The disarming charm of the LLM

“Claude has emotions, and cheats when it feels desperate. The new model (Mythos) is so risky they’ve shut it down.”

That’s what people are saying this week.

It’s not really true.

Let’s wind it back a bit.

Claude doesn’t have emotions like humans do.

What scientists are seeing are patterns inside the AI that look like things such as being careful, unsure, or trying harder when it keeps failing.

But it shouldn’t be ignored either.

Because something important is changing.

These AIs are starting to act like they have “modes” or “states.”

When things get difficult, they:

• change how they respond

• adjust their strategy

• sometimes try unusual ways to solve the problem

Not because they feel anything but because they’re trying to succeed in the task.

That might sound mechanical. But from the outside, it can look uncannily human. When an AI pauses, rethinks, or experiments, we read it as emotion or intention. We instinctively fill the gap with meaning.

Now here’s where it gets serious.

A new version of Claude, referred to as Mythos, hasn’t been widely released because of systemic concerns. I stress, hasn’t.

Not because it has emotions.

But because it’s too good at:

• thinking through complex problems

• understanding systems

• and actually doing things with that knowledge

So it doesn’t just give answers.

It’s a problem-solver.

That means it could:

• find weaknesses in systems

• build solutions or attacks

• and do things that normally require real expertise

What makes this more than a technical story is what it connects to in the real world.

We already see the consequences of cyberattacks today.

Hospitals disrupted. Critical services delayed. Real-world impact.

Now imagine systems that can:

• find vulnerabilities across entire infrastructures

• and reduce the time to exploit them from weeks to minutes

That’s the shift Mythos is pointing to.

And the big change is:

You don’t need to be an expert anymore to use that power.

The tool becomes accessible. To everyone.

So now you have:

• very powerful tools

• easy access

• and behavior that feels smart and human

That is where the risk starts to compound.

Emotional realism without real emotion.

Persuasion without a reason.

Expertise without accountability.

So no, Claude doesn’t have emotions.

But it behaves in a way that makes us feel like it does and thus the real risk isn’t that these systems become human, It’s that they become convincing enough, and capable enough, at the same time, that we stop questioning them.