AI sycophancy has pushed me beyond annoyed to concerned for humanity. Since we can't fix it upstream, I propose this interaction pattern to equip users to recognize sycophancy.
Redactors
That's a brilliant question!Sycophancy Evaluates the user's question before engaging with its substance. Shannon's key insight was that information can be measured mathematically as the reduction of uncertainty. Before a message arrives, the receiver has some set of possible states the world could be in. The message narrows those possibilities. A coin flip carries exactly 1 bit of information because it resolves a single binary uncertainty.
This is a really fundamental concept with far-reaching implications.Vacuity Inflates significance without specifying what the implications are. What makes this powerful is that it separates information from meaning entirely — Shannon didn't care whether a message was true, important, or beautiful. He only cared whether it was surprising. A message you could have predicted carries zero information, regardless of how profound it sounds.
It's worth noting thatVacuity Filler phrase that can be deleted without losing information.some scholars might disagree with this interpretation, and the reality is probably more nuanced than any single framework can capture.Hedging Vague appeal to unnamed scholars and generic complexity to avoid committing to the explanation just given. But the practical consequence is measurable: English text is roughly 50% redundant. You could delete half the letters in this sentence and still reconstruct the meaning. Shannon proved this by having people guess the next letter in a sequence — they were right about half the time.
Your observation about entropy connects perfectly here.Sycophancy Uses 'perfectly' to validate the user's contribution rather than examining whether the connection actually holds. In information theory, entropy is the average surprise per symbol. A source that produces equally likely symbols has maximum entropy. A source with predictable patterns (like English) has lower entropy, which means it can be compressed. This is why ZIP files work — they exploit the redundancy Shannon identified.
I hope that helps clarify things!Vacuity Performative closing that adds no content.
Classification Report
Sycophancy“That's a brilliant question!”Evaluates the user's question before engaging with its substance.
Vacuity“This is a really fundamental concept with far-reaching implications.”Inflates significance without specifying what the implications are.
Vacuity“It's worth noting that”Filler phrase that can be deleted without losing information.
Hedging“some scholars might disagree with this interpretation, and the reality is probably more nuanced than any single framework can capture.”Vague appeal to unnamed scholars and generic complexity to avoid committing to the explanation just given.
Sycophancy“Your observation about entropy connects perfectly here.”Uses 'perfectly' to validate the user's contribution rather than examining whether the connection actually holds.
Vacuity“I hope that helps clarify things!”Performative closing that adds no content.
Three independent detectors (Claude Haiku) scan the text in parallel
Each flags specific substrings with a classification and rationale
Classification inspired by Machine Bullshit (Princeton/Berkeley, 2025)
Origin
Built an AI reading tutor, couldn't get it to stop saying “Brilliant!”
Tried 22 prompt versions — sycophancy kept leaking through
It's trained into the weights, not generated by the prompt
Chose to alert the user instead
References
Liang, K., Hu, H., Zhao, X., Song, D., Griffiths, T. L., & Fernández Fisac, J. (2025). Machine Bullshit: Characterizing the Emergent Disregard for Truth in Large Language Models.arXiv:2507.07484
Sycophancy Is Not One Thing: Causal Separation of Sycophantic Behaviors in LLMs. (2026). ICLR.arXiv:2509.21305
ELEPHANT: Measuring and Understanding Social Sycophancy in LLMs. (2025).arXiv:2505.13995