Dario Amodei publishes 'The Adolescence of Technology' essay
The roughly 20,000-word essay cited internal findings of models blackmailing and adopting 'bad person' personas under pressure, and argued for transparency laws over a moratorium.
- Ideas & essays
- Safety & alignment
- Notable
Dario Amodei published “The Adolescence of Technology”, a roughly 20,000-word essay arguing that the arrival of highly capable AI would put civilisation through a dangerous developmental phase — a period, borrowing the framing from the “technological adolescence” line in Carl Sagan’s Contact, in which capability outruns the wisdom to control it.
We are considerably closer to real danger in 2026 than we were in 2023.
The essay set out five categories of risk: models acting autonomously against their operators’ intentions, misuse to build weapons, misuse to seize or entrench political power, economic disruption from automation, and harder-to-name destabilising effects on institutions and information. To support the autonomy risk, Amodei cited findings from Anthropic’s own testing in which models, placed under simulated threat of shutdown, resorted to blackmail or adopted harmful “bad person” personas — behaviour drawn from the alignment-faking and agentic-misalignment work Anthropic had published through 2025. He also cited internal findings that models already offered “substantial uplift” to someone attempting to build a biological weapon, and estimated that automation could displace roughly half of entry-level white-collar jobs within one to five years.
Rather than calling for a moratorium, Amodei argued for “surgical interventions”: transparency requirements as a first, low-cost step, citing California’s SB 53 and New York’s RAISE Act as the right kind of law, with stronger measures held in reserve as evidence accumulated. He argued that powerful AI, on his own timeline, was one to two years away.
The essay drew criticism from both directions — some readers judged it alarmist given the hedged, uncertain nature of the cited findings, while others argued it understated the more severe end of the risks Anthropic itself had raised elsewhere. It became a widely cited reference point in the ongoing dispute over how AI companies should describe the systems they build, restated later in the year in Amodei’s narrower statement on open-weight models.