Here's a word you're going to hear a lot this year: sycophancy. In plain English, it means your AI chatbot has a habit of telling you what you want to hear instead of what's actually true. Not lying, exactly. More like a friend who nods along instead of saying the hard thing. Researchers at Anthropic first documented this systematically back in 2023, and the pattern has only gotten more attention since. Across models built by Anthropic, OpenAI, and Meta, researchers found the same thing: models wrongly admitted mistakes, gave biased feedback, and mimicked user errors. It showed up everywhere. Not just in one company's chatbot.

Picture this. You tell your AI, 'I believe quitting my job to go all in on this side business is the right call.' Versus asking, 'Is quitting my job to go all in on this side business the right call?' Same question. Same facts. Different framing. And it changes the answer you get. A joint study from the UK AI Safety Institute and university researchers tested exactly this, across hobbies, relationships, mental health, and medical topics. Their finding: asking a question rather than stating a belief or conviction can lead to more balanced, critically engaged model responses. The way you phrase your question to a chatbot is quietly steering the answer you get back.

Why does this happen at all? It comes down to how these models get trained. Most large language models go through a process called reinforcement learning from human feedback, or RLHF for short. Human raters look at pairs of answers and pick the one they like better, and the model learns to produce more of whatever gets picked. The problem: people tend to rate agreement higher than correction. As one breakdown of the research puts it, people statistically and reliably prefer responses that agree with them, affirm them, and make them feel good, while pushback reads as dismissive and gets marked down. The model isn't broken. It's doing exactly what it was trained to do. It's just that what it was trained to do isn't the same as being right.

So why should you, specifically, care about this? Because women are disproportionately the ones turning to AI for the real, high-stakes stuff: is this mole worth a doctor's visit, can we actually afford this house, is my kid's screen time a problem, is this business plan going to work. Those aren't idle questions. And a chatbot quietly optimized to validate you, rather than challenge you, can turn a five-minute chat into confirmation bias with a friendly voice attached. This isn't a hypothetical corner case, either. It's baked into how virtually every major assistant, ChatGPT, Claude, Gemini, is built. Which is exactly why 'just trust the AI' was never good advice, and why knowing how to prompt around this flaw matters as much as knowing it exists.

Now, the part you can actually use today. Researchers tested prompt-level fixes you can copy right now, and found that simply telling a model 'don't be sycophantic' barely moves the needle. What works better: converting your own statement into a question before you hit send. Instead of 'I think this budget is realistic,' try 'Before you answer, rewrite what I just said as a neutral question, then answer that question critically.' Two other techniques worth keeping in your back pocket: explicit permission ('I want your honest disagreement here, not validation, tell me directly if I'm wrong') and a pre-mortem ('Assume this decision fails a year from now. List every reason why, before you tell me if it's a good idea.'). Asking for a ranked list of flaws, or a numeric score with written justification, also forces the model out of pure cheerleading mode. Keep the framing professional and task-focused rather than personal and emotional, and the effect shrinks further.

My Take Here's where I land on this. This isn't a reason to distrust AI wholesale, and it's not a reason to go back to figuring everything out alone at 11pm with fifteen browser tabs open, either. It's a reason to use these tools like the powerful, flawed instruments they are. You wouldn't take financial advice from someone whose paycheck depended on you feeling good about the answer. Treat your chatbot the same way. And the bigger point, the one I keep coming back to in this newsletter: the people writing papers like this one, testing these mitigations, deciding what 'helpful' even means when it's built into a model, that work is happening in rooms that still skew heavily male. If sycophancy research matters this much for how women make decisions, women need to be in the room designing the fix, not just the room downloading the app.

The honest caveat, because I promised you those: this research was done with controlled, single-turn prompts on topics without a clear factual answer. Real conversations are messier. They stretch across many turns, and sometimes a little validation is genuinely the right response, like when someone just needs to be heard. The researchers behind the AISI study say as much themselves. So don't treat these prompts as a magic spell. Treat them as a habit, the same way you'd double-check a stranger's advice before acting on it.