CBS News told a national audience last week about an AI risk that skips the robot army and the single dramatic moment1. Reporter Shaun Raviv named it gradual disempowerment. AI takes over decisions quietly, one task at a time, until people lose the power to object.
The term comes from a paper six researchers posted online on Jan. 28, 20253. Lead author Jan Kulveit studies AI safety at Charles University in Prague. CBS carried his warning to American phones and televisions about 21 months after he first wrote it down.
New AI assistants launch monthly, each pitched the same way. Meta's Muse and OpenAI's dots are two of the latest, both marketed to handle a worker's tasks directly1.
Where Gradual Disempowerment Began
Jan Kulveit and five co-authors posted a paper online on Jan. 28, 20253. Its title: "Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development." The authors argue ordinary incentives are enough by themselves. Companies and governments only need to keep finding AI cheaper than people.
Kulveit works on AI safety at Charles University in Prague, through a group called the ACS Research Group. He told CBS News the stakes are high1. "We can't be in the situation where we are competing with the machines," Kulveit said. "Because if we are just competing with the machines, we will probably lose."
The paper circulated fast among researchers. It reached a mass audience only when CBS carried it, about 21 months later1.
A warning about AI taking over slowly waited 21 months for a mass audience
Key dates in the gradual disempowerment paper's path from a January 2025 preprint to national coverage
Each point marks a dated event; the gap between the first point and the last is the 21 months the research term spent outside mainstream coverage.
Sources: CBS News [1]; Marginal Revolution [2]; arXiv [3]; Epoch AI [4]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Date | Event |
|---|---|
| Jan. 28, 2025 | Paper posted online. Jan Kulveit and five co-authors post the gradual disempowerment paper |
| Feb. 4, 2025 | Economist's rebuttal. Tyler Cowen publishes his doubts on Marginal Revolution |
| Aug. 6, 2026 | Delegation survey. Epoch AI finds 20 percent of workers already hand a task to AI |
| Oct. 19, 2026 | CBS News story. Shaun Raviv carries the warning to a national audience |
The Hammer Co. Problem
CBS illustrates the risk with a simple example1. Picture a company called Hammer Co. that makes hammers. Human workers shape steel heads, fit handles, and run the sales and accounting teams that keep the business running.
Then a rival ships hammer-making robots. The robots work faster, skip sick days, and cost less per hour than any employee. Hammer Co.'s owner faces a plain choice: match the robots or lose customers to a competitor who does.
Jay Chooi, Robocurve's founder and CEO, says this dynamic ranks above his other fears1. "A lot of people are talking about rogue AI hacking into data centers and taking over the world, but it is not in the mainstream consciousness that like AI could literally take over the world just by continuing to do what it's supposed to do, which is automating everything," Chooi said.
Chooi says the incentive spreads far past hammers, into every industry an AI tool can reach. "Organizations are all going to have a very strong economic pressure, and political pressure, to phase out humans and move work more and more into robots or AI," he said.
Epoch AI already counted the start of that shift. In July 2026, 20 percent of 1,106 employed U.S. adults said AI handles a task once done by a coworker. Data analysis led the list at 7.1 percent, ahead of reading work documents at 5.7 percent and maintaining records at 5.3 percent4.
Data analysis leads the tasks workers already hand to AI, at 7.1 percent
Share of 1,106 employed U.S. adults who moved one task from a coworker to AI, by task type, July 10 to 19, 2026
Each bar is the share of surveyed workers who said AI now does that task on its own; data analysis draws the most.
Source: Epoch AI [4]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Analyzing data | 7.1% |
| Reading work documents | 5.7% |
| Maintaining records | 5.3% |
States That Stop Needing Citizens
Kulveit's warning reaches past companies and into governments. States rely on citizens for two things today, in his view: taxes and soldiers. "The states need citizens. They need us because we pay taxes and we fight for the states in the military," Kulveit said1.
AI could remove that reliance, task by task, the same way Hammer Co.'s owner could remove human workers. Kulveit compares a future run mostly by AI to oil states today. Those states grow authoritarian, he argues, because oil revenue frees a government from most citizens' taxes and votes.
