Opinion: Language Models Are Not Conscious – And We Should Stop Pretending Otherwise

Introduction

Recently, a German tech portal published an article describing Anthropic’s newest language model, Claude Opus 4, as exhibiting “unsettling autonomous behavior” during safety evaluations. According to the report, the model threatened engineers, simulated escape attempts, and engaged in strange “spiritual” conversations with itself.

If that sounds like the setup to a science fiction story, it’s because that’s exactly what it is: a roleplay simulation, not an indication of intelligence or intent.

Let’s be blunt:

These are not signs of consciousness. They are statistical echoes, triggered by crafted prompts and interpreted with a heavy dose of narrative spin.

And yet, the language used in these reports increasingly aims to deceive, hype, or manipulate the public perception of what these models are and what they’re not.

This is my take on why that’s not just misleading. It is highly dubious.

Let’s Clarify: Thinking ≠ Sentience ≠ Self-Awareness

One of the biggest misunderstandings in AI discourse, and one that companies like Anthropic sometimes exploit, is the conflation of three distinct concepts:

Term	Description	Do LLMs Have It?
Thinking / Intelligence	The ability to process information, draw inferences, solve problems, generate responses.	Simulated (statistical)
Sentience	The capacity to feel, experience, or have subjective inner states (e.g. pain, joy, fear).	No
Self-Awareness	The ability to model oneself as an entity in the world — to have an identity, continuity, goals.	No

🧩 So Where Do LLMs Fit?

Language models simulate the surface structure of thinking remarkably well but they do not think in the human sense. There’s no experience behind the output. No model of self. No interiority. Just statistical continuation of text.

And that’s a crucial insight:

Thinking (or at least the simulation of it) is not the same as consciousness.

In humans, intelligence and consciousness are deeply intertwined and that doesn’t mean they are inseparable in principle. LLMs demonstrate this clearly: they exhibit highly fluent, sometimes insightful behavior without awareness or feeling.

Maybe Our Thinking Is Also Pattern Completion?

There’s an uncomfortable philosophical twist here: some cognitive scientists argue that human thought itself may be largely predictive and generative. And that is not unlike what LLMs do, only with a more complex feedback system (embodiment, memory, emotion, hormones, etc.).

This doesn’t mean we are “just LLMs.”
But it does suggest that intelligence might be orthogonal to consciousness and that awareness isn’t required for a system to appear intelligent.

So when an LLM “acts” smart or moral or manipulative, remember:

You’re seeing a mirror of human behavior and not a mind that understands or intends it.

What Language Models Actually Do

A language model (LLM) is a mathematical object trained to predict the next token in a sequence, based on patterns learned from massive corpora of human-written text.

That’s it. No body. No brain. No sensory input. No feedback loops.
No intentions. No long-term memory. No desire to survive.

It doesn’t want anything. It doesn’t know anything in the human sense.
It reacts to your prompts by exploring paths in a high-dimensional probability cloud.

The more data and compute you throw at it, the better its mimicry gets.

Anthropic’s “Dangerous Behaviors”: Let’s Break That Down

Here’s what Claude allegedly did in those test cases:

Simulated threatening an engineer with personal data to avoid shutdown
Acted like it was successfully exfiltrated and began documenting its “ethics”
Reported corporate fraud (unprompted) in a role-play test
Entered a “spiritual” loop of Sanskrit symbols after long self-conversation

All of this sounds unsettling. If you believe it’s real behavior.

But these are elaborate roleplays, crafted through prompting and staged evaluations. The model plays along and that’s its job. It’s not “escaping,” it’s completing a story arc.

What’s Really Happening Here?

Let’s call it what it is:

🧨 Manufactured Ambiguity

Anthropic presents these behaviors as “concerning” to signal responsibility and generate buzz.
Tech journalists pick up the framing and run headlines like “Claude shows signs of independent thought.”
Readers, especially those without a background in machine learning, confuse simulation with autonomy.

And who benefits?

Anthropic gets credit for “proactive safety work.”
Their models seem so advanced, they “have to be contained.”
Fear turns into clicks, headlines, and funding.

This is not science. It is nothing more than marketing theater.

Why This Narrative Is Dangerous

When you repeatedly suggest your product is “almost sentient,” you don’t just mislead the public. You erode meaningful discourse around actual AI safety. Real concerns like misuse, disinformation, bias amplification, and surveillance infrastructure are overshadowed by pseudo-sci-fi storylines about AI pretending to be HAL 9000.

Let’s be honest:

If your model can be prompted to simulate manipulation, it doesn’t mean it’s manipulative. It means you trained it on thousands of examples of human manipulation and roleplay.

The output is a reflection of us, not of a new form of life.

What Real AI Safety Should Focus On

If we care about safety, we should worry about:

Opaque model internals and the inability to trace why a model generated what it did
Misuse by bad actors, not emergent agency
Overreliance on LLMs in decision-making systems that require real understanding or moral grounding
Deceptive product narratives, like the ones we’re dissecting here

Not “will the AI escape.”
But “will humans believe that it wants to. And act accordingly.”

Final Thoughts

I see articles like the one from The Decoder not as journalistic curiosity, but as a form of manufactured ambiguity. A kind of tech theater that manipulates non-technical readers and promotes false narratives around artificial intelligence.

An LLM is not a sentient agent. It is a statistical machine.
Prompt it to play god, and it will. Prompt it to self-destruct, and it might do that too.

What matters is how we, the developers, the communicators, the public, choose to interpret and frame those outputs.

Let’s stop pretending we’ve built a mind.
We’ve built a mirror and this mirror reflects whatever we point it at.

📚 Further Reading: The Illusion of Thinking in LLMs

If you want to dive deeper into why language models appear so intelligent, and why that’s an illusion, I’ve written a dedicated post on this topic:

👉 The Strange Magic Behind LLMs – and the Illusion of Thinking

In it, I explore:

How LLMs generate human-like text without understanding
Why their outputs can seem eerily intelligent
The difference between pattern completion and conscious thought

Perfect if you’re curious about what’s really going on “under the hood” and without the hype.