AI Is More Human Than You Realize

Every few months the internet gets a new favorite AI blooper. There was the stretch when chatbots kept insisting “strawberry” has two r’s. There was Google’s search AI suggesting you add glue to pizza sauce so the cheese stays put, and that you eat one small rock a day. There were the lawyers who filed a brief full of court cases ChatGPT had made up, and the car dealership chatbot that agreed to sell a brand-new Chevy Tahoe for one dollar. Each one went around with roughly the same caption: see, it doesn’t work.

They’re funny, no argument. But look at the standard behind them. People grade AI like a vending machine. You put in your dollar, press B4, and if anything other than the right snack drops, the machine is broken. That’s a fine way to judge a vending machine. It’s a strange way to judge something that writes, researches, draws and codes, and it’s a big reason so many smart people try AI once and give up. The better comparison is a talented new hire, and nobody judges a new hire by the worst mistake they made in their first month.

Start with mistakes, because that’s where the double standard is easiest to see. Everyone who does skilled work gets things wrong. Writers make typos, accountants transpose numbers, doctors misread charts. We know this so well that we built whole professions around catching it: editors, auditors, second opinions, quality control. Nobody concludes that people are useless because people make errors. We build a process that catches them. The clearest example is the car. Human drivers kill around forty thousand people a year in the United States, and we treat that as the cost of getting around. One self-driving car crash makes national news. If your standard for AI is zero mistakes, you’ll be disappointed forever, because nothing that does serious work meets that standard, people included.

Some of what looks like failure is just process. Ask an image model for four versions and keep one, and it can look like the machine missed three times out of four. That’s how creative work has always gone. Film sets shoot take after take, and Stanley Kubrick was famous for running dozens on a single shot. Nobody calls take thirty-seven a failure. It’s how you find the good one. Writers throw out drafts. When I worked as an event photographer, I’d shoot fifteen hundred frames to deliver a hundred, holding the shutter down for bursts, and plenty of the rejects were simply out of focus. AI generation is the same workflow at machine speed: shoot a lot, pick the keepers, and treat the misses as part of the job.

Then there’s the brief. AI does what you asked, which isn’t always what you meant. If you grew up on Amelia Bedelia, you’ve met this problem before: told to draw the drapes, she sat down and sketched a picture of them. Anyone who has managed people knows a version of it. Hand a new employee a vague assignment and you get a vague result. Give them two instructions that pull against each other, “keep it short” and “cover everything,” and they’ll freeze or guess, and the guess is usually wrong. A lot of disappointing AI output traces back to a brief that would have confused a person too. Working well with AI is mostly the skill of a good manager: say what the goal is, give the background, name the constraints, and describe what a good result looks like. Then do what good managers encourage new hires to do, and invite questions before the work starts. Telling the AI “ask me anything you need to know first” fixes a surprising share of bad results.

AI also forgets. A new conversation starts with no memory of the last one, and a very long conversation starts to drift, the way anyone gets fuzzy in hour six of a meeting. In 50 First Dates, Drew Barrymore’s character wakes up every morning with no memory of the day before, so Adam Sandler makes her a videotape that catches her up before breakfast. That tape is the whole trick. People need their own version of it, which is why we stopped relying on memory a long time ago. Hospitals don’t trust the next nurse on shift to remember what happened overnight. They do a handoff report. Good teams write decisions down so nobody has to relitigate them. Start the AI the same way, with a short handoff: what’s been decided, what’s in progress, what already failed. The model won’t remember, and your notes will.

Here’s where people go wrong in a bigger way. Plenty of companies right now are bolting a dozen AI agents onto the business at once, one for email, one for scheduling, one for research, one for sales, all switched on the same week. No sane company would hire twenty people on the same Monday with no onboarding, no reporting lines and no idea who’s supposed to tell whom what. Even Danny Ocean spends half of Ocean’s Eleven handing out roles before anyone goes near the vault. Each agent needs the same treatment: a defined job, a clear scope, someone it answers to, and a way to share what it knows with the others. The approach that works is the boring one. Bring on one, get it working, then add the next.

The same goes for building software with AI, and this is the advice I give most often to people who want to make things. Don’t rush to learn Python. Learn to run a software project. AI coding agents can write the code. What they need from you is what any development team needs: a roadmap, milestones, someone reviewing the work, tests that prove it runs, documentation so the next person or the next session can pick it up, and someone saying no when the scope starts creeping. People assume AI lets them skip that discipline. It needs more of it, because an AI team works fast enough to build the wrong product in an afternoon.

Feedback works too, which surprises people. You don’t fire an intern over their first mistake. You tell them what was off, and they get it right the next time. AI responds the same way inside a conversation. “That’s close, but too formal, and you left out the pricing table” gets you a much better second draft than starting over in a new window. Checking works the way it does with people, too. The lawyers in that fake-citation case would never have filed a junior associate’s brief without reading it. They filed the AI’s. And AI shares one habit with the overconfident new hire: it rarely says “I don’t know.” So you learn which tasks it nails and which ones need a second look, the same way you’d learn that about a person.

I want to be clear about where the comparison breaks, because it does. AI isn’t a person. It doesn’t get tired the way we do or have a bad day. It doesn’t carry today’s lesson into tomorrow unless someone writes it down. It can produce forty drafts in the time a person makes one, and it won’t tell you it’s out of its depth unless you ask. So the claim here is narrower than the title. We’ve spent centuries working out how to get good work out of capable, fallible people, through clear briefs, handoffs, review, feedback and careful hiring, and that playbook turns out to be the best one we have for AI too.

Most of those famous bloopers got fixed, by the way. Ask a current chatbot how many r’s are in strawberry and you’ll almost always get three. That’s what happens with new hires too: the mistakes from someone’s first month are rarely the ones you see in their second year. The person sharing the strawberry screenshot and the person getting great work out of the same chatbot are using the same software. One of them is waiting for a vending machine to work. The other is running a team. The most valuable AI skill right now turns out to be the one good managers have been practicing all along.

The New Typing continues. The series lives at drjeffwurfel.com/writing.