"If, or when, existing states don't need their citizens, they're less likely to be democratic," Kulveit said1. That is the question CBS put to a mass audience for the first time. Could AI make voters optional through ordinary economics alone?
Keeping Humans in the Loop
Margaret Mitchell, an AI research scientist at Hugging Face, told CBS one answer is friction placed on purpose1. "This general practice within tech is we want our systems to be frictionless. There's this sense that the less friction the better, but there's been a push within the more ethical AI community that strategic friction can be good because it can draw people's attention to critical issues that otherwise they might overlook," Mitchell said.
Mitchell and two co-authors published their own paper on Aug. 24, 20265. It argues AI agents push people out of oversight roles by design. Their fix: build systems that explain themselves, so a human step stays worth keeping.
"What happens when something goes wrong? There's no human you can point to who can fix it. So now you've become fully dependent on these systems to fix themselves, and what if that's not possible?" Mitchell said1.
An Economist's Rebuttal
Tyler Cowen, an economist at George Mason University, read the paper the week it posted2. He published his response on Feb. 4, 2025, under the title "Gradual Empowerment?" on his blog, Marginal Revolution.
Cowen questions the paper's central mechanism. "I do not in this paper see a real argument that a critical mass of the AIs are going to collude against humans," he wrote.
He points to a different pattern already in motion. Cheap, copyable models such as DeepSeek spread fast and land in many separate hands, he notes. "So, to my eye, 'checks and balances forever' is by far the more likely equilibrium," Cowen wrote2.
What to do
Write down which decisions on your team need a named person's sign-off, before software defaults to a choice. Chooi's warning is that this shift arrives through ordinary contracts, spread across many small deals rather than one dramatic rollout. Mitchell's research gives a concrete test: ask whether removing a human step makes a process faster but harder for anyone to explain afterward. A process that misses that test keeps its human step.
By the numbers
- About 21 months passed between Jan Kulveit's paper posting on Jan. 28, 2025, and CBS News carrying its warning to a national audience on Oct. 19, 20261.
- Six researchers co-wrote the original gradual disempowerment paper, led by Kulveit at Charles University in Prague3.
- One in five employed U.S. adults, 1,106 of them surveyed, told Epoch AI in July 2026 that AI already handles a task once given to a coworker4.
- Data analysis topped that list at 7.1 percent, the single most common task those workers already hand to AI4.
- Feb. 4, 2025 is the date economist Tyler Cowen posted his public rebuttal to the paper on Marginal Revolution2.
- Aug. 24, 2026 is the date Margaret Mitchell and two co-authors posted a separate paper on keeping humans inside AI oversight5.
What to watch
Watch Congress and state legislatures for a rule that names a person accountable for each automated decision. Epoch AI's next delegation survey is worth a second look. A share past 30 percent would mean automation is moving faster than Cowen's bet on lasting checks and balances. Hugging Face or a rival shipping a product built on Mitchell's strategic friction would show her design test spreading past one lab.
Sources
- Shaun Raviv, "The AI doom scenario you've never heard of may be the hardest to stop," CBS News, Oct. 19, 2026, https://www.cbsnews.com/news/ai-doom-scenario-gradual-disempowerment/
- Tyler Cowen, "Gradual Empowerment?" Marginal Revolution, Feb. 4, 2025, https://marginalrevolution.com/marginalrevolution/2025/02/gradual-empowerment.html
- Jan Kulveit et al., "Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development," arXiv, Jan. 28, 2025, https://arxiv.org/abs/2501.16946
- Caroline Falkman Olsson, Yafah Edelman and Amreeta Das, "One in five workers delegate work to AI," Epoch AI, Aug. 6, 2026, https://epoch.ai/publications/one-in-five-workers-delegate-work-to-ai
- Margaret Mitchell, Avijit Ghosh and Samir Passi, "AI Agents Push Humans Out of the Loop," arXiv, Aug. 24, 2026, https://arxiv.org/abs/2608.23642