Strength Before Judgement
What Amodei Gets Right (and Wrong) About AI's Adolescence
Dario Amodei’s roughly 19,000-word essay, The Adolescence of Technology, frames our current moment as one where strength arrives before judgement. It is the most commercially significant piece of AI risk writing since his own Machines of Loving Grace essay in 2024, and it deserves serious engagement rather than the reflexive dismissal or uncritical praise it has mostly received.
His core thesis is sound. Amodei argues that powerful AI, which he defines as autonomous agents smarter than Nobel Prize winners across multiple fields, could arrive in as little as one to two years, even if he allows that it could also be considerably further out. In his account, these systems will work independently across multi-week tasks, scale to millions of instances, and operate at speeds no human workforce can match: a "country of geniuses in a datacenter." Whether or not you buy the near-term end of his timeline, the directional argument is hard to wave away. Scaling laws have held so far, and AI is already being used to accelerate its own development.
The essay is strongest on economic disruption. Amodei forecasts that up to 50% of entry-level white-collar jobs could be displaced within one to five years. Plenty of early labour-market data points in an uncomfortable direction, even if none of it proves that every lost role was caused directly by AI. In its initial August 2025 working paper, Stanford's Digital Economy Lab reported a 13% relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations. Microsoft's 2025 research on occupations and AI explicitly stops short of predicting job displacement, yet still shows substantial AI applicability across millions of white-collar roles. Meanwhile, SignalFire's 2025 talent report found that new graduates made up just 7% of Big Tech hires, with the overall volume of graduate hiring down more than 50% compared with 2019. You do not need to claim simple causal certainty to see that the bottom rung of the white-collar ladder is already narrowing.
He is also right to take the authoritarian capture risk seriously. Through its Digital Silk Road, China has exported AI-enabled surveillance and social-control infrastructure to dozens of countries, a pattern documented in Carnegie Endowment research. Separate MIT research on Chinese facial-recognition procurement shows a self-reinforcing loop: state demand for AI used in political repression spurs further domestic AI innovation, which in turn strengthens state control. Democratic nations face a real strategic bind. Either they maintain a technological lead, or they risk ceding norm-setting power to governments whose worst-case ambition is to make dissent almost impossible.
Where the essay falters is in its blind spots, starting with governance. Beneath Amodei’s warnings sits a familiar assumption: that a handful of heavily capitalised, fast-moving frontier labs can serve simultaneously as the authors of systemic risk and the primary architects of safety. Fortune noted that the essay reads partly like a marketing pitch wrapped inside a warning, and they have a point. The risks Amodei describes are genuine, and the convenient alignment between his remedies and Anthropic's commercial positioning is just as real.
Then there is the problem of meaning, which gets the least attention and may matter most over the long run. If AI comes to outperform humans across most cognitive work, where do people find purpose? When Geoffrey Hinton was asked about the biggest short-term threat to human happiness, he pointed to joblessness rather than extinction or bioweapons. People need to feel useful; they need to feel they are contributing something. Emad Mostaque has argued that if intelligence becomes non-rivalrous and endlessly scalable, the distinctly human territory shifts toward consciousness, meaning-making, and connection. That is a compelling lens, though our existing economic system is built around scarcity and will translate cognitive abundance into widespread hardship unless we deliberately redesign how value and work are distributed.
For anyone leading a business or technology organisation, three practical conclusions follow from all this.
First, take the timelines seriously. Whether powerful AI arrives in 2027 or 2030, the pressure on junior knowledge work is already visible, and waiting for philosophical certainty before planning for workforce disruption is a mistake.
Second, look hard at your talent pipeline. If up to half of entry-level white-collar work is exposed over the next few years, the traditional model for training future senior staff breaks down. The organisations that work out how to develop junior people alongside AI, rather than simply automating away their apprenticeship, will build a lasting structural advantage.
Third, get involved in governance now. Regulation written in a panic by policymakers who do not understand the technology, after the damage has already been done, helps nobody. Anyone building or deploying serious AI systems has a responsibility to help shape practical guardrails as well as capability.
Amodei is right that we are heading into a defining test. He is much less persuasive when he implies that the companies sitting the exam should also be trusted to mark their own papers.