How accurate have Ed Zitron's AI skeptic predictions been?

Scoring the record of a prominent AI skeptic

Ed Zitron has become one of the most frequently cited AI skeptics, with a newsletter and a media presence that regularly declares the end of AI progress. But how accurate are his predictions? To answer that, this review examines his claims against actual outcomes — and while the focus is on his record, his underlying reasoning matters just as much as the results.

A quick disclosure: the author of this review has no financial stake in AI companies beyond index funds, has a track record of criticizing AI hype (including a 2022 review that found futurist predictions largely wrong), and has repeatedly argued that AI's ability to displace workers is underestimated. The position here is plain: claims that current progress is impossible tend to be wrong.

A representative case: Meta, Google, and Microsoft

To understand Zitron's pattern of argumentation, it helps to look at one of his claims in depth. In a November 2024 talk, Zitron said that major tech companies like Meta and Google are dying, and that they are thrashing around on AI because they don't know how to grow. He called Meta, specifically, "a dying company." The company's actual revenue and profit numbers tell a different story:

He also named Google and Microsoft as companies that no longer know how to grow, suggesting they're shoving AI into everything out of desperation. But both companies have seen strong growth in revenue and profit:

Microsoft's numbers (here, calendar year rather than fiscal year) follow the same pattern:

One could argue Meta is dying but hasn't finished dying yet. But that doesn't rescue the core reasoning — these ecosystems are growing quickly in both revenue and profit. That growth contradicts the claim that AI is a desperation move from companies that are out of ideas.

Part of Zitron's case for Meta relies on an alleged Facebook MAU decline. His numbers appear to come from Similarweb, a third-party tracking service. Third-party estimates like these are notoriously unreliable and are generally only useful for rough order-of-magnitude comparisons. Facebook stopped publicly reporting MAU data in December 2023, but most estimates show usage increasing over time, and what Meta does report shows generally increasing usage. The cited decline appears to cherry-pick an outlier low estimate.

For Google, Zitron blames Prabhakar Raghavan — described as "truly evil" and "a computer scientist class traitor" — for grievous damage to Google search. But the claim that Raghavan is causing severe harm isn't credibly established, and Google search engineers who've commented on the rant don't seem to agree. Even if one accepted the premise that search quality has declined under Raghavan, that wouldn't demonstrate that Google's revenue growth is in trouble; other products like YouTube and Google Cloud could drive growth even if search stagnated.

The broader point: Google has indeed been incrementally making ads look more like search results over many years, each A/B-tested change small enough to avoid a fight, because each change makes more money. But that doesn't mean Google is running out of room to grow. Whether or not one approves, it's a revenue-positive move.

How his arguments are actually used

People who cite Zitron typically present him as someone who has "looked at the numbers" and concluded that AI is overhyped. But when you read his actual claims and know something about the topic, the numbers don't connect to a coherent argument. In many cases, they don't support his conclusion at all.

Others with strong attention to detail have reached similar conclusions. One Hacker News commenter, Juho Snellman, describes Zitron's work this way:

"His writing is certainly flamboyant, but the aggression and expletives seem more targeted at hyping up people who already believe the things he writes, not for making people change their minds. He found a niche in anti-tech grift, and is now exploiting the niche for all he can. But you might want to actually fact-check a few of the things he says that convince you, because at least for his written articles basically everything is made up or misrepresented. There's plenty of links to sources, sure, but if you follow them down to the primary source what they're saying is very different from what Zitron is implying."

A concrete example comes from Timothy B. Lee, who examined a spreadsheet Zitron used to project Anthropic's revenue:

"He doesn't count February 1-10, counts March 1-10 twice, counts August 21-October 21 as one month instead of two, and doesn't count October 21-November 1. [another commenter notes that his spreadsheet also contains February 30] ... Ed claims he tried to compute Anthropic's revenue for 2025 and came up with $3.6 billion, suggesting some funny business [but the numbers work out once you fix the errors]."

