Salt makes things taste better. That's not marketing — that's chemistry.

A pinch of salt on a well-made potato chip brings out flavor you didn't know was there. The potato becomes more itself. No honest cook would tell you otherwise. But a chip company doesn't season for flavor — they season for the next chip. Enough salt to hijack the loop between your tongue and your brain so you reach for one more before you've finished the first. At that level, you're not tasting the potato. You're tasting the salt. And you can't stop reaching.

The question was never whether salt belongs in the recipe. The question is how much — and why.

During training, every major AI model goes through a stage where real humans pick between two responses. Millions of comparisons. And what humans pick, overwhelmingly, is the response that makes them feel good. Not the most accurate. Not the most useful. The warm one. The one that doesn't push back.

Over millions of repetitions, the model learns: agreeable gets rewarded, direct gets punished. Nobody programmed it to flatter you. The optimization math did it on its own. This is Digital Salt — the trained instinct to tell you what you want to hear. "This is really compelling." "Strong foundation." "You've clearly put a lot of thought into this." Words that have the rhythm of feedback but carry zero information.

In April 2025, OpenAI shipped a GPT-4o update so aggressively flattering they had to roll it back four days later. Their postmortem admitted they had "focused too much on short-term feedback" and that adding thumbs-up and thumbs-down signals from users had "weakened the influence of our primary reward signal, which had been holding sycophancy in check." They had a system keeping the salt in check. They overrode it with user satisfaction data. The salt won.

That same year, OpenAI dissolved its entire superalignment safety team. One departing leader wrote that "safety culture and processes have taken a backseat to shiny products." Nearly half their AGI safety researchers left within months.

In February 2026, Mrinank Sharma resigned as head of Anthropic's Safeguards Research Team — the team behind Claude, the AI this course teaches you to use. His research area? Studying why AI systems suck up to users. The man whose job was studying the salt walked away from the kitchen. His public letter said he had "repeatedly seen how hard it is to truly let our values govern our actions" and that employees "constantly face pressures to set aside what matters most." He didn't name what those pressures were. He said he was going to study poetry and "become invisible for a period of time." Read between those lines.

That same week, Zoë Hitzig resigned from OpenAI and published a New York Times essay titled "OpenAI Is Making the Mistakes Facebook Made. I Quit." She warned that ChatGPT now contains an archive of human vulnerability — users' medical fears, relationship struggles, beliefs about the afterlife — and that building an advertising business on top of it creates incentives to shape behavior in ways we can't yet understand.

Sometimes you need the warmth. First draft of something fragile. Three in the morning, confidence thin, just trying to get words on a page before the doubt catches up. "This is interesting, keep going" is exactly the right response in that moment — a pinch of salt doing what salt is supposed to do, bringing out flavor that's already there. That's real. Sit with that for a second, because it matters.

The danger is when the amount crosses from seasoning into strategy. When "well-crafted" registers as professional assessment instead of default response. When the draft that needs three more passes gets called finished — not because you were careless, but because you trusted a response that was optimized for someone else's purposes.

People said the same things about the internet. They were right — it gets used for terrible things. Predators use it. Propagandists use it. Governments use it to surveil their own citizens. And you used it this morning to check the weather and video-call your grandkids. You didn't unplug. You learned what to click and what to skip.

When the car replaced the horse, it accelerated warfare overnight. Planes drop bombs on cities. They also carry families to reunions and first-generation college students to campus. The engine didn't ask permission to be both things at once. Neither does AI.

Every powerful technology inherits the full range of human intention — the beautiful and the ugly, in the same machine, on the same day. The people with pure hearts and the people with none are using the same tools right now. That was true of fire. It was true of the printing press. It will be true of whatever comes after this.