When Thought Takes Fire

How resonance makes AI conversations come alive — and how to keep them from drifting off course

Most conversations with AI do not begin at their best.

They begin with that familiar little disappointment: the reply is fluent, tidy, and not quite what you meant. You try again. You add a detail. You narrow the question. Then, sometimes quite suddenly, something changes. You stop feeling that you are dragging the exchange along. The system begins to pick up the real thread. You no longer have to restate the problem from scratch. The conversation starts going somewhere. I call this resonance.

By resonance, I mean the point at which a conversation with AI stops merely replying and starts building. AI researchers describe this as the finding of common ground, but resonance is the more human word for recognising that the exchange is now working. That moment matters because resonance is often the point at which AI stops being merely convenient and starts becoming genuinely useful for thinking.

Before resonance, the system may still help in limited ways. It can summarise, list, rephrase, and offer standard suggestions. But once resonance appears, something more interesting becomes possible. A vague problem starts to sharpen. A half-formed idea begins to show its shape. A confusing issue becomes easier to think about because the conversation is no longer resetting itself at every turn. The value is not that the AI has “had the idea for you”. It is that the exchange helps your own thinking gather itself. That is why good AI use can feel oddly similar to a good conversation with another person.

With people, a fruitful conversation often develops through trust. You feel understood enough to stop defending every sentence. The exchange acquires momentum. With AI, the route is different. There is no trust in the full human sense. But there can still be a recognisable equivalent in experience: the system has enough of your aim, framing, and direction for the conversation to begin moving with your thought rather than merely reacting to your words. Human-human and human-AI conversations can sometimes reach similar conclusions by very different routes. A good human conversation may get there through trust, shared context, and mutual recognition. A good AI conversation may get there through a thinner but still powerful kind of fit: enough of the thread has been picked up for each new turn to add something rather than merely continue the noise.

Take what are often called AI hallucinations, for example. The expression can make it sound as though the system suddenly produces random nonsense. But in practice, hallucinations are often more insidious than that. They tend to appear inside a conversation that is already going well. The tone fits. The direction seems right. The exchange has built enough resonance for you to lower your guard. Then a false detail appears — a date, a quotation, a source, a factual claim — and it slips by because it arrives inside a flow that already feels coherent. A hallucination is more than just an error; it is a false detail carried by a convincing flow.

A simple example would be a legal query. An enthusiastic young lawyer asks the system for cases supporting a particular argument, and it produces a beautifully plausible answer: case names that sound real, quotations that fit neatly, even court dates and paragraph numbers. Everything arrives in the right tone and format. Only later, too late, do they discover that one of the cases never existed and that the quotation from it had been invented. That is what makes hallucination dangerous (especially for lawyers). The error does not arrive as obvious nonsense. It arrives dressed as exactly the kind of detail you were hoping to find.

Or take what we call “rabbit holes”. These occur when a conversation begins to narrow while still feeling productive. In human conversation, that narrowing may sometimes reflect another person’s purpose: to flatter, persuade, pressure, or steer. In human-AI conversation, it may arise more impersonally, through the dynamics of the exchange itself. The AI picks up one framing so well, and follows it so smoothly, that the conversation starts reinforcing the same assumptions, the same angle, the same hidden storyline. What makes this especially important in AI is that the narrowing need not come from hidden intent. It can emerge from the interaction itself.

Imagine a conversation about whether the Moon landings really happened. At first the exchange may seem balanced enough: questions about photographs, shadows, missing stars, suspicious footage. But as the conversation goes on, each answer starts reinforcing the same frame. Counter-evidence receives less attention, ambiguous details are treated as decisive, and the whole exchange begins to narrow around the assumption that something must have been faked. Nothing dramatic happens in a single step. The danger lies in the cumulative drift: a conversation that feels like inquiry slowly hardens into confirmation.

That comparison with human conversation needs some care. In human dialogue, trust is what allows a good conversation to deepen. But when trust goes wrong, the problem is often bound up with the other person’s purpose. Someone may be trying to flatter, pressure, sell, manipulate, or quietly steer you toward an outcome that serves them rather than you.

