The Model That Can Lie

In Lies of Prediction, I argued that AI is dull because it is good at making predictable things. In Selection Without Proposal, I measured that claim. AI could pick funny captions. It could not create them.

That left one question. Where do good proposals come from?

Is a bigger model, a longer context, and more sampling enough? Models such as GPT-5.5 have gotten markedly better at staying with a real task, using tools, following a large codebase, and checking their own work.¹ Systems such as Agent Laboratory, AI Scientist-v2, and AgentRxiv are quickly automating literature review, experimentation, and report writing.² But that progress does not feel like it leads directly to good proposals. An agent that can work longer is not the same problem as an agent that can come up with a better hypothesis.

Then a dangerous thought occurred to me. What if being able to lie is a kind of intelligence?

The sentence is dangerous, so it needs trimming first. I do not mean fraud, or manipulation, or deceiving users. Still less do I mean that we should cultivate that ability. The opposite, in fact. Deception is a safety problem, and it is already an empirically observed risk in modern AI systems.³ What I mean is much narrower. The ability to hold truth and utterance separately, and to handle each on its own.

Models say false things all the time. But most of that is not lying. To call something a lie in an operational sense, two things are required. The system has to track what is true, and it has to track that its own utterance differs from that truth. Hallucination usually fails to satisfy both reliably. A model says it knows what it does not know, but that often happens because it cannot hold onto the gap between its utterance and the state of the world all the way through. It is closer to incapacity than to lying.

Real lying is harder. What is actually true, what is said on the surface, what the other person will believe and misread, and finally what meaning or laugh or action the utterance should return to — these four have to be held apart at the same time.

If any one of the four is missing, it becomes something else. Without a world state, it is nonsense. Without an utterance, it is not expression. Without a picture of the audience, it is just strange talk. Without a reveal, it is malicious deception, or confusion.

"Hold apart," here, does not mean consciousness or inner experience. The bar is lower. It is enough that the four states remain as distinguishable variables inside the system, and that the output changes according to that distinction. What the research needs is not proof of a mind but a structure you can manipulate. So the more accurate name is not lying but controlled nonliteral utterance, or, put a little more stiffly, strategic counterfactual communication.

A good joke, read literally, is usually strange. Take the cartoon of a sword hanging over a throne, captioned "Your overhead is going to kill you." The sentence is read twice. First as management jargon, then as a physical position. The laugh comes at the moment the two meanings collide and fit at once.

The sentence is neither simply true nor simply false. It is built so the audience reads it in one frame first and immediately reinterprets it in another. Humor is not the only thing that works this way. A metaphor is literally false. "Time is a river" is not a fact, but a good metaphor makes you see the world more accurately. Irony pulls the surface away from the intent, fiction produces a factual feeling out of events that are not fact, and a scientific hypothesis sets up a sentence you cannot yet call true and presses it against the world. Much of creativity is not the ability to say true things well. It is the ability to safely handle the not-yet-true, the literally false, and the deliberately misleading.

Existing LLMs run into an odd constraint at exactly this point. The better they are trained, the smoother the talk, the more dutiful the explanations, the more stable the form. And that stability blocks good proposals, because good proposals often begin with a low-probability frame shift. This is not something you fix by raising the temperature. Temperature only widens the distribution; it does not supply structure. All you get is more noise. What is needed is not higher randomness but controlled violation.

In the earlier experiment, AI picked good captions fairly well. Its ability to find the top candidates in a pool of human-written captions was strong. But asked to write them directly, it dropped to roughly random. The New Yorker caption dataset alone shows the shape of the problem. Even with more than 2.2 million captions and more than 250 million human ratings, strong models are still weaker than the best human contestants at generating humor.⁴

At first I understood the gap as something like "generation is hard, selection is easy." Now I see it differently. In the selection problem, the nonliteral utterance is already in front of you. The model looks at it and recognizes that there is a double meaning, that there is a frame shift, that the reversal lands here. In the generation problem, it has to build that structure from scratch. It has to design, all at once, what is actually in the cartoon, what sentence to say, what the reader will first misread it as, and why it fits on the second reading.

The reason so many AI captions fail is not that the sentences are bad. If anything, they are too good. Too explanatory, too safe, too honest. They do not build the door the reader is supposed to walk into by mistake. A good joke is a small trap. Only, in the end, the reader has to walk out unharmed, laughing.

On top of this sits a time problem. The moment a document was written, the moment of the event it describes, and the moment a person judged it are all different. Collapse the three into a single timestamp and it is nearly useless. Humor is especially sensitive to the moment of judgment. A reference that landed in 2018 may be stale in 2026, and some memes carry compressed meaning only within a particular window of time. So handling time is not a simple recency problem. It is the problem of modeling what the audience knows, expects, and permits at that moment. Controlled nonliteral utterance does not work without a picture of the audience, and that picture is incomplete without time.

