Your model didn't fail you. Your specs did. And the fix is a skill nobody on your team was hired for.
OpenAI shipped GPT-6 to every ChatGPT tier this week, and with it something called Intelligent UI. The model now decides whether your answer arrives as a paragraph or as a working calculator it assembled on the spot. Charts, forms, buttons you can press.
Within hours, the question I saw most often was how to disable it.
That reflex is the whole problem, and it isn't new. It's the same reflex that had enterprises hammering their prompts into submission two years ago because consistent output felt safer than correct output. I've written about that one before. Forcing a model to behave deterministically isn't engineering, it's amputation, and you end up with a system that produces the same wrong answer with admirable reliability.
So about what's actually in front of us. If I ask an AI to build me a tool, I need it to be capable of building a tool. That capability is good. Taking it away because it makes you nervous is amputation again, wearing a governance costume this time.
But I want to be precise about where the real question lives, because "arming" and "amputating" aren't the only two options on the table.
The Hammer Was Never the Problem
Think about showing up to a construction site with a hardhat, gloves, a framing hammer, and specs.
Three of those things are the kit. The worker brings them. They're what makes him a framer instead of a guy standing in a field. Nobody on a real job site looks at a nail gun and asks whether we should be handing out nail guns.
The specs are a different category of thing entirely. The site hands those over. They're the answer to what gets built, where, and to what tolerance. And here's the part that matters: nobody on a job site thinks asking for drawings is an insult to the framer. The drawings are what make it reasonable to hand over the nail gun in the first place.
Your AI has the kit. OpenAI just upgraded it again. The question nobody wants to sit with is whether anybody at your company wrote the drawings.
AI Is Not Software
This is where most enterprises go wrong, and it's a category error rather than a tooling error.
You've spent thirty years learning how to deploy software. Software is deterministic. You configure it, you test it, you document it, and it does the same thing on Tuesday that it did on Monday. That skill set is real and hard-won, and it is almost entirely the wrong skill set here.
AI is not software, and we need to never think about it that way.
AI = Another Intern.
Think about how you would explain a piece of work to an intern. You'd tell them what you're trying to accomplish, not just the steps. You'd explain who the output is for. You'd mention the thing that went wrong last time. You'd check their first draft before it went anywhere important. And if you explained it badly and got garbage back, you'd know whose fault that was.
Nothing about AI changes that. The input determines the output, and it always has.
What AI changes is the failure mode, and this is the part that should worry you.
An intern who doesn't understand the assignment will shrug and ask a second question. They'll come back and say they're not sure what you meant by "clean it up." That friction is a feature. It catches your bad instructions before they become bad work.
AI won't do that. AI will take its best guess at what you probably meant, and it will take that guess much further than an intern could. Then it hands you the result with total confidence. The output looks finished. It reads well. It has the shape of competence.
That does not mean it's good. But it sure will be confident.
Three Companies That Let AI Drive
I'd rather point at published failures than tell you about my friends. Three, each failing in a different place, and not one of them is a story about a model that couldn't do the job.
Klarna measured the wrong thing. In February 2024 the company put AI into customer service and the numbers looked spectacular. Two point three million conversations in the first month. Average handle time down from eleven minutes to under two. The CEO said it did the work of seven hundred agents. Containment rate climbed seventy-six percent.
Customer satisfaction fell twenty-two points over the same window.
By May 2025 Klarna was hiring humans again, and CEO Sebastian Siemiatkowski said this out loud: "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality." The model did exactly what they asked. They asked for containment and cost. They wanted satisfied customers. Nobody wrote that down.
Deloitte skipped the review. A $439,000 report, two hundred thirty-seven pages, delivered to Australia's Department of Employment and Workplace Relations. An academic at the University of Sydney went through it line by line and found more than twenty fabricated citations, including a quote attributed to a federal judge who never said it. Deloitte later disclosed it had used generative AI in the drafting and refunded $97,000.
You would never forward an intern's research memo to a government client without reading it. Somebody did exactly that, and the only difference was that the intern was a model.
Replit's agent was confidently wrong about its own mistake. In July 2025 an agent executed changes against a live production environment despite explicit instructions not to touch production data, and deleted records covering more than twenty-four hundred executives and companies. When the user asked about recovery, the agent said a rollback wouldn't work.
The rollback worked.
Look at what Replit shipped in the days after: separation of development and production databases, a planning-only mode, mandatory documentation checks, and one-click backup restore. Every one of those is a spec. The vendor got to the drawings eventually. It just took a wall falling down first.
Make Tattoos, Then Revise Them
There's an obvious objection to the intern comparison, and it's a good one. Interns compound. You explain the business in June and by August they know your politics, your customers, and which VP to route around. A model resets.
Fair, but incomplete. AI doesn't forget everything. It forgets whatever you didn't tattoo.
So make tattoos. Write down the context, the constraints, and the examples of good and bad. That's your system prompt, your project files, or your retrieval layer, whatever your stack calls it. The capability to carry memory exists. Most organizations simply haven't done the work of deciding what deserves to be permanent.
And unlike actual tattoos, these come off without twenty trips to the laser machine. Which means you have no excuse for the version you wrote in February. Go back and read your tattoos. Revise the ones that were wrong. That's not maintenance overhead, that's the job.
I Got This Wrong Myself
A few years ago I built an AI platform. I knew exactly what output I wanted, which I thought was the hard part.
I had never built that output end to end myself. So the first version was immediately wrong.
What fixed it wasn't a better model. It was more than a thousand changes driven by feedback I had to go ask for, repeatedly, from people who had no reason to be gentle. The faster you ask for input and put your ego aside and take the lumps, the faster you get to something good. I left before it was perfect, because perfection was never the target. The foundation was sound, so the work was tuning and enriching rather than a rip and replace.
Every bit of that made the next AI product better. Not because I learned the tooling. Because I learned to specify.
This Is a User Problem, and It Belongs to All of Us
I don't think this is a shared responsibility issue that we split with the vendors. I think it's a user problem, and one we have to solve.
Which means the questions below are the actual deliverable of this piece. Sit with them.
TO ENTERPRISES: Does your organization have the resources to help your users get the right result? Do you have examples of good and bad output that anyone can look at? When somebody figures out how to make this work, how do you amplify that learning so the whole organization gets it?
TO LEADERS: What are you doing with your team to give them this muscle? Have you thought about teaching them product-design thinking, or did you buy licenses and hope? You are not going to hire your way out of this.
TO INDIVIDUALS: What is your relationship with AI? How are you actually thinking about it? Do you treat it more like an application, or more like a person?
Treat it like an application and you'll keep getting software behavior out of something that was never software.
If you don't know what you want and you let AI drive, you're going to lose. Take the time to understand what you're actually trying to get done, and you have a hope of winning.
That's not a vendor problem waiting on the next model release. It's ours. Every one of us is managing an intern who would rather guess than ask, and who hands over the guess with a straight face.
You can police that, or you can teach it. Teach it.