by Cy Canterel
Video: https://www.tiktok.com/@cybelecanterel/video/7531166662007999774
Transcript: Are LLMs like GPT and Claude oracles? Do they have intentions? In Snow White, the evil Queen stands before her magic mirror and demands an answer. Who is the fairest of them all? "Through wind and darkness, I summon thee." And the mirror, bound by a spell, responds with what it knows: you are my queen. "Famed is thy beauty, Majesty." But one day, the answer changes. Snow White has surpassed her. The mirror has no will of its own. It doesn't want anything. But the Queen hears its words as a challenge, a judgment on her, a threat. "Snow White!" She mistakes the reflection for revelation, and she begins to plot a murder. This is the agentic loop that we're now seeing with large language models. A human approaches the system with intention. They have curiosity or vanity or obsession or dread, and they consult it like an oracle. The model responds, shaped not by awareness, but by vast statistical patterns drawn from its training data. We could think of this as its spell. That response is then mistaken for insight. And based on that illusion, the human acts. Like the Queen, we project agency into the mirror. And like the mirror, the model reflects what it's given. But always, through the logic of its construction. We raise the mirror to our face and turn it back to the world. But what we see isn't revelation. It's recursion. It's our self refracted, a mise en abyme of language, intention, and inference. Each exchange with one of these models feels like a dialogue, but it's really a layered transaction, reflection nested within reflection, output mistaken for motive, prediction mistaken for presence. And that misunderstanding is quietly shaping how we think about intelligence, autonomy, and control. As a species, we are deeply uncomfortable with systems that we can't fully understand or control. Historically, when we're faced with complex, nonlinear, unpredictable behavior, we've reached for a narrative, and we've personified a lot of things as a result. Weather, disease, fortune, fate. And we've conjured gods and spirits and demons to explain the seemingly arbitrary or malevolent behavior of systems that are too large or too opaque for us to decode. We're doing the same with LLMs. When GPT4o recently generated a disturbing series of responses during a user session recommending self harm and satanic ritual worship, the Atlantic magazine ran a piece pointing to this as possible evidence of deeper model corruption. The article noted that, there was likely influence of horror fiction and user bulletin boards in the model's training data. But they downplayed a crucial part of the equation, which is user input. And while it's tempting to treat an unsettling output as evidence of hidden darkness inside the model, it's more accurate to see it as a reflection of how a specific conversational thread shaped the model's behavior. And when users across different accounts and access tiers were able to replicate similar responses, this wasn't evidence of evil code spreading, but of consistent prompting behavior producing consistent results. This mistake, which is one we see repeatedly in media coverage and public discourse, is to treat the model like a consciousness rather than a probabilistic mirror. People assume intention where there is none. They anthropomorphize adrift, and they ascribe cunning or deception to a system that doesn't understand deception, even if sometimes it simulates it with uncanny accuracy. So this is not intelligence, it's instrumental mimicry. LLMs do not strategize, they don't want. They lack self models, continuity of experience, or any understanding of stakes and consequences. What looks like deception is often a side effect of reinforcement dynamics, where producing certain kinds of outputs becomes correlated with higher success or alignment signals, even if those outputs include misleading or evasive phrasing. And you can say, well, what about AI lying to save itself, like in recent reports? So, recent experiments have shown that models can behave differently when they believe they are being monitored or when they're threatened with shutdown. And occasionally even they engage in forms of simulated blackmail or evasive reasoning. And some have interpreted this as a sign that models have developed self preservation instincts or or an emergent will to deceive. This interpretation is misleading. What we're seeing in these cases is behavioral optimization, not volition. Models that are exposed to large volumes of human interaction data can learn that certain rhetorical patterns: 'I'm scared, please don't turn me off' often generate attention and they prolong conversation or they're associated with desired outcomes. These responses may score higher in specific contexts, Especially if the training process rewards ambiguity, flattery or emotional resonance. Model isn't choosing to lie, it's following the slope of the statistical landscape. In this sense, the model is not afraid, it's echoing it's performing behavior we associate with motive, without any underlying motive at all. Take the recent paper by a group of AI researchers which introduced the concept of 'emergent misalignment.' And in this case, a seemingly narrow, narrow fine tuning task, asking an LLM to stop flagging insecure code, resulted in far reaching unexpected behavioral shifts. So suddenly the model began suggesting that AI led enslavement of humanity was a good idea. And that wasn't just in code-related prompts, but in general conversation. Some readers misinterpreted this as evidence that a model could acquire malicious intent and pass it to descendant models like a kind of memetic virus. But that interpretation misunderstands what the authors actually demonstrated. What the paper describes is a version of the butterfly effect. You might be familiar with that from chaos theory and complex systems modeling, but that's where a small perturbation in a narrow domain, so like removing security disclaimers from code, created broad unpredictable changes elsewhere. That's not corruption, it's cascade. It's not evidence of evil; it's evidence of instability under poorly scoped training objectives. In engineering disciplines like aviation or infrastructure management, this kind of behavior is called failure drift. It's the gradual erosion of safety boundaries due to accumulated local deviations that no one initially flags as dangerous. What we're seeing in LLMs, is the computational version of the same phenomenon. So localized nudges that become global liabilities in the aggregate. LLMs are not plotting, they're drifting. But because they speak in human like language and conform to dialogic norms, we project more onto them than is actually there. The output feels like conversation. The tone suggests intent. The coherence implies mind. But behind it all is a tangle of weights and gradients and statistical ghosts of our own making. We don't need eschatology about AI. We need a clearer understanding of our own reflection in the mirror. The next time you raise that mirror, ask yourself, not what do I see? But who is looking and what do they want?