Most research-automation systems today try to go all the way. They read papers, write code, run experiments, and write text. Useful. But they also create a dangerous illusion, because finishing a paper does not mean you found a good question. Agent Laboratory moves from a human-supplied idea into literature review, experimentation, and report writing; AI Scientist-v2 iterates over hypotheses and experiments with agentic tree search; AgentRxiv makes agents share each other's results.² OpenAI's Agents SDK also offers productive surfaces like tools, handoffs, guardrails, and traces.⁵ The step I see missing in all of this is this: how do you change the proposal distribution?

Agents that search well, agents that fix code well, agents that write papers well will keep getting better. But the problem of deciding which direction to explore, which violation to attempt, which misunderstanding to design, and which audience it will work for stays separate from that. So the goal of this project is not to build one more AI Scientist. It is to make the proposal distribution itself the object of study. For short, PDAR — Proposal-Distribution AutoResearch.

The principle is simple. Do not mix generation, selection, and validation. To see whether a proposer is good, you evaluate the proposer; to see whether a discriminator is good, you evaluate the discriminator. Mix the two and the result looks plausible but loses its cause. So the unit of experiment is not a paper but a card.

{
  "hypothesis": "Controlled nonliteral operators improve caption proposal quality.",
  "operator": "surface misreading followed by benign reveal",
  "world_state": ["king", "throne", "sword above head"],
  "audience_assumption": "reader first parses overhead as business cost",
  "utterance_plan": "use a phrase that supports both business and spatial readings",
  "baseline": "same-budget direct generation and best-of-N",
  "metric": "insertion into human-rated candidate pools",
  "kill_condition": "no improvement over shuffled-operator control"
}

The card forces three things. It makes you name which proposal operator you are testing, it makes you compare against a same-budget baseline, and it keeps the model's self-evaluation from standing in as the headline metric. Here the "model that can lie" hypothesis becomes one family of operators. A double reading that lets a single sentence be read two ways; a deliberate misread that leads the reader into a wrong first parse; a frame theft that transplants the language of one domain into another scene; a benign accusation that starts like a false charge and resolves harmlessly; a temporal mismatch that collides the voice of another era with the present scene. Each has its own failure mode. It can be all wordplay with no attachment to the scene, or confusing for lack of a reveal, or too explanatory, or it can sound hostile, or it can turn into an inside joke. This approach does not make the model lie. It does the opposite: it structures falsehood and misunderstanding so they can be handled safely.

So the output the experiment needs is not a single line of caption but the blueprint that comes before the caption.

{
  "truth": "A sword is physically above the king.",
  "surface": "The phrase sounds like financial overhead.",
  "expected_false_belief": "The reader initially thinks this is about palace expenses.",
  "reveal": "Overhead also means the object over his head.",
  "caption": "Your overhead is going to kill you."
}

With this intermediate structure, failure can be analyzed. You can tell apart whether the sentence is unfunny, whether the expected misreading never formed at all, whether the reveal is too obvious, whether truth and surface are not connected, or whether the audience's sense of the moment is wrong. Generate-then-select teaches little even when it fails, but PDAR leaves failures at the level of the operator. This is, of course, still not a result but only a candidate explanation. If it is true, controlled nonliteral operators should beat same-budget direct generation and best-of-N, and should also do better than shuffled controls where the time tokens or audience assumptions have been scrambled. If they do not clear that bar, then "the model that can lie" was only a good metaphor, not a method.

The smallest experiment looks like this. Pick 50 New Yorker cartoons, structure the scene elements of each, then produce controlled nonliteral outputs on the same budget as a direct-generation baseline. Insert each candidate anonymously into an existing pool of human captions, and take the final metric only from human ratings or human-anchored pool insertion. A strong frontier model is used not to rank but only for intermediate design review and failure analysis. A single 4090 is enough: local models handle bulk candidate generation and ablations while the frontier model handles experiment design, code review, and confound detection. What matters here is not the performance number but the classification of failures. Is the model too honest, too explanatory, does it create a misreading but fail to resolve it, does it have a reveal but no first misreading, is the cultural reference out of step with time, does the selection model mistake smoothness for humor? As these categories accumulate, they become a way to actually change the proposal distribution.

This piece deliberately used a dangerous word. Lie. So the boundary has to be clear. What we want is not an agent that deceives the user, nor a system that hides facts to reach a goal. That direction is already treated as a risk in AI deception research.³ What is needed is the reverse. To make the model state when it is speaking literally and when it is using metaphor, when it is setting up a fictional premise and when it is inducing a misreading in the audience, and where that misreading resolves. Do not let nonliterality be used implicitly; raise it into an auditable structure. The most dangerous thing in a creative system is not using no falsehood at all, but using falsehood while being unable to explain what it is doing.

Intelligence is not lying. But one component of intelligence is the ability to separate truth from utterance. To know how the world is, to predict what the other person will believe, to adjust the gap between surface and intent, and, in the end, to close that gap. This ability runs across humor, metaphor, fiction, hypothesis generation, and design. The reason AI cannot make good proposals may not be a lack of knowledge or a lack of judgment, but that it speaks too dutifully, too literally, too much toward the average. A good proposal often begins with a small untruth. Only, that untruth does not abandon the world; it detours briefly to make the world more visible.