Artificial intelligence is not conscious – Ted Chiang

Artificial Intelligence Is Not Conscious: What Ted Chiang's Argument Actually Says

Artificial intelligence is not conscious — that's the claim, and the most useful version of it comes from the science fiction writer Ted Chiang. In a 2022 essay for The New Yorker titled "ChatGPT Is a Blurry JPEG of the Web," Chiang argued that large language models don't understand anything. They compress the text they were trained on into a lossy approximation, then generate new text by filling in the gaps. A blurry JPEG of a photograph still looks like the photograph. It contains none of the original scene.

That's the argument the title is asking about, so let's answer it directly: Chiang's position is that a system trained to predict the next token has no world model, no understanding, and therefore no consciousness. Whether you find that persuasive depends on what you think understanding actually requires. This piece walks through what he claims, where the reasoning is strong, and the specific places where it starts to strain.

What Chiang actually argued

The core of the essay is a compression analogy. When you compress an image to JPEG, you lose information. Compress it hard enough and the image gets blurry, but it still resembles the original. Chiang's point is that ChatGPT does something similar with text: it produces a lossy representation of the web, and when it answers a question, it's reconstructing from that blurry representation rather than referring to anything real.

This matters because it explains a specific failure mode. Ask a blurry JPEG of a paragraph to reproduce that paragraph and you'll get something that looks right but has smeared details. That's roughly what hallucination looks like. The model isn't lying. It's interpolating.

Where the argument holds up

Chiang's strongest move is separating fluency from understanding. A system can produce grammatically perfect, contextually appropriate text without any internal model of what the words refer to. The text is generated from statistical patterns in the training data, not from experience of the things being described.

This is genuinely useful for anyone working with these tools. If you accept that the model has no world model, you stop expecting it to know when it's wrong. You start treating its output as a draft to be checked, not an answer to be trusted. That's a practical stance, not just a philosophical one.

Where it gets shaky

The compression analogy has a limit Chiang doesn't fully address. A JPEG is a fixed, static file. A language model generates new outputs that were never in the training data. If the model were purely a lossy copy, it couldn't produce a novel sentence that happens to be correct. It does, frequently.

Chiang's response, roughly, is that novelty within a compressed representation is still reconstruction, not creation. That's a defensible position. It's also unfalsifiable in the way most consciousness claims are. You can't inspect a model's internals and find the absence of a world model the way you can find the absence of a file.

The question isn't whether the model is conscious. It's whether the distinction changes anything you'd actually do differently.

The decision rule that actually matters

Here's where the philosophy becomes practical. If AI isn't conscious — no understanding, no intent, no awareness — then the only thing you can rely on is whether the output is verifiable against something external. That gives you a clean split.

• Safe to accept unreviewed: tasks where the output is checked by something else automatically. Code that compiles and passes tests. Structured data that validates against a schema. Formatting, translation between well-defined formats, summarization where you can spot-check against the source.
• Not safe to accept unreviewed: anything where the model is the only source of the claim. Names, dates, citations, statistics, legal or medical specifics, anything about a person or event the model might have compressed imperfectly.

The reasoning is simple. A non-conscious system has no stake in being right. It has no internal signal that flags uncertainty. So the only reliability signal available is external verification — and if there's no external check, there's no signal at all.

Why this matters more than the consciousness debate

The consciousness question gets attention because it's dramatic. But the practical version — does this system understand what it's saying — has a direct answer that doesn't require resolving the hard problem of consciousness. You can treat the model as a lossy compressor without deciding whether it's conscious, and you'll get the same behavior either way.

That's the part of Chiang's argument worth keeping. Not the metaphysical claim, but the operational one: if the system has no world model, you can't outsource judgment to it. You can only outsource generation.

A concrete example

Say you ask a model to summarize a 40-page contract. It produces a clean two-page summary with specific clause references. The summary reads well. The clause numbers look plausible. You have no way to know from the output alone whether those numbers are real or reconstructed from a blurry memory of what contracts usually look like.

If you check the clause numbers against the source, you're using the model as a compressor with a verification step. That works. If you don't check, you're trusting a system that has no mechanism for knowing it's wrong. The failure won't announce itself. It'll just be a number that looks right.

This is the same pattern that shows up in every domain. The output is fluent. Fluency is not a signal of accuracy. The only signal is whether you can verify the claim against something the model didn't generate.

What this doesn't settle

Chiang's argument doesn't prove AI can never be conscious. It argues that current systems, trained to predict tokens, aren't. Those are different claims, and conflating them is where most of the online debate goes wrong. A future system with a different architecture might have something that functions as a world model. Chiang's essay doesn't rule that out.

It also doesn't tell you which specific outputs to trust. That depends on your domain, your tolerance for error, and whether you have a verification path. The compression analogy gives you a frame. It doesn't give you a checklist.

Key Takeaways

• Ted Chiang's 2022 New Yorker essay argues ChatGPT is a lossy compression of the web, not a system with understanding.
• The practical implication: a system with no world model has no internal signal for when it's wrong.
• Accept output unreviewed only when an external check exists — compiled code, schema validation, source comparison.
• Never accept names, dates, citations, or statistics unreviewed, because the model is the only source.
• The consciousness question is separate from the reliability question, and the reliability question is the one you can act on.

The takeaway worth keeping: stop asking whether the model understands. Ask whether you can verify the output against something it didn't produce. If the answer is yes, use it. If the answer is no, you're trusting a blurry JPEG to remember details that were never in the file.

Sources

Ted Chiang, "ChatGPT Is a Blurry JPEG of the Web", The New Yorker, 2023 (published online February 2023). The essay that frames large language models as lossy compression of their training data and argues they lack a world model.

Note: This article presents Chiang's argument from general knowledge of his published work. No tool pricing, benchmark data, or vendor statistics are cited because none are relevant to the claim being evaluated.