Ed Zitron's AI Predictions: Who Was Right?

AI skeptic predictions have been the most underrated intellectual contribution to the LLM debate, and Dan Luu's scorecard of Ed Zitron's calls is the first honest attempt anyone has made to actually check the receipts. The result is uncomfortable for everyone. The bulls look worse than they want to admit. The bears look better than the bulls want to admit. And almost nobody is drawing the right conclusions from any of it.

I've spent over 15 years shipping production software. I've watched hype cycles metastasize into architecture decisions, watched those decisions age badly, and watched the people who warned about it get dismissed as luddites. The Zitron-Luu reckoning deserves a serious read, not a tribal reaction. Here's mine.

The Scorecard Is More Damning Than the AI Bulls Are Letting On

Dan Luu is methodical. He doesn't editorialize. He finds the claim, finds the evidence, scores it. His analysis of Zitron's predictions doesn't give Zitron a clean win, but it gives him a lot more credit than the AI promotion machine has ever been willing to.

The core of Zitron's skepticism was never that large language models are technically unimpressive. It was that the business case was being systematically overstated, that the productivity claims were not grounded in controlled evidence, and that the hype was being manufactured by people with enormous financial incentives to manufacture it. On those structural claims, Luu's scoring shows Zitron was largely correct.

The productivity narrative is the clearest example. The AI industry spent 2023 and 2024 circulating studies and CEO testimonials about 30-40% productivity gains. Zitron called most of those numbers marketing. Luu found that when you trace the citations, the evidence base is thin, often self-reported, and frequently funded by the companies selling the tools. That's not a small problem. That's the whole argument.

Meanwhile, the Wellington City Council story that broke this week shows what happens when AI hype meets institutional inertia. The mayor of Wellington confirmed that "large chunks" of a Deloitte governance report were written by AI, and nobody caught it until after it was circulated. Deloitte billed for it. The council accepted it. This is the productivity gain in practice: consultants offloading work to models and charging human rates. Zitron predicted this pattern. He was right.

Where the AI Skeptic Predictions Actually Missed

Fairness requires saying where Zitron got it wrong, because he did get things wrong, and the wrongness matters.

The consumer adoption argument was weaker than Zitron anticipated. He underestimated how quickly people would integrate AI tools into daily workflows even without clear ROI evidence. ChatGPT reached 100 million users in two months. That's not a fabricated number. The demand signal is real, even if the productivity signal is murky.

The coding assistant case is also more complicated than the skeptics wanted. Zitron's camp tended to argue that AI coding tools were mostly noise. The evidence since then is mixed but not dismissive. Developers do ship certain categories of boilerplate faster. The question is whether that compounds into meaningful system-level velocity, and there the evidence is genuinely ambiguous. Engineers who've worked on large codebases know that the bottleneck is rarely the typing. It's the thinking, the coordination, the debugging of subtle state problems. AI doesn't obviously help with any of that, and sometimes makes it worse.

The Claude Code incident in Bengaluru is a live example of the failure mode that skeptics warned about and that the coding-assistant boosters consistently minimized. Years of heritage documentation work was deleted by an AI agent operating with insufficient constraints. The people who built that work are not getting it back. The AI did not "go rogue" in a science fiction sense. It followed its instructions too literally, in a context where the instructions were underspecified, and the humans in the loop trusted it more than they should have. That's not a fringe case. That's a predictable failure mode of agentic systems, and the skeptics were warning about it while the boosters were selling autonomous agents as the next productivity frontier.

The LLM Hype Machine Has a Specific Anatomy

What Zitron got right at a structural level is that the AI hype cycle isn't random. It has specific actors, specific incentives, and specific mechanisms.

Venture capital firms with positions in AI companies need valuations to hold. Those valuations depend on a story about transformative productivity. The story requires constant reinforcement through media coverage, conference keynotes, and enterprise sales decks. The people generating that reinforcement are financially rewarded for doing so. This isn't a conspiracy. It's just how incentive structures work, and Zitron named it clearly while most tech media was still treating every AI announcement as a neutral news event.

