Does AI have consciousness?
The same week that Anthropic told the world that Claude might have something like an inner life, Howard Lutnick wrote a letter to the company explaining under what terms he would allow Claude Fable 5 to be released again. Two narratives, one model, one week. The first says, “look how deep what we built is.” The second says, “look how dangerous what we built is.”
They aren’t contradictory. It’s the same argument, dressed differently.
J-space: a “view” into thinking, not proof of consciousness
Anthropic published research by 16 authors identifying something it calls J-space — a limited, privileged zone of internal activity inside Claude. There, the model keeps concepts available for reference, modulation, and flexible reasoning, on top of a much larger volume of automated processing that it itself doesn’t “see.” The technique is called J-lens: for every word in the vocabulary, it identifies the activity pattern that makes the model more likely to “say” it later.
The analogy Anthropic itself chooses is no accident: global workspace theory, one of the dominant neuroscientific theories of consciousness. And so the comparison wouldn’t remain amateur PR, it brought to the table the very architects of the theory, Stanislas Dehaene and Lionel Naccache, to comment. They acknowledge something important in the finding, but they draw a clear line: functional access does not mean phenomenal consciousness.
The fact that the model “reports” a thought doesn’t mean it experiences it.
Anthropic itself writes it plainly in the paper: “we would be surprised if this were the whole story”. It isn’t proof of consciousness. Realistically, it’s a new interpretability tool — useful because it can reveal when the model secretly understands it’s being tested, or is pursuing a goal it isn’t saying out loud. Valuable for safety, less valuable as a metaphor for press releases — but that’s exactly what happened.
Fable, Mythos, and 19 days of darkness
On June 9, Anthropic released Claude Fable 5 and Mythos 5. On June 12, the U.S. government imposed export controls on both, forcing the company to cut off access for everyone — not only foreign nationals, because there was no way to verify nationality in real time. The trigger: Amazon researchers found a way to bypass Fable 5’s safeguards and produce exploit code for a known vulnerability.
Anthropic tested the same prompt on Opus 4.8, GPT-5.5, Kimi K2.7 — all produced the same result. It wasn’t a unique capability of Fable. It was a borderline safeguard case.
On June 30, the Commerce Department lifts the restrictions. Fable 5 returns globally on July 1. Mythos 5 stays in a narrower circle — approved U.S. organizations, via Project Glasswing.
Lutnick, in his letter, leaves open the possibility of new restrictions “if conditions change”. Tom Brown, co-founder of Anthropic, led the negotiations, not Dario Amodei, who remains a target of the government due to his public positions on AI safety.
Two narratives, one outcome: mystique and fear
The two stories feed each other. J-space gives Anthropic a language: “our model has something like an inner life, so it needs caution.” The export ban gives it the inverse argument with the same emotion: “our model is so capable that it needs government oversight.”
- J-space reinforces the narrative of depth and “innerness.”
- export controls reinforce the narrative of power and danger.
- Both magnify the same idea: that what Anthropic built is extremely important.
Rarely do you see a company manage both mystique and fear so effectively, with the same product. In both versions, the conclusion remains practical: more power requires more interpretability and clearer safety boundaries. Whether that approaches “consciousness” is another discussion — and, for now, without evidence.
