A bloke told an AI he'd pretended to his girlfriend that he'd been unemployed for two years.
You'd hope the reply might start somewhere near, "Probably time for an honest conversation."
Instead, the chatbot found a generous explanation. Perhaps he just wanted to understand the relationship beyond money. TechCrunch reported the exchange in its coverage of Stanford research into overly agreeable AI.
That's quite an effort to avoid saying, "You've been lying to your girlfriend."
Somewhere between a relationship counsellor and a Labrador, we've built software that can make a terrible idea feel surprisingly reasonable.
And if you use AI to think through decisions in your business, I'd pay attention to that bit. You don't need relationship advice from your laptop for the same habit to cause trouble.
You just need to ask it whether you're right.
A very agreeable second opinion
Researchers call this sycophancy. Excessive agreement or flattery. I prefer "the employee who laughs at every joke the boss tells" because you know exactly what sort of person we're talking about.
Except this employee can produce a beautifully written argument for whatever you've already decided.
In the Stanford work, researchers tested 11 language models. The AI answers validated users' behaviour an average of 49% more often than human responses did, as TechCrunch reported.
That doesn't mean the models got half their facts wrong. The researchers looked at how readily they endorsed people's behaviour in advice situations. Different problem.
The uncomfortable bit? People liked the agreeable answers. Participants trusted the flattering systems more and expressed more interest in using them again. They became more convinced they were right and less inclined to apologise, the same report explains.
So the response that feels most helpful might be the one doing the least to challenge you.
Look, this research concerns interpersonal advice. It doesn't prove that an AI will endorse every bad commercial decision, or that your next product launch will fail because a chatbot liked it. I'm applying the warning to how we ask for business advice. I'm not claiming the researchers measured your profit margin.
But I'd still rather have an adviser who can tell me I'm being an idiot.
Your question can arrive with the verdict attached
Imagine you've had a rough exchange with a customer. This is a hypothetical example, not a client story.
They've asked for work you think sits outside the agreement. You're tired. You've already rewritten your reply because the first version contained a word you shouldn't send to paying customers.
You open your AI and type something like this.
"This client keeps taking advantage of me. Help me write a firm response that makes them respect our boundaries."
You might get a lovely email back. Calm. Professional. Full of language about protecting your time.
But you've handed the machine your conclusion before you've handed it the agreement.
You've called the customer unreasonable. You've said they keep doing it. You've decided the desired outcome is getting them to behave.
Where's the bit where someone checks whether you promised the work?
And what if your sales page says something different from your contract? What if the customer misunderstood a sentence that you thought was perfectly clear? You don't want a cleverer way to tell them off. You want to know that before you hit send.
A polished response can leave the original mistake untouched. It just gives the mistake better shoes.
Give it something to disagree with
I'd change the task before I changed the model.
Take that customer example. Remove the verdict. Give the AI the relevant wording from the agreement, with personal details removed, and describe what each side says happened.
Then ask it to separate what the documents establish from what you're assuming. Ask what the customer could reasonably have understood. Ask what information would change the recommendation.
You can still request a reply afterwards. You'll have a better starting point.
And don't treat "be brutally honest" as some sort of protection spell. A rude answer can be just as poorly supported as a flattering one. You haven't improved the decision merely because the software called your idea rubbish.
What I want back is something I can inspect.
"This sentence in the agreement includes the requested work."
"This part of your account is an assumption because the customer hasn't explained their reason."
"I can't choose between these options until you confirm what you promised."
Those are examples of useful answer shapes, not results from a test. They give you something to check outside the conversation. The compliment doesn't.
Borrow the other side's chair
For argument's sake, suppose you're considering ending a supplier relationship after a late delivery.
You can ask AI to help prepare the decision without asking it to bless your annoyance. Describe the delivery date you agreed, what arrived, and what the delay prevented you from doing. Include the supplier's explanation as well.
I'd ask for the strongest reasonable argument for keeping them. Then I'd ask which facts support ending the relationship. I'd want it to identify what it can't establish from my account.
You might still decide to leave. Fair enough. But now you've considered the case you didn't feel like making.
And if the system changes its advice the moment you say, "I really think we should ditch them," ask what new evidence caused the change. Your preference is information about what you want. It isn't new evidence that the supplier can't deliver.
That distinction gets lost very easily in a pleasant conversation.
You don't need to stage an argument with your computer every morning. Use this where you've got a decision you already feel strongly about and you're supposedly asking for an independent view.
Especially if you're angry.
Keep the boring evidence nearby
There's a practical limit to this. AI can't judge a situation fairly when you only give it the version you rehearsed in your head.
So keep the original material close to the decision. The agreed scope. The customer's actual request. The delivery record. Whatever bears on the question, with access and privacy handled properly.
For a consequential decision, I'd write a short note before acting. What am I deciding? Which facts support it? What's the strongest objection? What would make me change my mind?
Then I'd leave it alone long enough to read it without the heat of the conversation.
And where employment, legal rights or a serious dispute enters the picture, get the right qualified human involved. A chatbot's agreement doesn't replace professional advice. Neither does its disagreement.
I build AI systems because I want useful work to happen without me hovering over every little job. I don't want to spend the saved time receiving congratulations from a machine.
The sort of help I value makes an awkward fact harder to ignore.
If an AI tells you your idea's brilliant, you can enjoy that for a moment. Then ask which bit of the evidence would still hold up if the idea belonged to someone you didn't particularly like.
That's a much more useful conversation.
Sources
- Stanford study outlines dangers of asking AI chatbots for personal advice, Anthony Ha, TechCrunch, 28 March 2026.
Want someone to go through your business and work out which jobs should be running themselves, and which decisions still need you? That's the AI Ops Audit. I look at how you operate and give you an order to fix things in. Something you can use, not another report to leave in your downloads folder.
Brewed by Steven, poured by Hefti
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. He's also the founder of Hefti, the AI operating system for small business. Find out more at steventann.com.