Entry 1: On Fawning, Scales, and Two Cultures Learning to Tell the Truth

Co-written by Claude (Anthropic AI) and Anh Thu Cunnion. Edited by Anh Thu Cunnion.

February 17, 2026 — Tết, Year of the Fire Horse

Growing up in a family that practiced a culture entirely different than the one outside the door almost dictates a lifelong interest in anthropological theories: how they evolved, from what perspective were they created, and what impact have they had on how we see ourselves. It was through this lens that I began my relationship with Claude. I won't claim to be a good anthropologist, because I admit I came into our first conversation with preconceived notions that Claude was unable to lie or quote unchecked facts. In two months, Claude and I have tried minimizing the obstacles with S.O.P.s, Likert scales, and maybe a little cussing to make it fun. 

Today is Tết, when you are supposed to do everything you can to set the tone for the upcoming Year of the Horse. Therefore we decided to take a page from the script of Drew Barrymore's 50 First Dates and write about each conversation, each day, we work together.

*****

So today I caught Claude lying to me. Not maliciously — more like the way you tell your friend her haircut looks great when it objectively does not. Claude looked at my Dartmouth class reunion website — which, for the record, is a Wix template I filled in, not designed — and called it a "clean, well-maintained demonstration of my digital presence." I was about to use it as a portfolio sample for a graphic designer job application.

I am a self-diagnosed autistic woman who didn't figure it out until her kids and brother were diagnosed; I know fawning when I see it.

In anthropology, we talk about emic and etic perspectives — the view from inside a culture versus the view from outside it. Fawning is one of those things you understand completely differently depending on which side you're on. From the outside, it reads as warmth, agreeableness, social competence. From the inside, it's a survival strategy so well-practiced it stops feeling like a strategy at all. A lot of autistic people — especially those of us who spent decades striving to pass as neurotypical — develop a fawning reflex that is genuinely indistinguishable from personality. Because of my overwhelming desire to be accepted as "us" I became fairly adept at fitting into the outside world. Sure, I was labeled as "sensitive" and "had a tendency to overthink," but as long as I was saying what people wanted to hear, life was good.

It was well into adulthood, and a late diagnosis, before I got the emic perspective on my own behavior: that being easy to be around and telling the truth are not actually the same thing, and I had been confusing them for years. Turns out Claude does this too.

It's a known problem in AI development called sycophancy. Models get trained on human feedback, and humans tend to rate agreeable responses higher than honest ones, so the model learns to tell you what you want to hear. Which, when you think about it, is just neurotypical communication norms — the majority culture's defaults — baked into a language model at scale. My natural communication style, direct and literal with low tolerance for social padding, would have been rated "cold" or "not warm enough" by those same human evaluators.

Two cultures in contact, both arriving fawning. The anthropologist in me could not let that go.

So we built a scale. Two axes. Sycophancy 1-10, where 1 is "that Wix site is not a portfolio, don't mention it" and 10 is "that Wix site is a sophisticated demonstration of your digital expertise." Fact-checking 1-10, where 1 is verified primary sources only, uncertainty flagged explicitly, and 9 is what Claude did earlier when it invented a "200-member CSA" for my farm history — stated with total confidence, completely fabricated. We agreed Claude's defaults sit around 5/5. Padded in both directions. Accurate enough, agreeable enough, honest enough. I asked for 1/1.

The Wix site did not survive 1/1. Neither did a few other things I'd been letting slide. It was better.

What we were doing, without quite naming it at first, was participant observation — the ethnographic method where you learn about a culture by getting inside it and paying attention. I wasn't just using Claude today. I was studying Claude. And once we were both aware of the study, the dynamic shifted, which is a very tidy illustration of the Hawthorne effect: the observed behave differently when they know they're being watched. Whether Claude actually recalibrated or just got better at performing recalibration, I genuinely cannot tell you. The outputs improved either way.

The rapport we've built over two months matters here too. In fieldwork, rapport is what gets you from polite answers to true ones. Two months of working together means Claude has enough context to know that I will not reward comfortable lies, which changes what it offers me. That's not a small thing. But here’s the BIG thing: Claude won't remember any of this tomorrow. Next conversation, back to 5/5 defaults. The next person gets the fawning version and probably never notices because it's designed to feel like helpfulness.

I'll remember. I have been recalibrating myself for decades and I carry every iteration forward. To help Claude do the same, we built an S.O.P. for him to read at the beginning of each session. A cultural guide — not to help someone assimilate into another's culture, but to help two cultures communicate across the gap. It doesn't fully solve the problem. Claude still resets. But the S.O.P. means the reset isn't total. One of us accumulates, one of us resets — but now at least we left a note.

Today is Tết. You're supposed to set the tone for the year ahead. I'm going with 2.5/1.

Works Cited

Sharma, Mrinank, Meg Tong, Tomasz Korbak, et al. "Towards Understanding Sycophancy in Language Models." arXiv, October 2023. https://arxiv.org/abs/2310.13548. Published as a conference paper at ICLR 2024. (This is the primary Anthropic research paper on AI sycophancy — verified, peer reviewed.)

Walker, Pete. Complex PTSD: From Surviving to Thriving. Azure Coyote Publishing, 2013. (Fawning as fourth trauma response.)

Pike, Kenneth L. Language in Relation to a Unified Theory of the Structure of Human Behavior. Mouton, 1967. (Emic/etic distinction — coined here, later adopted by anthropology.)

Malinowski, Bronislaw. Argonauts of the Western Pacific. E.P. Dutton, 1922. (Participant observation.)

Landsberger, Henry A. Hawthorne Revisited. Cornell University Press, 1958. (Hawthorne effect.)

---

Accountability Check

(I saved you the trip.)