OpenAI's CEO Says the World Should Accept Some Bad Things From AI. Decide What You Won't.

Sam Altman, the CEO of OpenAI, told Politico on Sunday that "the world should accept some bad things happening for the benefits of this technology and people having the agency." He also said there's broad daylight between OpenAI and Anthropic on how much to worry, and he pushed for lighter regulation. I'm not here to argue with him. I use ChatGPT every day. But I read that quote the way a business owner should read it, which is that the company selling you the tool just told you it expects some of the damage to happen.

The "world" in that sentence includes the guy running a three-truck landscaping company in Maumee who let AI write his estimate and never looked at the number.

The vendor isn't going to catch it for you

Nothing in that interview is a scandal. A company can believe the benefits are worth the risk, and plenty of reasonable people agree. But it does settle something I get asked in workshops all the time, which is whether the AI companies are going to make these tools safe enough that I don't have to think about it. Pretty clearly no. They're building the engine. What it does on your customer's invoice is on you.

Picture a dental office where an AI tool drafts the appointment reminders. One day it starts telling people their cleaning is covered by insurance when it isn't, and nobody's reading the reminders before they go out. That's the kind of bad thing Altman is talking about, and it doesn't make the news. It just costs you a few patients and an awkward week.

Write down the bad things you won't accept

Altman's version of this is a philosophy for a trillion dollar company. Yours can fit on an index card. Sit down for twenty minutes and write the handful of bad outcomes that would actually hurt your business: a wrong price going to a customer, a client's private info getting pasted into a free tool, a made up fact in a proposal, an email sent in your name that you never read. Then next to each one, write who looks at the work before it goes out. A name, not "the team."

That's the whole exercise. If a bad outcome has nobody's name next to it, that's the spot where you're accepting the bad thing by default, same as the CEO is asking everybody to do, except you didn't get any say in it.

Pick the checks that match the stakes

You don't need a person reviewing every AI draft of a social post. You do want a human on anything with a dollar figure, a legal promise, or a customer's personal information in it. I tell people to sort their AI use into two piles. Pile one is stuff where a mistake is embarrassing and fixable, like a rough blog outline. Pile two is stuff where a mistake costs money or trust. Pile one can run on its own. Pile two gets a set of eyes every single time, and that rule shouldn't bend just because everybody's busy on a Friday.

It also helps to pick your tools with this in mind. Some vendors let you turn on approvals before an agent sends or buys anything. Some don't. Given what Altman said Sunday, I'd lean toward the ones that do, and I'd turn those approvals on before I turned on anything else.

The good news is the CEO is right about one thing, which is that the benefits are real. Plenty of small offices I work with get hours back every week. You just get to keep those hours if you decide ahead of time what you're not willing to lose for it.