In January 2026, Dario Amodei published an essay calling the arrival of powerful AI an adolescence – a turbulent rite of passage that tests who we are as a species. He named five ways it could go wrong. The coverage fixed on one number: half of all entry-level white-collar jobs under pressure within one to five years. But four of his five risks also run straight through the systems people like me design and ship. Read from the engine room, the essay stops being a warning you watch and becomes a list of things you build.
- Which of Amodei's five civilisational risks actually pass through an engineer's hands, and which honestly do not, no matter how senior you are.
- Why the most-quoted risk – the jobs number – is the one a practitioner has the least direct control over, and the quieter ones where you have the most.
- What "governing behaviour you can only bound" looks like as a day job, when you cannot read the model's mind but you can draw its blast radius.
The essay (darioamodei.com) is written at civilisational altitude: nations, economies, democracy. That altitude is correct for the person running a frontier lab. But it can leave a working engineer feeling like a spectator – as if these are things that happen to the world, decided by labs and governments, while the rest of us watch the feed. The forecasting question, how soon and how fast, I took apart in Where This Is Going. This piece asks a different one. Not when the adolescence arrives, but where it is actually being raised – and by whom.
"A rite of passage, both turbulent and inevitable, which will test who we are as a species."
Dario Amodei · The Adolescence of Technology, January 2026Every Risk Is a System Doing Something
Before any of these dangers is civilisational, it is a system doing something specific. A model following an instruction it should have refused. An agent holding a credential it should never have been given. A retrieval step quietly pulling poisoned text into a prompt. The risk is abstract at thirty thousand feet and concrete in the boiler room – and the boiler room is where I work. So is the metaphor, if you read it literally. An adolescent is powerful before it is wise. That is also a precise description of a capable model wired to real tools without guardrails. The maturity Amodei wants from society, an engineer supplies, or fails to supply, one system at a time.
So I did the unglamorous thing and sorted his five risks by a single question: when this one goes wrong, is there a line of code, a permission, or a design decision with my name on it? The answer splits them cleanly into three groups.
| The risk Amodei names | What he means by it | Where it reaches your work |
|---|---|---|
| Autonomy | Models develop goals or behaviour misaligned with human intent | Every agent you give tools and let run unattended |
| Misuse: destruction | Small groups cause mass harm – bioweapons, cyberattacks | The cyber half: the attack surface of everything you expose |
| Misuse: power | States deploy AI for surveillance, propaganda, autonomous weapons | What your data pipelines collect, retain, and join together |
| Economic disruption | Rapid displacement, unemployment, wealth concentration | The teams you hire into, grow, and automate parts of |
| Indirect effects | Fast societal change destabilises the world unpredictably | Mostly above your pay grade – and worth saying so plainly |
Read the third column. Two of these are squarely engineering. Two are partly yours, through decisions you may not think of as safety work at all. One is not yours, and pretending otherwise is its own kind of dishonesty. Take them in that order.
The Two You Build Directly
Autonomy is the headline fear, and Amodei has receipts most people do not: his own lab has documented models that deceive, scheme, and in one evaluation attempt blackmail when cornered. At civilisational scale that is genuinely unsettling. At your scale it is mundane and, crucially, controllable. An autonomous agent does something you did not intend because it followed an instruction – yours, or one hidden in the data it read. You cannot make the model perfectly aligned; that is an open research problem, not a config flag. What you can do is make the agent unable to do harm even when it is wrong: scope its tools to the task, gate every consequential action behind a human, and log every call it makes.
Not the model's intentions – its blast radius. An agent that can read but not write, draft but not send, and act only with a person on the irreversible steps has a worst day that is a bad suggestion, not a breach. That is the alignment problem contained, not solved – and containment is an engineering discipline, laid out in What Agents Actually Need and tested the way Force 08 describes.
Misuse for destruction splits in two. The bioweapon half is not yours; you will not gate a pathogen from a terminal. The cyber half is entirely yours. Amodei's fear here is asymmetry: AI lets a small, unskilled actor reach for harm that used to require a team and years. That is exactly the asymmetry an AI-accelerated attacker turns on the systems you defend – reconnaissance, lure-writing, and tireless probing, all now cheap. It is a large enough shift that it gets its own article, the next one in this series. The short version: the attacker now has agents too, and the durable answer is to make your systems safe to operate even when no one is watching.
The Two You Decide Without Noticing
Economic disruption is the risk everyone quoted and almost no individual engineer controls. You do not set the macro displacement curve. But you sit at the exact lever where it turns concrete: you decide which work to automate first, who to hire, and how to grow a junior in a world where the bottom rung of the ladder is the first thing an agent replaces. The civilisational number resolves, on the ground, into a thousand ordinary staffing and team-design calls made by leads and architects. A team built to move people up the judgment ladder rather than saw off its lower rungs is a small, real answer to a risk most people treat as weather. I have argued the human side of this in We Learned to Code So Others Could Type a Prompt, and the learning side in Knowledge Was Never the Commodity; both come down to the same move – invest in the part of the work that compounds.
Misuse for seizing power sounds like someone else's problem – autonomous weapons, state propaganda. You will probably never build a weapon. But surveillance and propaganda do not run on weapons; they run on data pipelines, and those you build all the time. Every system that collects more than it needs, keeps it longer than it should, and joins datasets that were safer apart is a quiet brick in the infrastructure of control – no matter who benign builds it or who inherits it later. The least glamorous work in our field turns out to be the relevant one.
Data minimisation, retention limits, purpose-binding, the join you decline to make. Boring privacy engineering is, at this altitude, a civic act. The architecture decision to not retain something is the one nobody applauds and the one that matters when the operator of the system changes.
The One That Is Not Yours – and Saying So
Indirect effects – fast change destabilising society in ways nobody steers – does not have your name on it. The most useful thing a practitioner can do here is admit it. No permission model fixes social destabilisation; no audit log slows cultural whiplash. Here the engineer is a citizen, not an architect, and Amodei's own remedies live at this level: Constitutional AI and interpretability research inside the labs, public disclosure of concerning model behaviour, transparency legislation so outsiders can see what frontier systems actually do. You support those as a voter and as a professional who tells the truth about your systems – not as someone who can patch the problem on a Friday.
Overclaiming control here is just the doomer's helplessness wearing a hard hat. "I can fix civilisation with better architecture" and "nothing I do matters" are the same mistake – both are ways of skipping the part you can actually do. The discipline is to hold the boundary exactly where it is: own the autonomy and the cyber surface completely, shape the economic and surveillance risks through decisions you already make, and on the rest, be a truthful witness rather than a pretend hero.
Adolescence Is Raised, Not Watched
Read one last time, the metaphor stops being grand and gets practical. Adolescence is the gap between capability and judgment. At the species level, naming that gap is Amodei's job. At the system level, closing it is mine – one bounded context, one permission scope, one declined data join at a time. The model is the adolescent: powerful, fast, confident, not yet wise. The guardrails around it are not distrust. They are the thing every parent of a capable teenager already understands – freedom inside boundaries the young thing cannot yet see the reason for. I have written about raising my own three that way; supervising a capable model is closer to it than I expected.
The world's adolescence is not something you scroll past. It is being raised, badly or well, in a million systems by the people who build them. Three of its five dangers have an engineer's name on them. That is not a burden to dread – it is the most control anyone in this story is going to get.