A prediction record

Below is a list of Zitron's major predictions and their outcomes:

• Feb 2024: "I believe we're reaching the upper limits about what generative AI can do and how accurate its outputs can be." — Wrong.
• March 2024: "Have We Reached Peak AI?"; hallucination limits mean AI progress is capped at then-current levels. — Wrong.
• April 2024: "artificial intelligence companies are running out of data"; models can't improve. — Wrong.
• June 2024: OpenAI growth is stalling, leading to collapse. — Wrong (a collapse might happen, but not from a 2024 growth stall).
• July 2024: "Generative AI, as I said back in March, is peaking, if it hasn't already peaked." — Wrong.
• July 2024: "Generative AI models aren't getting more energy-efficient, nor are they getting more 'powerful' in a way that would increase their functionality." — Wrong; models continued to improve.
• Aug 2024: "generative AI is a dead-end technology that has peaked." — Wrong. Also restated that the AI bubble had 3 quarters to prove itself (from March 2024). — Wrong, though arguably right in that AI proved itself; Zitron also argues no improvement, so wrong by his own accounting.
• Sep 2024: "o1 shows that OpenAI is both desperate and out of ideas." — Wrong.
• Oct 2024: OpenAI's forecasts of $3.7B (2024), $11.6B (2025), and $100B (2029) are absurd — "a statement so egregious that I am surprised it's not some kind of financial crime to say it out loud." — Wrong; OpenAI exceeded these goals.
• Oct 2024: "[OpenAI revenue] growth is already slowing, and will slow dramatically as we enter the new year." — Wrong; growth continued.
• Jan 2025: "I believe we're at peak AI." — Wrong.
• Jan 2025: "DeepSeek has commoditized the [LLM]." — Wrong; OpenAI and Anthropic retain pricing power.
• Feb 2025: Anthropic's $34.5B revenue forecast for 2027 is "laughable on many levels." — Wrong; Anthropic's 2026 ARR makes the 2027 estimate plausible.
• Feb 2025: Google targeting 500M Gemini users by end of 2025 is "a number so unrealistic that someone at Google should have been fired, and that someone is Sundar Pichai." — Wrong; Gemini hit 750M users.
• Feb 2025: "Sam Altman deputizing Orion from GPT-5 to GPT-4.5 suggests that OpenAI has hit a wall." — Possibly right about GPT-4.5 being unexciting, but wrong if implying models can't improve (GPT-5 was a major step up).
• Feb 2025: "I will keep writing this stuff until I'm proven wrong." — He continues writing despite repeated refutations.
• Mar 2025: CoreWeave "will not be able to survive for six months except with fundraising." — Wrong; CoreWeave IPO'd at $1.5B and trades well above that.
• Apr 2025: "generative AI isn't going to do much more than it does today." — Wrong.
• Aug 2025: "These models have clearly hit a wall where training is hitting diminishing returns." — Wrong.
• Aug 2025: Cursor, "is not good enough to be sold at a price that doesn't require Cursor to incinerate hundreds of millions of dollars"; a $10B exit is implausible. — Wrong; Cursor got a $60B exit.
• Oct 2025: Asked when the AI bubble pops, answers: "No later than Q2 2026." — Wrong (the "no later than" phrasing makes this a strong, testable upper bound).
• Nov 2025: "these models aren't getting better. What we're seeing today is pretty much what they're always gonna be like." — Wrong.

After this point, most further predictions were either unfalsifiable or scheduled to resolve in the future. Note that backward-looking false statements — like the December 2024 claim that AI products have been "trapped in amber for over a year" and the January 2026 claim that models are "basically the same as they were a year ago" — were excluded, as were other statements that were simply wrong at the time.

Comparison with futurists

Zitron's performance resembles that of futurists like Ray Kurzweil, but inverted. Where futurists predicted progress that didn't arrive, Zitron predicts stagnation that doesn't arrive. In style, Zitron relies more on anger than any futurist in a prior review. In reasoning quality, he's about average — not more unreasonable than Buckminster Fuller, who thought we'd send people by radio because both people and radio waves have frequencies.

