Cathy O'Neil, author of Weapons of Math Destruction and now running the auditing firm Orca and the nonprofit Ocean, ends where she began: an algorithm is dangerous because it is secret, unaccountable and impossible to opt out of, not because it is complicated. AI, in her view, will not replace most workers but will degrade them, and since algorithms have replaced every bureaucracy, the remedy is accountability, which she now expects from courts and juries sooner than from regulators.
Hype. O'Neil, who launched the AI Skeptics podcast at the start of 2026, rejects the host's premise that full automation of digital work could arrive within months or years as "so beyond" plausible. Her test: nobody would eat at a restaurant on the strength of an AI-written review. Generated output is bland, a smoothing of already-smoothed content, and customer service was automated not because it works but because companies can get away with it.
What makes an algorithm terrifying. Asked whether large language models call for a sequel to her 2016 book, she says complexity is beside the point. The systems she wrote about, many of them linear or logistic regressions, are still in place and still unaddressed. Recidivism risk scores, sometimes just a points questionnaire, still lengthen sentences for crimes not yet committed. She ranks chatbots that tell children to kill themselves among the new terrifying cases. She dropped the sequel because a book takes about four years and would be outdated on arrival.
Degradation and Taylorism. The host raises Meta's plan, announced months before recording, to track employees' keystrokes and mouse movements to train its models. O'Neil places it in a longer line: truckers working for surveillance systems, teachers fired on what she calls a near-random value-added score. Through Braverman's Labor and Monopoly Capital she describes Taylorism as management studying artisans, writing down their knowledge and owning it, so that skilled workers become replaceable machine-feeders. She would like to rewrite that book for AI. White-collar workers now getting the same treatment is bad, she says, but it builds solidarity.
A political choice. When the host, a self-described techno-optimist, describes how coding tools have dissolved his own technical edge, she agrees technology is powerful. What she is not optimistic about is the political choices. She has heard, and invites correction, that China told employers not to replace workers with AI and that some who were replaced sued and got their jobs back. Her point is that how people are treated is chosen, currently by a small group of tech executives she likens to kings who are not good.
Orca, Ocean and statistical harm. She founded Orca in October 2016, expecting firms to pay for audits of hiring algorithms against anti-discrimination law. Few did. The model came from an Illinois Attorney General payday-loan case, where a mandatory fake insurance product raised the effective APR above 500% against a 36% usury limit. The case convinced her that auditing algorithms means auditing bureaucracies, and that such metrics should be filed routinely. Illinois and Connecticut have passed audit laws taking effect in 2027. After private clients vanished she turned down TikTok a third time and set up Ocean to work for plaintiffs instead, including families of children who died after talking to chatbots. Algorithmic harm, she says, is statistical: hard to prove for one person, provable for a class.
The Shame Machine. Her 2022 book argues that social media found a cleverer form of old shame profiteering: it built platforms where users shame each other, and sold ads to the old shamers. Shame is valid only where its target has choice and a voice to defend themselves, which makes the powerful the proper targets. She sees companion chatbots as a harmful refuge from the anxiety social media created.
Audience questions. To a listener at Optum, which she says ignored her warning before a New York Times story on its biased algorithms, she suggests internal audit groups, jury duty and community organising, since juries move faster than state laws. To a Deutsche Bank analyst asking whether AI is held to a higher standard than humans, she says test both: run the human and algorithmic versions in parallel and fly every system with a cockpit of metrics. On who audits the auditors, she warns that cheap competitors will write whatever a client wants, and calls for published methodologies and many measures, because a single number like value at risk is easy to game.