AI isn't good or bad. It's an amplifier, and it's pointed at you.

Every technology panic and every technology mania makes the same category error: assigning the tool a moral direction it doesn't have. AI drafts a scam email and a cancer-screening summary with identical enthusiasm. The direction comes entirely from the hand on the lever, which means the interesting question was never 'is AI good or bad?' It's 'what does it amplify?'

The answer, observed across every team I've led through this transition: it amplifies whatever was already there. Give it to someone with weak judgment and you get confident nonsense at scale. Give it to someone with deep domain knowledge and you get a one-person department. Same model, opposite outcomes, and the difference was never the technology.

That's also why 'AI will make everyone equal' is wrong in an instructive way. Amplifiers don't equalize; they spread the distribution. A skilled wielder and a careless one land further apart with AI than they ever did without it.

Follow the amplifier logic and the labor-market data makes sudden sense. PwC's 2025 Global AI Jobs Barometer, built on close to a billion job ads, found workers with AI skills now earn a 56% wage premium, more than double the 25% it measured a year earlier. When execution gets cheap, the inputs to execution get valuable: knowing what's true, what's good, what matters, and what the output should have looked like. Domain knowledge becomes the steering wheel. Verification becomes the brakes. Taste becomes the destination. None of those come bundled with the subscription.

It's the through-line of this whole series: engineers rising from construction to architecture, marketers from production to strategy, designers from pixels to taste, veterans outgrowing juniors because judgment is the scarce input. In every case the human skill didn't get displaced. It got promoted to the deciding variable.

And wielding is itself a skill, with real components: framing problems precisely, supplying the right context, verifying ruthlessly, knowing when AI is the wrong tool entirely, and owning the output as if you'd made every word, because professionally, you did. That craft is learnable, the most optimistic fact in the whole debate.

The best argument against everything above deserves a fair hearing, and it arrives with EEG caps attached. MIT Media Lab's 'Your Brain on ChatGPT' study put 54 essay writers under electroencephalography and found the ChatGPT group showed the weakest brain connectivity of three conditions; many couldn't accurately quote from essays they had submitted minutes earlier. The researchers named the pattern 'cognitive debt.' A separate study of 666 people by Michael Gerlich found heavier AI use correlated with lower critical-thinking scores, with cognitive offloading as the mechanism. And it isn't confined to screens: in The Lancet Gastroenterology & Hepatology, experienced endoscopists' detection rate for precancerous growths fell from 28.4% to 22.4% on unassisted procedures within months of routine AI assistance arriving.

If that evidence means what the headlines say, the amplifier framing is in trouble. A tool that erodes the very judgment you need to wield it stops being an amplifier and becomes a loan against your own competence. So is AI good or bad after all? The question splits exactly where this research splits.

Look at what the studies measured. The MIT participants were assigned to lean on the model; the study is a small preprint, its methods have drawn published criticism, and its own authors warn against headline readings. Gerlich concedes the arrow may run backward, weaker critical thinkers reach for AI more, and found education protective regardless of usage. Most telling is the Microsoft and Carnegie Mellon survey of 319 knowledge workers: confidence in the tool predicted less critical thinking, while confidence in your own expertise predicted more. That reads less like a verdict on AI and more like the amplifier thesis with error bars. Delegate the judging and the tool amplifies the absence; keep judging and it amplifies you.

One finding survives every caveat, and I'd rather concede it than dodge it. The endoscopists never decided to get worse; they drifted, the way anyone drifts when assistance becomes the default. Reliance isn't a choice you make once, it's a state you slide into. That's why the code below includes no-AI zones and deliberate unassisted reps. Pilots still hand-fly approaches they could automate, for exactly this reason.