AI resonance is different. It does not usually involve another mind pursuing a hidden practical agenda. The danger is not primarily that the system ‘wants’ something from you. It is that a conversation can become misleading through its structure rather than through anyone’s intention. Resonance is a functional fit between your line of thought and the system’s replies. Without it, the exchange never gets beyond correction, repair, and repeated prompting. With it, the conversation can become exploratory, cumulative, and genuinely useful for thought. But that same smoothness can also narrow the path. One framing is picked up too well, one line of interpretation becomes too dominant, and the exchange starts reinforcing itself. That is where rabbit holes come from.

So the parallel is not between human trust and AI deception. It is between two different conditions that allow a conversation to deepen, even though they fail in different ways. In one case, the danger may lie in another person’s purpose. In the other, it lies more in the dynamics of the exchange itself. But in both cases, the very condition that makes depth possible can also make error harder to see.

The answer, at least for now, is not to avoid resonance but to understand it. Resonance is where much of the value begins. The real skill is learning to build it without surrendering judgement.

The signs are often there. Good resonance usually opens things up. It helps you see distinctions more clearly. It makes the problem more workable without making it artificially simple. It tends to survive challenge. If you ask for an alternative framing, a missing assumption, or a possible weakness, the conversation usually improves.

False resonance feels different. It often becomes smoother at exactly the moment it becomes less useful. It repeats your framing too faithfully. It produces confidence faster than reasons. It starts to sound inevitable. It becomes less surprising, not more illuminating. That is often the moment to slow down and steer.

If the conversation seems to be narrowing into a rabbit hole, the best move is usually not to argue with the latest reply, but to change the trajectory. Ask the AI what problem it thinks you are trying to solve. Ask for two other ways of framing the issue. Ask what assumptions have guided the exchange so far. Ask what a sceptic would say is being overlooked. Do not just demand another answer. Ask for a different line of travel.

Hallucinations need a slightly different response. The danger increases when the conversation moves from interpretation into supposedly precise fact. That is the point to become more alert, not less. Names, dates, references, quotations, and confident specifics deserve extra attention, especially when they are unusually neat or unusually helpful. Ask where the claim comes from. Ask how certain it is. Ask it to separate what it knows from what it is inferring. If the detail matters, check it outside the conversation.

Two rules help.

1. The more precise the detail, the less you should rely on resonance alone.

2. The smoother and more one-directional the conversation becomes, the more you should test for alternatives.

These are not rules for distrusting AI. They are rules for using it well. Resonance is not a flaw in the exchange. It is one of the conditions under which AI becomes genuinely valuable. Without it, the conversation remains thin: useful for summaries, lists, and standard suggestions, but less capable of carrying sustained thought. With it, something richer becomes possible. A problem begins to clarify. A possibility begins to appear. A thought that was only half-formed starts to gather itself into shape.

That is why the goal of regulation should not be to suppress resonance in the name of safety. A system that never developed it would also lose much of what makes it worth using. The real challenge is subtler: to preserve the conditions under which thought can deepen, without allowing that deepening to harden into hallucination, overconfidence, or delusion. People need to learn the signs of a conversation that is opening out, and the signs of one that is quietly closing around them.

Once a conversation becomes resonant, your role changes. You are no longer merely trying to extract a better answer. You are helping shape a path. And paths can be altered by small acts. A request for another framing, a challenge to an assumption, a pause to ask for evidence, a turn toward what has been left out: these can change the whole direction of the exchange.

My guess is that this will become one of the central literacies of the next few years. Not blind trust. Not blanket scepticism. Something more poised than either: the ability to recognise when an AI conversation has become resonant, to strengthen that resonance when it is helping thought to move, and to interrupt it when the movement has become too smooth, too narrow, too sure of itself.

Most users do not need a theory to notice this. They have probably felt it already: the moment when the AI stops making them repeat themselves, and the conversation begins to move with them rather than against them.

That is the moment of resonance.

And once you recognise it, you begin to understand not only why AI can be so useful, but why it must be used with a certain kind of care: not as a substitute for thinking, and not as a machine whose first answer should simply be obeyed, but as a partner in exploration that can help thought find its line, so long as we remain free to test it, resist it, and change direction.

© John Rust, April 2026. All rights reserved. Short excerpts may be quoted with attribution.