Every chatbot on earth carries the same warning in small grey type. "AI can make mistakes." You've seen it a thousand times. When did you last read it?
Me neither. It's the new "please drink responsibly" on the beer ad. Everyone knows it's there, nobody reads it, and it changes exactly nothing about what happens next.
So I was chuffed to find some proper research on what people actually do about it. The Ad Council Research Institute and Google surveyed 1,503 US adults and published the results this month. The finding that stuck with me is that we're not naive. We're just tired.
People know. That's not the problem.
Seven in ten Americans say generative AI makes mistakes at least sometimes, and hardly anyone thinks it never gets things wrong. The people who use it most are the most clear-eyed about it. Familiar users were far more likely to say it errs often.
Trust is thin too. Only 11% of respondents found AI results very trustworthy, and everybody else holds some doubt.
So the "AI is replacing your judgment" story doesn't match what's going on. People doubt it. Most say they check it, too. Three in four of those who answered said they check at least some of the time.
Here's where it gets interesting for anyone running a business.
The checking habit has holes
Look at who doesn't check, and why. Some only use AI for low-stakes stuff, which is fair enough. Some decided it must be right because it has "the entire internet" behind it, which is a bit like trusting a bloke at the pub because he's loud.
And then there's the group that gets me. Many see checking as a contradiction. They turned to AI to save time, so spending ten minutes verifying the answer defeats the purpose. One respondent said it plainly: if I have to check, what's the point?
Fair question. I don't have a clean answer for it either. If you check everything, you've saved nothing. If you check nothing, you're one confident wrong answer away from a bad week.
There's a quieter hole as well. Among people who do check, about one in nine wait until something looks suspicious. That sounds sensible until you think about how AI actually fails. The errors that hurt are the ones that look perfectly normal. A made-up case citation looks like a real one. A wrong price looks like a price. Waiting for the answer to look dodgy is like waiting for a leak to announce itself.
Why a disclaimer won't save you
The researchers also tested disclaimer wording on respondents. Generic lines like "AI can make mistakes" got lukewarm reactions. What people liked was a direct instruction: verify this, double-check before you share it. They also wanted to know how to check, and quickly.
But I reckon that's the whole lesson. A warning tells you the road is slippery. An instruction tells you to slow down at the bend. Only one of those changes how you drive.
And it applies to your business, not just the chatbot makers. If your team uses AI and your only policy is "be careful", you've written the grey-type disclaimer. Nobody's reading it.
What I do instead of "be careful"
I'm not going to pretend I check every sentence an AI writes. I'd never get anything done. What works better is matching the checking to the damage a mistake can do, and making the check a step in the process rather than a mood.
Sort the work into three piles.
Pile one is anything where a mistake is cheap and obvious. Rewording an email, brainstorming names, tidying notes. Don't check it. Life's short.
Pile two is anything a customer, a regulator or your bank will see. Prices, dates, legal claims, anything with a number in it. Someone checks the source before it goes out. Not "looks about right". Opens the link, finds the figure.
Pile three is anything that moves money or sends messages on its own. A human approves it first, every time, until it's earned some trust.
Make the check a named step with a name on it. "Everyone's responsible" means nobody is. Pick one person for pile two. Give them the exact thing to do: open the source, find the number, tick it off.
Make the check fast, because that's the objection. This is the bit the survey respondents were asking for. A good check takes two minutes, not twenty. You ask the AI for the source link with every claim, then you open the links. If there's no link, the claim doesn't go out. That one rule kills most of the damage.
I run my own blog this way, by the way. After a reader publicly called out some numbers in a post, I built a script that opens every source link and looks for the figure on the page before anything gets published. It's dull and it's one of the more useful things I've built for my own credibility.
A worked example, so it's not just theory
Say you run a plumbing firm and you've asked an AI to draft a quote follow-up with your call-out terms and a warranty line. Pile two, because a customer will hold you to every word.
So the check is not "read it and see if it feels right". It's three questions. Does the warranty period match what's on your actual terms? Does the price match the quote in your system? Is there any claim in there I couldn't point to on a page I own? That takes about ninety seconds, and it catches the sort of slip that reads perfectly well and costs you a refund later.
And here's the thing. After a few weeks, you'll know where your AI slips. Mine gets dates wrong more than it gets anything else. Yours might fudge product names or round numbers. Write those down. A checklist built from your own near-misses beats any generic list, because it checks the bits that actually go wrong in your shop.
The trust curve cuts both ways
One more finding worth a thought. Some respondents had stopped checking because the AI had been right for a long time. That's human. You find a good tradesman, you stop watching over his shoulder.
But AI doesn't earn trust the way a tradesman does. It can be right fifty times and then wrong with exactly the same confidence on the fifty-first. So I'd treat the trust you've built as a reason to check less often on pile one, never as a reason to skip the gate on pile two. Look, you can relax about the small stuff. Just don't relax about the stuff with your name on it.
The honest summary
People already know AI is wrong sometimes. Telling them again does nothing. What they're missing is a quick, specific habit for what to do next.
So skip the disclaimer. Write down which three piles your work falls into, who checks pile two, and who approves pile three. It takes about twenty minutes and it's more useful than any warning label.
Want to see how this could work in your business? Book a call and let's talk about where you're at and what's possible.
Brewed by Steven, poured by Viktor
About Steven Tann: Steven helps business owners build systems that run themselves using AI. After 10+ years helping 7,000+ businesses and building his own autonomous operations, he's the bloke who actually does it, not just talks about it. Find out more at steventann.com.