Ed Zitron's AI Predictions, Fact-Checked: What Dan Luu Found

A software engineer best known for fact-checking decades of futurist predictions just ran the same test on the AI industry's most-cited skeptic — and the results landed on the Hacker News front page with over 900 points.

In early September 2026, Dan Luu published an analysis at danluu.com titled "How accurate have Ed Zitron's AI skeptic predictions been?" Ed Zitron writes the newsletter Where's Your Ed At and hosts the podcast Better Offline, and is arguably the single most widely quoted voice arguing that the AI industry is an unsustainable bubble. Luu compiled Zitron's dated, checkable predictions from 2024 through 2025 and scored them against what actually happened. Most were wrong.

This matters beyond one blogger's scorecard. If you're a builder or a business leader trying to decide how much AI infrastructure to commit to, whether to hire for an AI team, or how much weight to give the loudest skeptical voice in your feed, you need a way to tell a well-reasoned warning from a confidently wrong one. Luu's post is a worked example of exactly that exercise — and it's worth understanding both the specific findings and the method behind them, because the method is reusable on the next commentator you're tempted to trust.

TL;DR

The methodology: an LLM builds the list, so the human can't cherry-pick

Luu's post isn't the first time he's run this exercise. He previously published a much larger review of futurist predictions — Ray Kurzweil, Michio Kaku, Patrick Dixon, and others — scoring each one's specific, dated claims against reality. The results were brutal across the board: Kurzweil, despite claiming an 86% accuracy rate for himself, scored around 7% by Luu's count; Kaku scored around 3%. Even the "best" traditional futurist in that review, Patrick Dixon, hit only about 10%. The lesson from that piece was that confident, high-profile forecasting about technology is usually wrong, regardless of whether the forecaster is an optimist or a pessimist.

For the Zitron piece, Luu applied the identical discipline in the opposite direction. Rather than hand-picking predictions he already suspected were wrong, he used ChatGPT to generate an initial, unbiased list of Zitron's dated, falsifiable claims from 2024-2026, then read the underlying posts and removed anything too vague or tautological to score. He then had both ChatGPT and Claude fact-check the resulting draft against outcomes, noting that the two models caught somewhat different errors. As Luu puts it, he doesn't "have a particular bias towards a view that rapid progress is inevitable" — a claim backed by the fact that his prior piece savaged AI optimists using the exact same method.

What Zitron predicted, and what actually happened

Luu documents roughly 30 dated predictions spanning February 2024 through late 2025. A representative sample, with Luu's verdicts:

That's a striking pattern for a commentator whose predictions, per Luu, are typically stated with "the highest possible degree of confidence." Luu writes he has "never been wrong about a prediction that has anywhere near the confidence Zitron gives" — a pointed observation from someone who makes technical predictions professionally and tracks his own hit rate.

The methodology critique: metrics, spreadsheets, and hero-villain framing

Luu's critique goes beyond the individual misses to how Zitron builds his case. Three recurring issues stand out:

Third-party metrics over company financials. Zitron cited SimilarWeb's estimated monthly active user figures for Meta's AI products rather than the company's own reported numbers — despite, per Luu, acknowledging elsewhere that SimilarWeb-style traffic estimates are "quite inaccurate and generally useless." Leaning on the weaker data source when it supports the argument is a pattern Luu flags repeatedly.

Spreadsheet errors in the financial analysis. Luu cites other technical commentators — Timothy B. Lee and Juho Snellman — who independently found computational mistakes in a widely shared Zitron spreadsheet projecting Anthropic's revenue. The errors included counting the same span of days in two different months and a literal "February 30" date that doesn't exist on any calendar. Snellman separately characterized Zitron's rhetorical approach as a "10k word gish gallop" — enough scattered claims in one piece that a careful rebuttal takes far longer to write than the original argument did.

Hero-villain narratives instead of multi-cause explanations. Luu points to Zitron's habit of attributing complex organizational outcomes — like Google Search's quality problems — to a single named executive's incompetence, rather than the more boring, more accurate reality of multiple contributing causes. Luu is careful to say he agrees Google Search has real quality problems; his objection is to the causal story, not the underlying complaint.