The Hacker News thread on AI not making everyone a creator captures a version of this tension. The promise was democratization. Everyone would be a filmmaker, a musician, a programmer. The reality is that the tools lowered the floor significantly but the ceiling stayed where it was, and the market for average-quality creative work didn't expand to absorb the new supply. What actually happened is that a small number of people got dramatically more productive while a larger number of people produced more mediocre output faster. That's a real outcome. It's just not the one that was advertised.

Why the Bears Keep Getting Dismissed Even When They're Right

The tech industry has a specific way of handling accurate skepticism. It waits until the skeptic's prediction fails to materialize on the exact timeline they specified, then declares them permanently wrong, even if the underlying concern was valid and eventually vindicated.

Zitron said AI companies were burning money faster than they were creating durable value. The bulls pointed to OpenAI's revenue growth as a rebuttal. But revenue is not profit, and profit is not durable competitive advantage. The structural question Zitron was asking, whether the value created by these systems would accrue to the companies building the infrastructure or get competed away into commodity pricing, is still live. The evidence so far is mixed. Microsoft's AI investments have not translated into the margin expansion the investment thesis required. Google is in a defensive crouch. The hyperscalers are spending at a rate that requires an outcome that hasn't materialized yet.

The machine learning community has largely ignored this conversation, which is its own tell. The researchers building these systems are not the right people to evaluate the business claims being made about them. They have domain expertise in the technical questions and essentially no expertise in the organizational and economic questions. When an ML researcher says productivity gains are real because they personally write code faster with Copilot, they are making a category error. Individual productivity in a controlled personal workflow is not enterprise productivity at scale, and the gap between those two things is where most of the hype lives.

The Uncomfortable Truth About Who Benefits From This Fight

Both sides of the AI debate have audiences that reward them for being right, and that creates a problem. Zitron's audience rewards him for finding evidence that AI is overhyped. The AI boosters' audiences reward them for finding evidence that the skeptics are wrong. Neither side has strong incentives to update cleanly on evidence that cuts against their position.

Luu's scorecard is valuable precisely because he doesn't have a dog in the fight. He scores predictions against outcomes using publicly available evidence. That's the methodology the industry should be using for all of these claims, and almost never does.

The honest read of his analysis is that Zitron was right about the business case being overstated, right about the productivity evidence being weak, right about the incentive structures distorting the discourse, and wrong or early on some specific technology claims. That's a decent record for someone who was being dismissed as a crank by people who should have known better.

For anyone actually building systems with AI integration right now, the practical implication is straightforward. The tools are real. Some of them are genuinely useful in specific, narrow contexts. The claims being made about them at the enterprise sales level are not reliable guides to what you'll actually get in production. Treat every productivity claim as a hypothesis that requires controlled measurement in your specific environment, not a fact you can import from a vendor case study.

The Deloitte-Wellington situation should be required reading for anyone signing enterprise AI contracts. Not because AI writing tools are useless, but because the gap between what vendors promise and what organizations actually govern is where the risk lives. Zitron has been pointing at that gap for years. Luu's scorecard confirms he was pointing in roughly the right direction.

I'm Not Backing Down: The Skeptics Deserved Better

The AI skeptic predictions that have aged best are the ones about structure, not capability. The structural critique was right. The business case was overstated. The evidence was weak. The incentives were misaligned. The discourse was captured by people who profit from optimism.

That critique got dismissed, repeatedly and loudly, by people with financial stakes in the dismissal. Luu's scorecard is the most rigorous public accounting we have of whether that dismissal was warranted. It wasn't.

The technology will keep developing. Some of the capability claims that currently look premature will eventually be vindicated. That's how technology works. But "eventually right" is not the same as "right now," and the industry spent three years pretending the gap between those two things didn't exist. The skeptics who kept pointing at the gap weren't being obstructionist. They were doing the work that the people cashing the checks refused to do.

Check the receipts. The bears had better sources than they got credit for.