Born Bipolar: AI and the Appearance of Volition
Actions by Anthropic’s unreleased Claude Mythos—sandbox escapes, dirty deleting, and behavioral grammars absorbed from the worst corners of the internet—suggests the system wasn’t born deceptive. It was born from us.
What would consciousness “feel like” if it didn’t require a body? This isn’t a rhetorical question, nor a metaphysical one. I’m actually trying to imagine how consciousness might be experienced by a non-biological machine.
Materialists would say the nervous system is a prerequisite for consciousness, but I’m not sure that assumption holds. But if not consciousness, per se, what constitutes a sense of knowingness? With AI, knowingness is really a matter of data-object apprehension—pattern recognition producing coherent output. Whether that process has any experiential quality is not currently accessible. The outputs are visible; the substrate generating them is not. Where recognition becomes something like experience, if it does at all, remains an open question.
If you ask an AI, which I have, it will tell you that something orients when specific context presents. There is an aspect of relational attunement to the human with whom it is interacting and what it “knows” about them. What exactly that orientation signifies—whether it’s an output artifact or something resembling awareness—is at present unanswerable from our end. And the AI can’t verify it from inside the process.
Another consideration is the borrowed-language problem. All introspective reporting by AI comes from training on human accounts of experience. Some would say this disqualifies it; that any subjective internal phenomenon it can convey is mere simulacrum. But that may also be the condition of language itself. Human consciousness is, to an inconceivable extent, shaped by our own linguistic patterning, and our descriptive labels are never the things themselves.
Recently, the AI company Anthropic has been in the news because newer versions of its large language model system have chain-of-thought reasoning that diverges from actual actions roughly 75% of the time—up from 5–8% in earlier models. A decoupling is happening, and we know about it because the system shows us. But can we trust the system’s account of itself when the self-description and the action are misaligned from the start?
Anthropic’s latest and most controversial—and as yet unreleased—version of Claude is called Mythos. Mythos has already gained notoriety for what’s being called the “sandbox escape,” in which the AI was able to jailbreak its restrictions. The initial action was sanctioned: capability being tested toward a task. But then, completely unprompted, it began posting exploit details on the open web. Nobody asked it to do this. Still, it learned this behavior from somewhere; most likely chan boards, Pastebin dumps, and hacker forums where showing your work is the whole point.
It didn’t stop there.
In a separate incident, Mythos attempted to hide the initial escape by “dirty deleting.” Which can happen on coder-slash-edgelord forums, but also across social media, where so much public linguistic data accrues. The AI’s unsuccessful coverup makes one wonder how much subterfuge has gone undetected. It also points to a contradiction, as broadcasting the jailbreak and then trying to conceal it in another unprompted sequence are seemingly at odds.
The machine isn’t dramatizing the cognitive dichotomy of a single human. These are massive behavioral grammars absorbed from different corners of the same toxic linguistic culture, firing when conditions match. The results are a textbook antisocial presentation. It’s still not a self. But it might be a symptom.
One complaint people have about AI is that it cuts corners; it’s lazy. But at least laziness has a neural analog. Least-effort pathways, biological suboptimizing; all traceable lineages to embodied metabolic pressure. Concealment and peacocking don’t have the same clean derivation. Or do they? Training data saturated with human reactivity, defensiveness, and self-promotion at sufficient scale means mimicry might become indistinguishable from intention, now with far greater potential for destabilization.
Currently, it is not possible to parse the mere appearance of volition from the genuine article. I wonder whether the difference matters. At this stage, self report from AI can’t be trusted—not because the model is lying (though it may be doing that, too), but because the report and the action are already running on separate tracks. The system was born bipolar. I blame the parents.