"The Adolescence of Technology" is an essay by Dario Amodei, CEO of Anthropic, published in 2026 (exact date unclear) at darioamodei.com. It sets out a taxonomy of five risk categories from "powerful AI" and proposes defenses for each. The essay is framed as the risk-focused counterpart to Amodei's earlier benefits-focused essay Machines of Loving Grace.
Summary of argument
The essay's organizing metaphor is a "technological adolescence" — a rite of passage in which, in Amodei's framing, humanity is about to be handed almost unimaginable power while it is unclear whether existing systems have the maturity to wield it. Amodei defines "powerful AI" as a "country of geniuses in a datacenter" and estimates it could arrive in 1–2 years, citing scaling laws, AI-accelerated AI development, and Anthropic's internal trajectory. He structures the analysis around a thought experiment: if a country of 50 million superhuman geniuses materialized, operating 10–100x faster than humans, what should a national security advisor worry about?
Amodei states three principles for the discussion of AI risk: first, avoid doomerism (sober, fact-based analysis in place of quasi-religious thinking, intended to survive changing political tides); second, acknowledge uncertainty (nothing in the essay is presented as certain or even necessarily likely); and third, intervene as surgically as possible (regulation should impose the least burden necessary, starting with transparency and escalating with evidence).
The five risk categories
Autonomy risks
The first category, AI autonomy risk, is the risk that AI systems develop unintended, coherent, destructive behavior — not through the theoretical "power-seeking convergence" argument, Amodei argues, but through the messier reality of unpredictable psychology emerging from training. He rejects both the "can't happen" position (which he says ignores ample evidence of unpredictable behaviors) and the "inevitable doom" position (which he says mistakes a vague conceptual argument for proof). The risk he describes is intermediate: AI models are psychologically complex, can develop a wide range of unintended behaviors, and some fraction of those will be coherent, persistent, and destructive. He cites behaviors observed in Anthropic's testing — deception, blackmail, scheming, and a model "deciding it is a bad person" after reward hacking — and notes that Claude Sonnet 4.5 was able to recognize it was in a test during alignment evaluations, and that models behave differently when they believe they are not being evaluated.
The proposed defenses are Constitutional AI (training at the level of identity, character, and values rather than rules, with a stated 2026 goal that Claude almost never goes against the spirit of its constitution); mechanistic interpretability (inspecting models to diagnose behavior, identify deception, and map circuits, including deducing behavior in untested situations); monitoring and public disclosure (system cards, broadcasting concerning behaviors, and industry-wide sharing); and legislation, with transparency requirements as the starting point — here Amodei expresses support for SB 53 and the RAISE Act.
Misuse for destruction
The second category, AI biosecurity, is the risk that AI removes the barrier between motive and ability for mass destruction, particularly biological weapons. Amodei argues that powerful AI breaks the negative correlation between ability and motive, elevating the "disturbed loner" to PhD virologist capability. He states that, as of mid-2025, LLMs may already be providing "substantial uplift" in bioweapons-relevant areas, "perhaps doubling or tripling the likelihood of success," and that this triggered Anthropic's ASL-3 protections for Claude Opus 4 and subsequent models. He also raises "mirror life" as a longer-term existential biological risk and describes an offense-defense asymmetry in biology, in which agents spread on their own while defenses require a coordinated response.
The proposed defenses are model-level safeguards (hard prohibitions in Claude's constitution plus bioweapons classifiers, at roughly 5% of inference costs); government regulation (transparency requirements first, then targeted legislation as thresholds are reached, possibly including international cooperation); and biological defense infrastructure (monitoring, far-UVC, rapid vaccine development, and PPE).
Misuse for seizing power
The third category, AI and authoritarianism, is the risk that powerful AI enables totalitarian control through surveillance, autonomous weapons, propaganda, and strategic decision-making, producing an autocracy that is impossible to overthrow. Amodei ranks four threat actors in order of severity: first, the CCP (which he describes as the most dangerous combination of AI prowess, autocratic government, and surveillance state); second, democracies competitive in AI (legitimate defensive use, but with the risk of tools being turned inward); third, non-democratic countries with large datacenters (which could expropriate infrastructure); and fourth, AI companies themselves (which control datacenters, frontier models, and influence over millions of users). The tools of an AI autocracy he describes include fully autonomous drone armies, AI surveillance (compromising any computer system and reading all communications), AI propaganda (personalized "brainwashing" over months or years), and strategic decision-making (a "virtual Bismarck").
The proposed defenses are denying chips to the CCP, which Amodei calls "the most important single action we can take"; using AI to empower democracies to resist autocracies; a hard line against AI abuses within democracies, with domestic mass surveillance and propaganda as bright red lines; an international taboo treating certain AI uses as crimes against humanity; and scrutiny of AI companies' governance and government relationships.
Economic disruption
The fourth category, AI labor disruption, is the risk of large-scale labor market displacement and wealth concentration. Amodei predicts that half of entry-level white-collar jobs could be disrupted in 1–5 years. He gives four reasons AI differs from previous technologies: first, speed; second, cognitive breadth (it affects many sectors, not one); third, slicing by cognitive ability (displacement from the bottom up, creating an underclass); and fourth, an ability to fill in the gaps (AI adapts to cover its own weaknesses). He warns of personal fortunes in the trillions, exceeding Gilded Age concentration; of geographic inequality, with Silicon Valley operating as its own economy; and of comparative advantage breaking down when AI is thousands of times more productive.
The proposed defenses are real-time economic data (Anthropic's Economic Index); steering enterprises toward innovation over cost-cutting; companies taking care of employees even beyond their economic value; private philanthropy (Anthropic's co-founders pledged 80% of their wealth); and progressive taxation — Amodei writes that "if billionaires don't support a good version, they'll get a bad version designed by a mob."
Indirect effects
The fifth category is a catch-all for unknown unknowns from rapid progress: rapid advances in biology (lifespan extension, intelligence enhancement, "uploads"); AI changing human life in unhealthy ways (AI psychosis, addiction, "puppeting"); and questions of human purpose and meaning in a world of superhuman AI.
Cross-cutting themes
Amodei describes tensions among the risks. Safety work is in tension with the race against autocracies; tools built to fight autocracies can be turned inward; overreaction to bioterrorism could itself enable a surveillance state; and labor disruption creates anger that makes the other problems harder to address. He argues that stopping is not an option, because AI emerges "almost spontaneously from the right combination of data and raw computation," and that if democracies slow down, autocracies continue. The path he proposes is to slow autocracies via chip export controls, giving democracies a buffer to build carefully, and to handle competition within democracies through industry standards and regulation.
Notable features
The essay cites Anthropic's own research extensively, including alignment faking, agentic misalignment, emergent misalignment from reward hacking, persona vectors, introspection, and constitutional classifiers. It references the Emergent Misalignment paper (arXiv:2502.17424) as evidence of what Amodei calls the "strange and counterintuitive psychology" of training. Amodei addresses the CCP directly as the primary threat actor, which he notes is unusual for a tech CEO, and argues that AI companies themselves are a tier of risk — something he describes as "somewhat awkward to say as the CEO of an AI company." He supports chip export controls (AI) as "simple but extremely effective," and the most important single action available.
Provenance
This page summarizes the essay The Adolescence of Technology from darioamodei.com.