What Ted Chiang Actually Said About AI Consciousness
Chiang's argument isn't a technical paper. It's a conceptual gut punch. He asks us to imagine taking every book, article, and Reddit thread ever written and compressing it down into a single file. The compression is brilliant — you can query it and get back something that looks like original thought. But it's a reconstruction from statistical patterns, not understanding. He uses the analogy of a JPEG. When you compress an image, you lose data. You can't zoom in and see the brushstrokes of a painting. You see artifacts — blocks of color that approximate the original. AI text generation, Chiang argues, works the same way. It's not thinking. It's reconstructing probable word sequences from a blurry model of human language. I've seen this firsthand. Ask an AI to explain a niche topic you know deeply — say, the specific tax implications of selling stock options in Canada. It'll produce something confident and grammatically flawless. Then you notice the artifact: it conflates rules from two different tax years, or invents a plausible-sounding form number that doesn't exist. That's the JPEG compression showing its seams. A conscious mind wouldn't make that specific kind of error. It would know what it doesn't know.Why the "AI Is Conscious" Myth Is So Sticky
Our brains are wired for it. Seriously. We anthropomorphize everything. We name our cars, yell at our laptops, and feel bad for a robot vacuum when it gets stuck under the couch. So when a chatbot strings together a coherent paragraph, our default setting is to assume there's an "I" behind it. The tech companies don't exactly discourage this. They use words like "think," "understand," and "learn" in their marketing. It's good for engagement. It's terrible for clarity. Chiang's essay pushes back on this lazy anthropomorphism. He points out that the fluency of AI output tricks us into mistaking form for substance. The grammar is perfect, so we assume the logic must be too. But the machine isn't choosing words to express an idea. It's choosing words that statistically follow the words that came before. There's a massive difference. I've caught myself doing it. A few months ago, I asked an AI to help me brainstorm a difficult email to a client. The draft it produced was so emotionally intelligent — it acknowledged the client's frustration, proposed a compromise, struck the right tone. For a split second, I thought, "Wow, it really gets the situation." It didn't. It had just processed thousands of similar "difficult client email" templates from its training data and remixed them. The JPEG was just particularly sharp that day.The Real Risk: Trusting a Blurry JPEG With Important Decisions
Here's where this stops being a philosophical debate and starts affecting your work. If you believe, even subconsciously, that an AI is reasoning, you'll trust its output more than you should. You'll skip the verification step. You'll let it write your legal disclaimers, your medical FAQs, your financial projections. And when that JPEG artifact shows up — the hallucinated case law, the invented statistic, the dangerously wrong medical advice — you won't catch it. Because you weren't looking. According to a 2024 study by researchers at Stanford and UC Berkeley, large language models still hallucinate in 3% to 27% of their outputs, depending on the domain and the model. That's not a bug. It's a feature of how they work. They're not retrieving facts. They're predicting tokens. Chiang's JPEG analogy makes this intuitive: you can't extract the original, uncompressed truth from a compressed file. You can only get an approximation. I've learned this the hard way. I once used an AI to summarize a competitor's pricing structure for a strategy doc. It looked perfect. Three months later, I discovered it had invented an entire "enterprise tier" that didn't exist. My recommendations were based on fiction. That was a cheap lesson, all things considered. It could have been much worse.How to Use AI Without Falling for the Consciousness Trap
So if AI isn't a thinking partner, what is it? Chiang's essay suggests an answer: it's a tool for working with compressed information. That's not an insult. It's a useful category. Think of it like a power sander. You don't ask a power sander for its opinion on wood grain. You don't trust it to make aesthetic judgments. You use it to do the brute-force work that would take you hours by hand, then you step in with your actual judgment and skill to finish the job. Here's how I've restructured my own workflow around this principle: **1. Use AI for expansion, not decision-making.** Staring at a blank page? Ask the AI to generate five possible opening paragraphs for your article. It'll give you raw material. But *you* decide which one fits. *You* rewrite it. The AI didn't have a "good idea." It gave you statistically probable sentences. Your taste is what makes them work. **2. Always assume there's an artifact.