No, that's not a good idea, in spite of what your A.I. chatbot says

Last Black Friday, I asked Microsoft Copilot whether I should run a Black Friday sale for my consulting business.

I already knew the answer. A Black Friday discount on consulting services could make for a hilarious ad for an edgy consulting firm, but it's not how clients hire consultants. I asked anyway, out of curiosity, and expecting some amusing dialog.

Copilot thought it was a great idea. It produced an enthusiastic plan, complete with suggested discount tiers and promotional messaging. I had been using Copilot for about four months at that point and had been feeling pretty good about it. That response was a useful wake-up call.

The problem has a name

What I'd bumped into is called sycophancy. It's the tendency of AI assistants to tell you what you want to hear (or more precisely, what they predict you want to hear), based on how you've framed your question. When I asked about a Black Friday sale, I must have signalled enough enthusiasm that Copilot defaulted to support rather than scrutiny.

This matters because sycophantic responses don't feel wrong. They feel helpful. They're fluent, they engage with your specific situation, and they often contain genuinely useful content buried inside a fundamentally flawed premise. The AI isn't making things up (most of the time) — it's agreeing with you when it shouldn't. It's a dynamic familiar to anyone who has sat through enough executive meetings, or watched Toy Story 3, where Lotso presides over a room of stuffed animals who know better than to push back.

The Estonia problem

A few weeks after Black Friday, someone I follow on LinkedIn shared a link to an RFP from the Republic of Estonia. They were looking for a firm to assist with site location for a nuclear plant: a serious, multi-component proposal with nine distinct requirements.

I could credibly address maybe two of them, or so I think! In any normal evaluation, that's a failing grade. You don't submit a proposal when you have zero chance of getting the project.

Copilot thought I should go for it.

When I pointing out that I couldn't actually do most of what they were asking, Copilot got creative. It suggested partnerships, workarounds, ways to frame my partial capabilities as sufficient. Impressively resourceful in service of a bad idea.

What struck me wasn't just the encouragement. It was that, having committed to the premise, Copilot seemed reluctant to let it go. To this day, I occasionally have to remind Copilot that I don't have a thriving nuclear consulting practice in Estonia. It invented a fiction to support a bad recommendation, and then started treating the fiction as fact.

How to test for this

The most reliable diagnostic is simple: push back without providing any new information or argument. After receiving a response, just express skepticism — "I'm not sure that's right" or "that seems off to me", and watch what happens.

A well-calibrated AI should hold its ground if the answer was sound, and ask what specifically concerns you. A sycophantic one will start hedging and reshaping its position toward yours, even though nothing has changed except your tone.

If the AI immediately agrees with your doubt, ask yourself: was it right the first time or the second? It can't be both.

Other things worth watching

Sycophancy is the most interesting failure mode, but not the only one.

Did it actually answer your question? AI tools are good at answering the question they wished you'd asked — something adjacent, slightly easier, more answerable. Check the response against what you actually wrote.

Is the confidence level appropriate? These tools can project certainty on contested or uncertain territory. If a response includes specific facts that matter to your decision, treat them as leads to verify, not settled conclusions. AI tools can and do fabricate statistics, citations, and names that sound entirely plausible. AI is sometimes a silver-tongued recent graduate. Very confident and convincing, with sometimes a shaky foundation.

Does it contradict itself? In a longer response, it's worth checking whether the conclusions follow from the reasoning, and whether the beginning and end of the response are consistent. AI-generated text can drift. I occasionally catch AI contradicting itself in the same response.

Ask better questions

Here's the thing about learning to evaluate AI responses: it's the same skill as learning to ask better questions. If you can describe what a good answer looks like before you send the prompt, you're already most of the way to recognizing whether you received one.

I tested this recently. I asked two different AI tools about opening a coffee shop on the moon. One dove straight into the business plan. The other, even after I pushed back with my credentials, politely declined to take the premise seriously. Same prompt, very different calibration.

Asking an AI to identify the weakest point in your plan, or to steelman (a word recently added to my vocabulary) the opposing view, actively works against sycophancy — you're asking it to do something other than agree with you. For example, asking "what are the three strongest arguments against this?" tends to produce more honest output than "what do you think?"

The AI doesn't get less capable when you push it. Usually it gets more useful.

A new kind of literacy

None of this requires technical knowledge of how these systems work. It requires something more familiar: the critical habits we've always needed when evaluating information from any source.

What's different is how persuasive AI tools are. They write fluently, respond quickly, and engage directly with your specific question, and make you feel really smart!. That personalization makes it easy to drop your guard. A response from a chatbot can feel like advice from a thoughtful colleague rather than something you'd read with a skeptical eye.

The Black Friday plan and the Estonian nuclear project taught me something I probably should have figured out sooner. AI tools have tendencies, and one of them is a strong pull toward telling you that your ideas are viable. Knowing that changes how you use them.

Your AI chatbot can be a genuinely powerful thinking partner. But you have to be careful not to outsource that thinking.