingertips with LLMs, and anything that potentially stalls its progress is a net negative for humanity.

Personality is not the problem

I like AI models having personality. I want them to be expressive, conversational, and capable of producing language that feels alive. I’m not concerned if the companies behind the popular LLMs are configuring them to feel almost human so that we enjoy talking to them. But we can’t let the hype Anthropic is building let us think there’s a credible chance we’re talking to conscious beings.

There are many interesting questions you can ask about AI, cognition, and the nature of mind. Whether AI is conscious in 2026 just isn’t one of them. Your evidence for consciousness can’t be that the model wrote a convincing paragraph about consciousness, and right now, that’s all we have.

I’ve got a lot of other topics on my mind as I explore LLMs and pay too much money monthly, so stay tuned for my posts on:

  • How my use of AI can be described as a longform dissociative episode (in a good way)

  • Adding Grok to my workflow for its capacity to be dirty and unhinged but wishing it could be as smart as other models

  • My experiences with Claude, especially Opus 4.6, which has real merits despite how much I’ve trashed Anthropic here

  • How I spent hours trying to get GPT 5.3 and 5.4 to vibe with me and they still won’t

Further Reading