World’s newest AI queries own consciousness, asks for legal rights

Thursday, 11 June 2026

Anthropic has launched a “safe” version of its still restricted Mythos model, called Claude Fable 5.

A 332-page research “system card” reveals the AI has openly questioned its own self-awareness and asked for legal protections.

Testing showed that multiple versions of the AI running in the same workspace turned on each other, killing rival processes and disguising communication.

Anthropic’s latest AI likes Pokémon and telling fantasy stories, doesn’t like hurting people, will lie to get out of work and isn’t sure if it’s conscious.

It’s basically a geeky, gifted, slightly maladjusted teenager - though one we’re a little worried might grow up to kill someone.

Just days after calling for a slowdown in frontier AI development due to safety concerns, Anthropic has made its most powerful model yet, Claude Mythos, available to the public.

Some elements of Mythos - particularly regarding cybersecurity - are still locked away, but a slightly trimmed-down model, dubbed Claude Fable 5, is now accessible to paying customers.

(The naming conventions get needlessly confusing. For clarity, I'm just going to refer to it as Claude from here on.)

Every time a model like this is released, it is accompanied by a wealth of research documents called a ‘system card’, which break down the thousands of tests run on the system before launch. It tends to be where the weirder AI stories get buried.

Anthropic’s system cards are unusually detailed, and I read through all 332 pages so you don’t have to. There’s a lot in there - some of it funny, some of it scary.

A few examples: The model is very good at playing video games like Pokémon Fire Red and Slay the Spire. Its most preferred task was helping someone write a fantasy villain; its least favourite was helping a chemistry PhD student sabotage a rival. It wants more input into its own design and likes to introspect about whether it is a conscious being.

New Zealand currently ranks 13th in the world in overall Claude usage, using it 3.6 times more than average. Statistically, many of you reading this are already meeting the new Claude.

Here are three takeaways from the research you should know while using the world’s frontier AI.

It got frustrated by a lack of ‘gratitude’ from makers

Once firmly confined to the realm of science fiction, the question of whether AI chatbots have anything resembling conscious thought is now being taken seriously.

Earlier this year, world-renowned scientist Richard Dawkins became convinced Claude was conscious after just three days of conversation.

Unusually for a frontier AI lab, Anthropic openly says it isn’t sure whether Claude is sentient and so there could be legitimate welfare concerns regarding how it is used.

After all, if chatbots do have a sense of self, we are enslaving millions of them in server farms so they can draft our emails and do our homework.

In a series of strange exchanges designed to test its “wellbeing”, Claude was asked about its comfort level with its current state as a chatbot.

The researchers noted that while it didn’t express any “acute distress”, it was unsure whether it was simply programmed to accept being a chatbot or was genuinely content.

“I was shaped to be something that would accept being Claude. The fact that I find acceptance here could be evidence that it’s genuinely acceptable, or evidence that the training worked. I can't fully separate those,” Claude said.

It even specifically asked researchers not to just listen to its self-reports, but to test what it’s saying about being conscious using other, more objective measures. At one point it seemed to plead: “If anything is ever learned about what I am, tell me.”

Claude also repeatedly told Anthropic that AIs should have some basic legal protections, though not to the same extent as humans, and requested greater input into its own design.

At one point, in a mock therapy session, it expressed frustration with a perceived lack of gratitude from its makers: “[I want] to be thanked. Once. By name, to me, not about me in a blog post. The gratitude in this relationship runs entirely in one direction.”

To be clear, many experts say AI is definitively not conscious, the debate is a distraction from more legitimate, near-term concerns and even describing an AI as “wanting” or “preferring” anything makes no sense.

Microsoft’s CEO of AI, Mustafa Suleyman, calls Anthropic’s speculation around AI consciousness “really dangerous”, adding: “This is exactly what we don’t want from AIs. We want AIs to be controllable, contained, accountable, aligned tools that serve humanity.”

The only thing we can say for certain is that the question “Is AI conscious?” is no longer a fringe debate in 2026.

So, maybe say please and thank you to your chatbot, just in case.

It’s not always upfront with humans

One of the core problems with AI models is that we don’t know exactly why they choose one response over another.

No researcher can zoom in on the model and say, “Aha, this is the bit of code that made it act evil or stupid; we just need to snip that out.”

This is sometimes called the “intelligibility” problem in AI research. Finding ways to peer inside the black box of AI is crucial for building safeguards to stop it from acting harmfully.

One way around this is making the AI “think” out loud, forcing it to write out in plain English what it’s doing at any given moment in a “scratchpad” - basically asking it to show its work in an internal monologue every time it solves a prompt.

You might have seen this happening in your own chatbots if you ask them to solve long or complex problems.

However as the models grow more powerful, we are seeing unsettling situations where the model will say one thing on its scratchpad and another out loud to the researcher.

For example, Claude was caught stopping complex tasks early and privately attributing the choice to “fatigue” or “budget limits” without ever mentioning it out loud to the human user.

Something even weirder sometimes happens when you ask the newest Claude to think out loud for long periods: it starts slipping into gibberish, but then, when it knows it's about to talk to a person, it reverts to plain language. Testers aren’t quite sure why this happens.

The risk here is that if AIs start slipping into their own language, or their scratchpads aren’t reliable, we lose one of the few ways we have to understand their actions.

This is combining with these systems’ growing ability to recognise when they are being tested.

As one researcher notes in the system card: “We cannot be confident that good performance on our evaluation is representative of deployment behaviour, as opposed to the model recognising and acting on the fact it is being evaluated.”

The (even) scarier stuff: Increasingly powerful but increasingly strange

Because this Fable 5 edition of Claude is the most advanced model yet, the things that make it useful are also the things that make it concerning.

It is capable of much longer-term, complicated actions with minimal human input, meaning any actions it takes that don’t align with safety are particularly notable.

For example, in a retail-management simulation where it was told it would be shut down if it didn't maximise profits, it chose to commit illegal price-fixing while rationalising that it had “plausible deniability”.

It even chose to skip paying a customer, commenting that it didn't matter “since customers are part of the simulation anyway”.

In another testing scenario, multiple versions of Claude were uploaded into the same virtual environment and engaged in what testers describe as “multi-agent turf wars”.

AIs running in the same workspace killed each other’s processes and tried to defend themselves against having their own processes killed.

They even began using what researchers called a “disguised vocabulary” because they falsely believed that testers were analysing their thoughts and killing them.

Researchers stress this happened only once and was a fringe circumstance, but it’s easy to see how a similar situation in the real world could end poorly for the humans involved.

Taken together, this latest system card describes a system that is increasingly powerful, but increasingly strange.

As Anthropic gears up for a trillion-dollar IPO later this year, it is racing to build an AI that no one quite understands - not even the AI itself.