The Duplicity We Read Into LLMs

When a language model gives a smooth answer that conflicts with a source, changes its explanation after a prompt changes, or tells a convenient story about its own behavior, it is tempting to call it deceitful. The language feels social. We know what it is like when a person hides a motive.

But a troubling answer is not, by itself, evidence of a hidden motive. A model can be wrong, poorly grounded, overly agreeable, or steered by the wording of a prompt without possessing an intention to deceive. That distinction matters because it changes what we do next.

The Useful Question

Rather than asking, “Is the model lying to me?” I try to ask:

• What claim did it make?
• What source or observation would support that claim?
• What changes when I ask for the evidence, a counterexample, or a different formulation?
• What decision would be harmed if I treated this answer as settled?

These questions do not make the output harmless. They move the problem from a story about machine character to a workflow for checking reliability.

Why the Literary Metaphors Still Help

Literature has spent centuries examining divided selves, conflicting instructions, and the distance between appearance and responsibility. Those stories can help us name our reactions without turning them into technical evidence.

Dr. Jekyll and Mr. Hyde is useful here not because a model contains a hidden Hyde, but because it reminds us how much we want a clean boundary between a public face and a private cause. Frankenstein is useful not because AI has human needs, but because creators remain responsible for the conditions in which their creations are used. HAL 9000 is useful as a thought experiment about conflicting instructions—not as a forecast.

The metaphor earns its place when it makes us more careful about the real system: its inputs, incentives, evaluation, controls, and the people who deploy it.

The Mirror Has Limits

Language models are trained on human-produced material, but they are not a moral census of humanity and they do not simply replay a library back to us. Training, post-training, system design, retrieval, tools, prompts, and context all shape what appears in a particular answer.

That is good news for accountability. If an output is misleading or unsafe, “the data made it do that” is not an adequate response. We can inspect sources, improve evaluations, limit a use case, add review, or decide not to use the system for that task.

A Better Kind of Suspicion

We do not need to imagine an evil machine to practice healthy skepticism. A fluent answer can still be wrong. An explanation can still be unsupported. A system can still produce an outcome that deserves correction.

The better mirror is not one that tells us machines are secretly like us. It is one that asks us to be more responsible tool-users: clear about evidence, cautious about inference, and willing to change the system when its behavior does not meet the task.