Like Kurzweil, Zitron uses numbers to project credibility that evaporates on inspection. And like Kurzweil's fans, Zitron's fans use similar tactics — citing alleged accuracy rates that don't hold up when the specific predictions are examined (Kurzweil's claimed 86% accuracy is actually 7% when the exact predictions are judged).

Michał Zalewski explains the incentive structure:

"The surest way to build [a] popular following is to articulate positions that are crisp, strong, and leave no room for doubt. You can't get too many podcast or TV appearances out of 'well, the market could go either way', 'both political parties make good points', 'there's some merit but also some hype to AI'. Or, to tap into the example in the post, 'Harry Potter is an OK book'... It's also why Ed Zitron has a blockbuster blog about how it's all just one big scam. If you take a more nuanced view, you will at best get no reaction, or at worst, you'll invite scorn from both sides."

The pattern of wrongness and confidence

A striking feature across Zitron's output is the very high level of stated confidence. His "OpenAI forecast is a financial crime" claim, his "Sundar Pichai should be fired" claim, and his "no later than Q2 2026" bubble prediction are all delivered at near-maximum rhetorical certainty. Nearly every one of those high-confidence predictions has failed. In standard Brier-score-style evaluations, a high-confidence wrong prediction costs substantially more than a low-confidence one. Zitron's record is extremely poor on any such metric.

There's also a characteristic internal contradiction to his arguments. In one breath he says Google doesn't know how to grow and is recklessly shoving AI everywhere as a result; in the next, he says Pichai's idea of getting 500M Gemini users is absurd malpractice. As Dennis Snell pointed out, if Zitron's first claim were true, Google could trivially hit any user number it wanted by doing exactly what Zitron accuses them of. The 750M figure confirms both claims were wrong: Google wasn't so desperate it had to stuff Gemini into everything, and the 500M target wasn't absurd at all.

The name-calling and hero-villain framing (Raghavan the villain, Ben Gomes the hero) is another recurring pattern. Whatever one thinks of search quality, the dynamic described isn't a single-person failure; the ads-looking-like-results change was a slow, deliberate, multi-year process that predates any one executive, and each step increased revenue.

Defense patterns and dismissal of evidence

When commenters point out that Zitron has repeatedly predicted no improvement and been proven wrong, a common defense is that "nobody refutes his actual points." This is generally accompanied by ignoring refutations that exist in the same discussion. A related move is to simply deny Zitron ever said what he's being challenged on, even when it's directly on the record. Another popular trick is whataboutery: "what about all the AI hypists who are wrong?" Other people's wrongness doesn't make Zitron right.

Some say "Zitron isn't wrong, he's just early." The implication is that we will eventually conclude that 2026 models are no better than 2024 models. For anyone who has used modern coding tools or seen side-by-side video generation comparisons — where a 2023 example looks primitive next to 2025 output, and 2026 output is better still — that seems like an untenable claim. Coding models have improved most dramatically. In one review, a new regex engine with an interpreter and a native code compiler was created and used to speed up ripgrep in minutes of human time, a project that would likely cost seven figures in human expertise before LLMs.

On the interview circuit, Zitron tends to deny improvement and then pivot — after the interviewer points out that users notice improvements and asks about it directly, the response can be "have they?" Even broader claims about models not improving fall apart against what's visible in practice.

What actually matters

Even if one of Zitron's financial collapse predictions eventually lands — some AI company will no doubt fail at some point — that would be beside the point. The societal changes driven by current and next-generation model capabilities will continue whether or not specific companies survive. Particular companies failing doesn't undo the existence of capable models; it only affects who benefits.

This is, of course, precisely why a prediction record review like this matters. Zitron isn't just wrong about outcomes. The reasoning behind the predictions is wrong in ways that are structurally similar to the wrongness of futurists who kept predicting imminent breakthroughs. Taking the flip side of a bad bet doesn't make it a good bet.