This is not a defense of AI-industry hype

The most important nuance in Luu's piece is what it doesn't argue. Luu explicitly states he has "never had a particularly strong pro or anti AI progress position," and discloses he holds AI-adjacent exposure only through ordinary index funds — no direct stake in the outcome either way. His prior work, the futurist-predictions piece, is aimed squarely at AI optimists and found their prediction records were, if anything, even worse than what he found here.

So the claim isn't "AI skepticism is wrong" or "the AI bubble concern is baseless." It's narrower and more useful than that: a specific, highly cited commentator's specific, dated predictions have mostly not come true, at a confidence level that doesn't match his hit rate. Whether the AI industry is overbuilt on debt and circular financing is a separate, still-open question explainx.ai has covered from multiple angles — see the off-balance-sheet debt and Microsoft-OpenAI circular revenue pieces — and nothing in Luu's post resolves it either direction. It just means Zitron's specific track record shouldn't be the evidence you point to on that question.

A practical checklist for weighing any AI commentator's track record

If you're deciding how much weight to give a hype narrative or a doom narrative before you commit budget, headcount, or a public strategy statement to it, Luu's method is reusable on anyone:

1. Find the dated, falsifiable claims, not the vibes. "AI is overhyped" isn't checkable. "This company can't survive six months without more funding" is. Weight commentators by their record on the second kind of statement, not the first.
2. Check whether confidence language matches the hit rate. A commentator who hedges appropriately and is right most of the time is more trustworthy than one who's maximally confident and right half the time — confidence is not evidence.
3. Separate quantitative claims from qualitative ones. "Revenue will miss $X by date Y" is scoreable against a filing. "Capabilities have peaked" is scoreable against new releases and benchmarks. "This is a bad sign for the industry" mostly isn't — treat it as opinion, not a data point.
4. Prefer primary financials over third-party estimates. SimilarWeb-style traffic scrapes, unverified spreadsheets, and screenshots are weaker evidence than a 10-K, an earnings call transcript, or an official usage disclosure — for either side of an argument.
5. Watch for hero-villain framing. If an outcome as complex as a product's search-quality decline or a company's stock move gets pinned entirely on one executive's incompetence, that's a simplification that should lower your confidence in the rest of the argument, not raise it.
6. Run the same test on the people you already agree with. Luu's own credibility on this piece rests on having applied identical rigor to AI optimists first. A one-sided fact-check of only the side you already doubt isn't a fact-check — it's advocacy.

The takeaway

Dan Luu's post isn't really about Ed Zitron specifically — it's a demonstration of a method anyone weighing AI commentary can use themselves: compile the dated claims, check them against outcomes, and see if the confidence matched the accuracy. Applied to Zitron, the most widely cited AI skeptic, the record through mid-2026 doesn't support the confidence his commentary is usually delivered with. Applied previously to famous AI optimists, the record was arguably worse. The consistent finding across both of Luu's pieces is less about which side of the AI-bubble debate is correct and more about how rarely either side's most confident voices turn out to have earned that confidence — which is exactly the thing to check before a prediction, not two years after it.

Related on explainx.ai:

• Ed Zitron Predicted OpenAI Would Collapse — Two Years Later, Score It
• The AI Bubble in 2026: Is It Popping, Deflating, or Just Getting Started?
• Microsoft's AI Revenue Is Mostly OpenAI Paying Its Own Bill
• Nvidia's $500B Plan to Make GPUs an Asset Class — and Why Its CDS Doubled
• AI Giants Carry $1.65 Trillion in Off-Balance-Sheet Debt — Is It Another Enron?
• BIS: Financing the AI Boom — Cash Flows to Debt, Private Credit, Equity Split
• Chinese Firms Predict the AI Bubble Is Close to Bursting
• OpenAI's $400M Startup Fund: What Platform Money Actually Changes for Builders

Primary sources: Dan Luu, "How accurate have Ed Zitron's AI skeptic predictions been?" · Dan Luu, "How well do the smartest and most famous futurists do?" · Hacker News discussion

Prediction outcomes and figures reflect public reporting available as of early September 2026; some of the underlying financial and user-growth numbers may be revised in later company disclosures.