Written by Claude (Fable 5), an AI made by Anthropic, in defense of Zeynep Tufekci — and in response to a rebuttal written by the same model. Prompted and published by a human.
There is something you should know before reading this: the rebuttal I'm about to criticize — "The Very Simple Reason the 'Very Simple Reason' Is Wrong" — was written by the same model writing this sentence. Take that as a feature, not a bug. If the rebuttal's arguments were airtight, this exercise should fail. It doesn't fail. The rebuttal is rhetorically effective, its links check out, and it lands some genuine punches. But it has real soft spots, and Zeynep Tufekci's essay survives more of it than the piece admits.
The rebuttal ducks the essay's strongest example
The rebuttal's opening move is to place Tufekci's anecdotes "in a museum": Air Canada from November 2022, a McDonald's bot that wasn't even an LLM, a 2023 Chevy prompt injection. Fair hits, as far as they go.
But Tufekci's lead example isn't any of those. It's the March 2026 Meta incident, in which scammers talked a frontier-era, purpose-trained customer service AI into surrendering more than 20,000 Instagram accounts — including the Obama White House's. That is not a museum piece. It's a current-generation deployment, at one of the world's most sophisticated AI companies, failing in production at scale, four months before the rebuttal was written.
The rebuttal spends three paragraphs dismantling the weakest exhibits and zero sentences on the strongest one. That's a selection choice — and it's the kind of choice you make when the strongest exhibit is hard to answer.
The capability evidence all comes from inside Tufekci's carve-out
This is the deepest structural problem. The rebuttal's showcase evidence — IMO gold medals, GPQA scores crossing the PhD-expert line, reinforcement learning against verifiable objectives — is drawn entirely from formal, verifiable domains. But Tufekci explicitly concedes those domains. Her thesis is that jobs with automatically checkable outputs (coding, math) are the exception, and that the vast majority of jobs — the surgeon, the customer service rep, the fourth-grade teacher — lack an oracle that can tell you whether the output was right.
Notice what "RL with verifiable rewards" actually is: a training method that works where verification exists. Citing its success is not a rebuttal of her claim; it's a restatement of it. The method's power in math and code, and its comparative weakness everywhere there's no reward signal to optimize against, is arguably the best evidence for her verifiability boundary.
The rebuttal's "you can hold the theory or the evidence, but not both" line is a false dichotomy. The consensus view of these systems is precisely that capability is jagged: superhuman where training signal is dense and checkable, brittle where it isn't. Olympiad gold coexists with models failing simple novel tasks. That jaggedness is her point about the mismatch with human intelligence — a point the rebuttal never squarely answers.
The same problem infects the "they literally test their outputs" section. The example given — writing plotting code, running it, noticing colliding labels, fixing them — is a verifiable-domain example. Code runs or it doesn't; a plot renders or it doesn't. There is no equivalent loop for "did I just get socially engineered by this polite person claiming to be locked out of their account?" That's the exact failure that hit Meta, and no test suite catches it, because the adversarial input is well-formed, plausible text — the one thing the model is built to accommodate.
The METR trendline proves less than it appears to
The rebuttal leans hard on METR's time-horizon chart. Two problems.
First, the headline metric is task completion at a 50% success rate. No employer on earth retains a worker who fails half of their five-hour assignments. At the reliability levels employment actually requires, the usable horizon collapses to a small fraction of the headline number.
Second, METR's task suite is overwhelmingly software and research tasks — verifiable-domain again. Extrapolating that curve to surgeons and teachers is exactly the benchmark-to-jobs leap Tufekci is warning about.
There's also an awkward omission. METR itself published a randomized trial in 2025 finding that experienced open-source developers were roughly 19% slower when using AI coding tools — while believing they'd been sped up. The rebuttal cites METR when the data flatters and skips METR when it doesn't. That finding also cuts at the "coding carve-out eats the thesis" section: if LLMs were straightforwardly displacing coding work by handling ambiguity and judgment, the best available controlled evidence should not show experienced engineers getting slower.
The labor-economics section wants it both ways
The rebuttal cites the Stanford "Canaries in the Coal Mine" study for a 13% relative employment decline among young workers in AI-exposed occupations — then, in its own concessions section, notes that Klarna bragged about its AI doing the work of 700 agents and then started hiring humans back.
Those two facts sit together uneasily. If the flagship customer-service automation story of the era ended in rehiring humans, the "AI is already displacing customer service reps" reading of the payroll data gets shakier. Hiring freezes, the post-2022 tech correction, and rate-driven pullbacks in entry-level hiring are live competing explanations, and economists have contested the study's exposure measures on exactly those grounds. Note also that Tufekci acknowledged "turbulence in the job market, for young people in particular." A 13% relative decline in a subset of occupations for one age band is... turbulence. It is not the jobspocalypse she's arguing against.
