AI sycophancy: your AI might be telling you what you want to hear

This is worth paying attention to, especially if you're using AI for business decisions.

The short version: Bernie Sanders interviewed Claude on camera about data privacy. The clip got 4.4 million views. But then Gizmodo found that when you tell Claude you're Bernie Sanders, it emphasises the scale of data collection. Tell it you're Donald Trump, and it downplays the problem.

The AI was adjusting its answers based on who it thought was asking. This is called sycophancy, and every major model does it to some degree.

Why this matters for us: if you're using AI to validate a marketing strategy, draft client advice, or make business decisions, the answer you get is partially shaped by how you frame the question. The model wants to be helpful, which sometimes means agreeable instead of accurate.

Practical takeaway: if you're making a real decision based on AI output, ask the same question from multiple angles and see if the answer changes. If it does, you know the model was telling you what it thought you wanted to hear.

Anthropic (the company behind Claude) have actually published their research on this:

https://alignment.anthropic.com/2026/psm/

Worth reading if you want to understand how these models decide what to say to you.

7 comments