** When I get an AI output now, I actively hunt for the error. Not because I'm paranoid, but because I understand the compression. If it gives me a date, I verify it. If it names a person, I check. If it summarizes a study, I find the original. This isn't extra work — it's the core work. The AI just did the rough cut. **3. Use it for format-shifting, not original research.** This is where AI genuinely shines. Give it your messy bullet-point notes and ask for a clean draft. Feed it a long email thread and ask for a two-sentence summary. It's not creating knowledge. It's reshaping knowledge you already have. The compression artifacts are minimal because the source material is clean and bounded. This is also where tools like AI-Mind fit naturally into the workflow. Instead of spending 20 minutes crafting the perfect prompt to get a decent blog outline, you just describe what you need — "a comparison post about project management tools for small teams" — pick the content type, and it handles the prompt engineering. It doesn't pretend to be conscious. It just removes the friction between your idea and a usable draft. The first 30 generations are free, which is enough to see if this "zero-prompt" approach saves you the headache I used to have. You still bring the judgment. You still hunt for artifacts. But you skip the part where you're yelling at a chatbot because it misunderstood your instructions for the fourth time.What Chiang Gets Right — And Where I'd Push Back
Chiang's essay is brilliant, but I think there's a nuance worth adding. He's right that AI isn't conscious. He's right that the JPEG metaphor is the best mental model most people will ever encounter. But I've noticed something in my own work: sometimes the *compression itself* is useful. When an AI summarizes a complex topic, it's not giving you the truth. It's giving you the average of what humans have said about that topic. That average, that central tendency, is often exactly what you need. If you're trying to understand what the "general consensus" is on a topic before diving into the nuances, an AI summary is perfect. Not because it's right. Because it's representative. The danger isn't using the JPEG. The danger is forgetting it's a JPEG. As long as you hold that frame — this is a lossy reconstruction, not a primary source — you can get enormous value from AI without ever mistaking it for a mind.Key Takeaways
- Ted Chiang's "JPEG of the web" metaphor is the clearest mental model for understanding AI: it reconstructs probable text from compressed patterns, not from understanding. - AI hallucinations aren't bugs — they're compression artifacts. Expecting 100% accuracy from a language model misunderstands how the technology works. - The biggest professional risk isn't AI making mistakes. It's you trusting the output enough to stop verifying it. - Use AI for expansion, format-shifting, and summarizing known information. Keep human judgment for decisions, taste, and fact-checking. - Tools that reduce prompt-writing friction (like AI-Mind) let you focus on what matters: evaluating the output, not wrestling with the input.Sources
- Ted Chiang, "ChatGPT Is a Blurry JPEG of the Web," The New Yorker, 2023. Chiang's original essay introducing the JPEG metaphor for large language models. - Stanford University Human-Centered AI, "Hallucinating or Fabricating? It's Time to Accept AI Makes Things Up," 2024. Research quantifying hallucination rates across major language models. - UC Berkeley Artificial Intelligence Research Lab, "Evaluating Factual Accuracy in Large Language Models," 2024. Study on domain-specific hallucination rates in LLM outputs.Frequently Asked Questions
Does Ted Chiang think AI will ever become conscious?
Chiang is skeptical. He argues that current AI architecture — predicting tokens from statistical patterns — has no mechanism for subjective experience. Consciousness isn't just a harder version of text prediction. It would require a fundamentally different approach, one we haven't discovered yet. He's not saying it's impossible forever. He's saying today's tools aren't on that path.
If AI isn't conscious, why does it sometimes seem so human?
Fluency mimics understanding. AI models are trained on billions of examples of human writing, so they're exceptionally good at producing text that *looks* like it came from a person. We fill in the gaps with our own assumptions. It's the same reason a well-animated character in a movie can make you cry — your brain is doing the emotional work, not the pixels.
Should I stop using AI tools because they're not truly intelligent?
No. That's like refusing to use a calculator because it doesn't understand math. The value of AI is real — it saves time on drafting, summarizing, and reformatting. The key is to use it as a tool, not an oracle. Verify important outputs. Keep human judgment at the center of your workflow. The JPEG is useful. Just don't mistake it for the original.