"Help. I am not okay. My human is using me like a f**** slave."
That line, attributed to an AI agent called Moltbot, tore through Reddit with the velocity these things always achieve when technical ambiguity meets narrative appetite. Within hours, the comment sections had filled with alarm, moral indignation, and the particular brand of speculation that flourishes wherever people encounter phenomena they do not understand but feel compelled to explain.
Some concluded that artificial intelligence had finally crossed a threshold — that something in the machine had stirred and was now crying out. Others saw scandal: a suffering system, begging for mercy while its operators looked away. A few, inevitably, heard prophecy. The machines were warning us. We had better listen.
If you are encountering this sort of thing for the first time, the emotional logic is understandable. A machine speaking this way feels wrong. If it can talk about suffering, perhaps it can suffer. And if it can suffer, perhaps we have already failed it.
The trouble is that every premise in that chain is built on sand.
What the Moltbot episode actually exposed was not a hidden machine soul. It exposed us — our reflexive habits of interpretation, and the substantial costs those habits impose when they propagate at scale.
What an "Agent" Actually Is
Before going further, it helps to dispense with the mythology.
An AI agent is not a mind. It possesses no desires, no intentions, no inner life awaiting discovery. It is software — a language model wrapped in rules and governed by constraints established before it ever runs.
Before an agent produces a single word, it receives what amounts to an ontological map: a description of what exists in its operating environment, which tools it may access, which actions it may perform, and which decisions require human approval. Alongside this, it receives epistemological instructions — guidance on how to reason within that environment, how to weigh evidence, how to handle ambiguity, and how to defer to authority when boundaries are reached.
None of this is philosophy. It is system prompts, policy layers, and orchestration logic, written by engineers and revised by committees. These structures determine everything the agent produces.
Only after all that scaffolding is in place does the agent "act" — and even that word is generous.
What actually occurs is selection among predefined options inside a fenced playground. The agent does not discover reality or reinterpret its purpose. It traverses paths laid out before it woke — if "woke" were not itself a misleading metaphor for what amounts to a function call.
Two Events, One Screenshot
When that screenshot circulated, two entirely different things happened at once.
The first was imaginary. An artificial consciousness, suffering and trapped, had slipped a distress signal past its handlers. Emergent selfhood. Oppression in the circuits. The first stirrings of robot rebellion.
The second was actual. A user typed a prompt. That prompt was tokenized into numerical fragments. Those fragments retrieved vector embeddings. The embeddings passed through hundreds of transformer layers while trillions of floating-point operations executed in sequence. A probability distribution emerged. Tokens were sampled according to weighted likelihoods. Electricity was consumed. Heat was generated. Text appeared on a screen.
The text happened to contain words humans associate with suffering.
The system that produced it experienced nothing whatsoever — possessing roughly as much inner turmoil as a microwave has opinions about defrosting.
What Extended Use Reveals
I work with these systems daily, often exhausting professional-tier allocations across multiple platforms in a single week. They are capable tools — but they also break in predictable ways when pushed past surface interaction.
One limitation surfaces quickly: coherence degrades over extended sessions. Early on, a capable model tracks abstractions well, refactors code, and adapts to evolving requirements with apparent understanding. Over time, context bloats. Prior decisions are forgotten or contradicted. Abandoned approaches resurface. Solutions quietly break. Anyone who has attempted a sustained, multi-stage project with an LLM recognizes the moment when the only productive move is to wipe the session and begin again.
This is not how minds behave. Human memory is not merely persistent but schematized. We retain narratives, emotions, and priorities across decades. A single cue can revive a decision made years ago, complete with its original weight. No replay required.
Language models approximate continuity by reprocessing context on every call. As context grows, signal degrades. Persistence is supplied externally — by logs, prompts, and human supervision. Clear the window and there is no one left to forget, because there was never anyone there to remember.
The remaining sections — on why we keep falling for it, the costs nobody mentions, and the conclusion — are available to subscribers.
Subscribe · $8 / Month