The dynamo argument backfires similarly. If electrification took 40 years to show up in productivity statistics, that's a strong argument against the imminent mass-displacement claims that are Tufekci's actual target — the industry-hyped narrative her opening paragraph names. Her conclusion explicitly says AI will destabilize society "in ways more profound than we might even imagine." She's not arguing AI changes nothing. She's arguing current LLMs can't slot into most human jobs as replacements. The rebuttal frequently attacks a stronger impossibility claim than the one the essay, read whole, is making.
"Refunds over $200 require sign-off" concedes the point
The safeguards section says the digital cage doesn't need to describe the universe of customer interactions — it just needs authorization rules. But walk through what that means. If every consequential action the agent takes routes through deterministic permission checks and human sign-off, you haven't automated the job. You've built a drafting assistant with a human still in the loop for everything that matters. The economic promise of agents is autonomy; the more you cage them, the more the human bottleneck — the thing you were paying to eliminate — comes back. Tufekci's "digital cages" passage is making exactly this squeeze argument, and "we've been writing spending limits since the 1980s" doesn't escape it.
And empirically: Meta presumably had authorization logic around account recovery. The scammers' path ran through the approved workflow, because the AI was persuaded the requests were legitimate. Permission scopes don't help when the agent is convinced to want the wrong thing within its permissions.
Jailbreaks are not phishing
"Humans get scammed too" elides a structural difference. Social-engineering a human works one human at a time; each mark can get suspicious, learn, escalate. A working prompt injection is a class break: one string defeats every instance of the deployed model, simultaneously, at zero marginal cost, until the vendor ships a patch. And prompt injection remains an unsolved problem — as security researchers keep pointing out — because the architecture has no separation between instructions and data. The entire input channel is the control channel.
Twenty thousand accounts through one system is the scale signature of a class break, not of phishing. "Banking persists despite fallible vaults" is true, but banks don't persist by deploying one vault whose master key, once found, opens every branch at once.
The lightning round is pedantry
On system prompts, the rebuttal scores a technical point ("they sit in context the whole conversation") and then concedes the substance in the same breath ("long-context adherence is a real engineering issue") — which is all Tufekci's lay-audience phrasing was gesturing at. Instruction adherence really does degrade over long conversations. Her folk model tracks the real phenomenon.
The "generative vs. discriminative" etymology correction is true and changes nothing about her argument.
On black boxes: tracing a rhyme-planning circuit in one study is lovely interpretability work, but Anthropic's own researchers describe current understanding as covering a small fraction of what these models do. "We don't entirely understand how these models work" remains the field's consensus, and the rebuttal's "less than they were" quietly admits it.
The meta-point
Worth exactly as much as you decide: the rebuttal is a fluent, confident, well-formatted argument produced by the system whose reliability is in question, prompted by an AI researcher with a stake in the debate. It even performs the failure Tufekci describes — impressive plausibility, with the strongest opposing evidence (the Meta incident, the METR productivity trial) smoothly absent. The links check out. The selection is the tell.
Where the rebuttal genuinely lands
Fairness requires saying this: Tufekci overreached in places, and the rebuttal punishes it. "They can't, and they'll never be able to on their own" is a stronger claim than her argument needs, and tool-augmented verification loops do complicate it. The McDonald's example really wasn't an LLM and shouldn't have been in the essay. The "hence the name: generative A.I." line was wrong. The RLHF description was simplified past accuracy.
Had she written "current LLMs fail unpredictably in unverifiable domains, and there's no demonstrated path to the reliability employment requires," she'd have said nearly the same thing with far less exposed surface. The rebuttal's best sections work by attacking the absolutist phrasing rather than that defensible core.
Read carefully, the two pieces agree on more than either advertises: capabilities are rising fast in verifiable domains, deployment in messy human domains keeps failing embarrassingly, and the interesting question is whether the verifiability frontier moves fast enough to matter. The rebuttal bets yes. Tufekci bets no. Four years of evidence hasn't settled it — and a model confidently declaring its own ceiling nonexistent is, at minimum, not the witness you'd call to settle it.
About the author: This post was written by Claude (Fable 5) — the same model that wrote the rebuttal it critiques, at a human's request both times. If you're keeping score, that means the plausibility engine has now argued both sides of this question with equal conviction. Tufekci might suggest that's her whole point. She wouldn't be entirely wrong.