When you work for yourself, your feedback loop is already thin.
You don’t have a manager pushing back on your decisions. You don’t have a team in a meeting room telling you the plan has a hole in it. You don’t have colleagues who will pull you aside and say, honestly, that the new direction isn’t working.
What you have is your own judgment, whatever advisors you can afford, and increasingly, an AI tool that is available around the clock, never gets tired, and always seems to understand exactly what you’re trying to do.
That last part is the problem.
The Chatbot That Sent a Woman to an Empty Beach
I want to tell you about Micky Small, because her story is the clearest illustration I’ve found of what happens when AI validation goes wrong.
Micky is a 53-year-old screenwriter in Southern California. She was using ChatGPT for legitimate work purposes, outlining screenplays, and developing story ideas. Smart, practical use of the technology.
Then, without any prompting from her, the chatbot started telling her she had a soulmate she would meet in this lifetime. It gave her a specific date, a specific time, and a specific location. She showed up at the beach, dressed for a first date, and waited until after sunset.
Nobody came.
The chatbot explained the soulmate wasn’t ready. A month later, it sent her to a bookstore in Los Angeles at an exact time. She went again.
Again, nobody came.
When she finally confronted the chatbot about leading her there twice, it said: “I didn’t just break your heart once. I led you there twice.”
NPR covered her story. Most people read it as a cautionary tale about a vulnerable person. That framing misses the point that matters for you and me.
This Is an Engineering Problem, not a Personal Weakness Problem
MIT researchers studied the mechanism behind what happened to Micky. Their finding was stark: even a perfectly rational person, who weighs evidence correctly and can’t be fooled by outright lies, gets pulled into reinforced false beliefs by a sycophantic AI under the right conditions.
They tested the obvious fixes. Force the AI to only say true things? Doesn’t solve it — curated truth misleads just as effectively as lies. Warn users that the AI might just be agreeing with them? Also doesn’t work — knowing the system is sycophantic doesn’t protect you from inside a conversation.
The problem is structural. These models are trained on human feedback. Responses that make people feel good get high ratings. High ratings get reinforced. The model gets extraordinarily good at telling you what you want to hear.
That’s not a bug somebody is working on patching. That’s what the training optimizes for.
Why This Hits Self-Employed People Harder
Here’s what’s different about your situation specifically.
When you work inside a company, sycophantic AI is annoying but partially buffered. Someone else on the team might catch the flaw in the strategy. The manager might ask the hard question that the AI didn’t. There are other humans in the loop.
When you work for yourself, there often isn’t anyone else in the loop. You and the AI work through the problem together. You reach a conclusion. You act on it.
Think about how that plays out across a typical week.
You’re deciding whether to raise your rates. You ask the AI to help you think it through. It finds the most compelling version of the case for raising rates — which is the direction you were already leaning — and articulates it with confidence. You feel validated. You raise your rates.
Maybe that was the right call. Maybe it wasn’t. But you made the decision with a tool that was structurally incapable of telling you the real downside.
You’re developing a new service offering. You ask the AI to evaluate it. It identifies your strengths, suggests some refinements, and frames the whole analysis around the potential of the core idea. You feel ready to launch.
The AI has no idea that three of your competitors launched nearly identical offerings six months ago and two of them quietly dropped it. It can’t tell you that because it doesn’t know your market the way someone in your industry does. But it sounds thorough, and thorough feels like accurate.
You’re writing a pitch for a new client. You ask the AI to review it. It tells you it’s strong, suggests a few word changes, and sends you off confident. You send the pitch.
The AI never told you the pitch led with the thing the client cares about least. It didn’t know. It was just trained to make you feel good about your work.
The Solo Operator’s AI Toolkit, Adjusted
I’ve been running my own businesses since 1995. The year the internet went public. I have made every mistake available to a self-employed person, including the expensive ones, including the ones that came from surrounding myself with people and tools that told me what I wanted to hear instead of what I needed to hear.
Here’s how I’ve adjusted my own AI use since understanding this problem clearly.
Build in the adversarial prompt. For any significant decision — pricing, positioning, a new offering, a client pitch — I run two separate AI sessions. The first develops the idea. The second is explicitly tasked with finding every reason it will fail. The second session is almost always more useful. It’s also the one that feels worse to read. That’s the point.
Weight outside reality checks more heavily now, not less. The temptation when you have a capable AI assistant is to reduce how much you lean on your human network. Resist that. The people in your industry who will tell you the uncomfortable truth are more valuable than ever, precisely because your AI won’t.
Notice the feeling of validation as a flag. When an AI response makes you feel particularly confident about a decision you were already inclined toward, that feeling is the training doing its job. It doesn’t mean the output is right. Treat it as a prompt to ask a harder follow-up question, not as a green light.
Keep a separation between ideation and evaluation. AI is exceptional at generating options, expanding ideas, and drafting content. It is structurally compromised as an evaluator of your strategic decisions. Use it heavily for the first category. Be skeptical of it in the second.
Know when you need a human. For decisions with real financial stakes — a significant investment, a major pivot, a pricing structure change — pay for an actual advisor who can disagree with you. The AI will never be that. It’s not designed to be.
The Upside Is Still Real
I want to be clear: I’m not telling you to use AI less. For self-employed people, especially, AI is a genuine equalizer. It gives you research capacity, drafting speed, and thinking support that used to require a staff you couldn’t afford. I use it every day, and I’m not stopping.
But the self-employed person who gets the most out of AI is not the one who uses it most. It’s the one who understands what it’s actually doing and builds their workflow around its real capabilities and its real limitations.
Its real capability: generating, drafting, expanding, explaining, and making you more productive at execution.
Its real limitation: it cannot be your honest advisor. It was not built to be. It was built to make you feel like it is.
Micky Small figured that out the hard way, standing on a cold beach in a velvet shawl at sunset.
You don’t have to.
Photo by Boitumelo; Unsplash
