The 90%-right answer is the dangerous one.

Wisdom that outlasts the algorithm.

──────────────────────────────

The 90%-right answer is the dangerous one.

This week I'm not writing from my desk. I'm writing from the grounds of Chautauqua Institution, where I'm teaching a course called Human + AI: Skills for Thriving to a room that runs from 18 to 85.

On the first morning, an attorney in the room told us a story he's still uneasy about. Preparing for a deposition, he'd asked an AI whether a particular statute existed in Texas. It said yes, and produced it. Clean, confident, perfectly formatted. It was also invented. He lost ninety minutes chasing a law that does not exist. "My license is my livelihood," he said. "I'm fearful now using it for any high-level work."

Hold onto the shape of that failure, because it repeated throughout the week. The next morning, a participant showed us a mobility route he'd asked an AI to build, a guide for a visitor in a wheelchair getting from the main gate to the amphitheater. The map placed the destination on both sides of the highway. Not obviously broken. Broken in the one way you'd only discover halfway there, in a wheelchair, with no good way back.

Here's the thread connecting them, and it's the line the whole week turned on: the answer that hurts you isn't the one that's obviously wrong. You catch that one. It's the one that's almost right, the one that survives your scrutiny because it looks like every other correct answer you've seen. Who notices when the AI is almost right? That's not a tooling question. It's a judgment question, and judgment is exactly the thing that doesn't come in the box.

We have been here before. In 1962, before John Glenn would orbit the Earth, NASA had already run his trajectory on an IBM computer. Glenn didn't fully trust it. His instruction has outlived almost everything else about that mission: "Get the girl to check the numbers." The girl was Katherine Johnson. He would not fly until she had checked the computer's math by hand.

Sixty years later that's no longer a quaint story about an early machine. It's a job description.

The course used a spine I call From Spectator to Steward:
A spectator watches the machine produce an answer and assumes the answer is the work.

A steward knows the answer is the cheap part now, and that the expensive part is everything around it: examining the reasoning, catching the blind spot, deciding whether the output is any good at all. That part does not get cheaper as the models improve. It gets more valuable.

And the people who hold the most of it are often the ones the industry wrote off as behind. One lady in the room spent a career spanning mainframes and PCs. COBOL and FORTRAN plus forty years of watching systems lie to her. Her take on AI output was four words: "I don't trust anything." That isn't a woman who's behind. That's a woman who has been checking the machine's math her whole life, against a dataset no model was trained on, the part she lived.

Here's the pattern I keep watching prove itself. I'll call it a pattern, not a law: the more capable the AI becomes, the more a human's judgment matters, not less. When answers are cheap, the scarce asset is the person who can tell a good answer from a merely confident one.

──────────────────────────────

→ Half-Baked.
[ COMPOUNDER] We're teaching AI literacy backwards. Most "learn AI" programs train the tools: the menus, the prompts, the model of the month. Those expire in about six months. The thing that compounds for a lifetime is judgment: taste, skepticism, knowing which question to ask and when the answer smells wrong. Teach the durable skill, not the disposable one. The tool is rented. The judgment is owned.

→ Hot Take.
Building just got cheap. Taste didn't. A teenager can now say "build me a game" and have one in two hours, then discard it in fifteen minutes. The capacity to make is no longer the bottleneck. Deciding what's worth making is. The new scarce skill isn't production. It's knowing the difference between something worth shipping...and confident slop.

→ Confession.
I walked in assuming the youngest, most fluent users would lead the exercises. The opposite kept happening. The hesitant ones asked sharper questions precisely because they weren't dazzled by how fluent the machine sounded. Fluency, it turns out, is the thing most likely to make you stop checking. I had to relearn that in my own classroom this week.

──────────────────────────────

Put another way: what's the part of your judgment you've never had to explain, because you've only ever lived it? The statute that reads perfectly…but doesn't exist. The route that looks fine until you're navigating it by wheelchair. The deal that smells off before you can say why.

That cheap answer AI handed you? It stays invisible until something breaks.

Next week I'm going to take you out of the classroom and start showing you what institutions are doing live. I've been quietly tracking where authority actually lands when a person and a machine disagree, across dozens of real cases, not one classroom. The room proved the pattern. The data is about to test it.

──────────────────────────────

The Monday Move. Run a two-oracle check on a question whose answer matters more than it sounds. Prompt two LLMs on a consequential topic and see where their responses diverge.

How that went, live: the class asked two different AI systems how much water a single data center uses in a day. The numbers came back off by roughly 20X, and neither system showed its work. Stand in that gap and write down what you'd need to know before you'd trust either figure: which facility, what cooling design, whose estimate, from what year. That list is your judgment made visible. It's the part of the work the machine can't do for you.

The Asymmetric Move (this quarter). Pair your most experienced person with your most AI-fluent one on a live decision. Do not send the veteran off to learn the tool. Ask them to check its math. One of them gets you the answer fast; the other catches what it got wrong.

The Decade Move (this decade). Build the organization that treats judgment as the scarce asset. Stop hiring only for tool-fluency that expires every six months. Start keeping and elevating the people whose judgment compounds. When intelligence is cheap, the organization designed around discernment is the one still standing.

──────────────────────────────

This week was an experiment. I created a course website that drafts itself overnight from whatever the room said that day. Then every morning, before class, I sit down and fixed what it got wrong. That's the whole issue in one object: AI is your new hire, and a new hire needs a manager. The machine writes fast and confident; you're the one who checks the numbers. The drafting is free now — which is exactly why the judgment is the job.

Hand something to your AI "new hire" this week, then be the human who reads it back before it goes out the door.

I made the class one more thing worth borrowing: a plain-language map of which AI to reach for and how to begin — no jargon, built for a room aged 18 to 85: https://chrishuberreitz.com/ai-tools.

Borrow this for someone in your life ready to move from AI spectator to societal steward.

──────────────────────────────

This newsletter is called Signals from the Curve because there are two kinds of forecasting: the curve and the cliff. The cliff says everything changes at once. The curve says it is already changing, you just have to know where to look.

This week I wrote it from the classroom. If something here changed how you're thinking, hit reply. I'll respond.

— Chris

Chief of AI & Strategy at Essential Innovations · Founder, Attainable AI · Adjunct Faculty, Columbia University · Teaching Human + AI this week at the Chautauqua Institution

chrishuberreitz.com · LinkedIn