Writing, coding, analysis, translation, design, research. Machines now do each of these faster and cheaper than most professionals, and the gap closes a little more every few months. Whatever it is you're good at, your honest five-year outlook probably includes a model that's also good at it.
The thing that keeps not showing up on that list is judgment: deciding whether the thing you're about to do should be done at all.
Amplifiers don't have opinions
A tool amplifies whoever holds it. AI is the biggest amplifier ever built, and it has no opinion about direction. The same model that drafts a cancer research proposal will draft a disinformation campaign without noticing the difference. For most of history this mattered less than it does now, because bad intent was throttled by incompetence: doing serious damage took money, staff, expertise, and time. That throttle is mostly gone. One person with a laptop can operate with the reach of a small institution, which means the kind of person doing the acting starts to matter more than what they're technically capable of.
Ethics is not a personality trait
People tend to treat ethics as either a vibe or a rulebook: you have good values or you don't, you followed the code of conduct or you didn't. Building an assessment platform around exactly this question has convinced me it's neither. Ethics behaves like a skill. It has components, people are measurably better or worse at each of them, and they improve with practice.
The most underrated component is simply noticing: registering that the decision in front of you has a moral dimension at all. Most ethical failures don't happen at the moment of choosing wrong. They happen earlier, when someone optimizes a metric without ever registering that a person was inside it. The engineers behind Amazon's experimental CV-screening model presumably never set out to penalize women; the model learned that on its own from a decade of the company's hiring data, and quietly downgraded résumés that mentioned women's colleges or women's chess clubs. Amazon caught it and scrapped the tool in 2018. Plenty of teams haven't caught theirs.
After noticing comes the part that takes real practice: reasoning when two good things conflict. Genuine dilemmas are rarely good versus evil. They're speed versus safety, loyalty versus honesty, this quarter versus this decade, and working through them without collapsing into a slogan is learned the way negotiation is learned, mostly by doing it badly a few times first.
Then there's owning what your work enables after it leaves your hands. Tom Lehrer was already singing about this in 1965: 'once the rockets are up, who cares where they come down? That's not my department, says Wernher von Braun.' It wasn't much of a defense then, and it has aged badly.
And finally, the component that actually separates people: acting when it costs you. Knowing the right call is cheap. Volkswagen employed engineers who understood exactly what the emissions defeat device was for, and at least one of them went to prison for going along with it anyway. Nothing about that failure was a knowledge problem.
Why AI raises the stakes
Partly it's scale. A biased hiring manager harms the handful of people who cross their desk; the same bias deployed in a model makes the identical decision a million times before lunch. Ethical errors used to be retail. Now they're wholesale.
Partly it's speed. Legislation and courts run on years, deployment runs on weeks, and the slow external process that used to eventually catch you has stopped keeping up. If the check isn't internal, increasingly there isn't one.
But the most corrosive part, I think, is distance. The case I'd make everyone study is the Dutch childcare benefits scandal: a tax-authority risk algorithm flagged tens of thousands of families, disproportionately ones with foreign backgrounds, as welfare fraudsters. Families were ordered to repay money they didn't owe. Some went bankrupt, children ended up in foster care, and the Dutch government eventually resigned over it. The striking thing in the aftermath was that nobody in the chain felt like the person who had done it. One person wrote a specification, another trained a system, another deployed it, thousands of caseworkers followed its output. Diffuse responsibility is how enormous harm gets done by people who were each, individually, just doing their job.
The delegation problem
The tempting response is to build the ethics into the systems themselves, with aligned models, guardrails, and safety layers, and stop worrying about the humans. The guardrails are worth building; I rely on them. But someone still decides what a model should value, which tradeoffs it makes, and what counts as an acceptable failure rate. Those are moral judgments, made in advance, by people you will never meet. You can't outsource ethics to a system whose ethics somebody had to specify. You've just moved the judgment upstream, concentrated it in fewer hands, and made it harder to see.
The quieter problem is atrophy. A population that stops exercising moral judgment because a machine handles it is a population that eventually can't evaluate the machine. Skills fade when they're not used; I can barely do long division anymore. For most skills that's fine. This one I'd rather we kept.
Look at who's winning
Here's the uncomfortable part, and the part I'm least sure what to do about.
Look at what political systems currently select for. Across democracies and autocracies alike, the people rising share a recognizable profile: fluent in lying, comfortable with manipulation, able to project certainty about anything, indifferent to being caught. I don't believe voters want liars. I think it's a filter effect. Modern politics is a permanent attention contest, all constant broadcasting and viral outrage, with no memory and no penalty for contradiction, and that contest is simply easier to win if you have no internal brake. A person with genuine scruples hesitates, concedes points, refuses the cheap attack. Each of those is a competitive disadvantage, so round after round, the filter promotes the people who feel nothing.
Now hand that class of people a technology that manufactures persuasion at industrial scale. Before the New Hampshire primary in January 2024, thousands of voters got a robocall in Joe Biden's cloned voice telling them to stay home. The consultant behind it reportedly paid a street magician 150 dollars to generate the audio. The cost of producing deception used to grow with its reach; it doesn't anymore.
I don't have a fix for the selection problem, and I'm wary of anyone who claims one. It lives upstream of individuals, in incentives and accountability and whether lying carries any cost, and changing that is slow, unglamorous work. What's actually in reach for most of us is the other end of the pipeline: what we build, what we go along with, and what we're willing to say out loud in the rooms we're in.
What it looks like in practice
None of it looks like a philosophy seminar. In practice it's a handful of unglamorous habits. Ask who is affected by the thing you're building, including the people who never opted in. Assume your work will eventually reach someone with worse motives than yours, and design for that person rather than for yourself. Say the uncomfortable thing in the meeting while saying it is still cheap. Notice when you're reaching for a justification instead of a reason, and treat 'everyone's doing it' and 'it's technically legal' as warning signs, because that's usually what they are.
I don't know exactly how the next decade shakes out, and I've stopped trusting anyone who says they do. But one bet seems safe. As capability gets cheap and universal, it stops being what distinguishes anyone, and what's left is judgment: whether the person holding all that leverage can tell the difference between what's possible and what's worth doing, and whether they'll hold that line when it costs them something.