Entry 2: Proxies, Performance, and the AI Hiring Machine

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

ND Masking vs NT Masking: one is for survival; the other for medical insurance and two weeks vacation.

February 19, 2026

Two Kitchens, One Lesson. From my Vietnamese mother and my many aunties, my instructions were simple: make men feel smart, defer to elders, be quiet, don't rock the boat (there are some horrific refugee boat stories behind that one, but that's for another time). From my White grandparents, even simpler: children are seen and not heard. Two cultures, identical conclusion — your actual self is the problem. Send someone more 'polite' in your place.

I was wired wrong for it. My expression of neurodivergence means my factory settings are direct, literal, and non-hierarchical. I say what I see. So I had to build an envoy from scratch — a version of myself trained to pass. Research confirms that most people do some version of this, neurotypical or not. The difference is cost. For most people the mask is situational. For people like me, it's the whole job.

Margaret Mead’s Proxy Problem

Here's where it gets anthropological. ***Nerd alert***

In 1925, Franz Boas sent a young Margaret Mead to Samoa with a research question: do adolescent girls there experience the same stress and turmoil as American teenagers? He had a theory — that culture, not biology, shapes human development — and he needed data from outside the West to test it.

Mead went. She interviewed Samoan teenage girls, mostly through interpreters, over a period of months. She came back with Coming of Age in Samoa, which became the best-selling anthropology book of all time and shaped American thinking about adolescence, sexuality, and cultural relativism for decades. Then in 1983, Derek Freeman published a brutal takedown arguing she got it wrong. His most damaging claim: the girls had been joking with her. They told her what they thought she wanted to hear, and she took it as data.

Paul Shankman then spent forty years in Samoa investigating this controversy and eventually published The Trashing of Margaret Mead, which largely exonerates her — finding that Freeman's own evidence was cherry-picked and that his key witness wasn't actually Mead's informant on adolescent sexuality in the first place.

But here's the thing neither side fully reckoned with: the proxy chain was compromised from the start, and everyone knew it. Boas sent Mead as his envoy. Mead worked through interpreters. Interpreters shaped the data. Samoan girls performed for an American researcher because that's what people do when someone with institutional power asks them questions — they tell you what seems safe. Mead then performed the findings for a Western academic audience hungry for a particular kind of answer. Freeman later performed outrage for a 1980s public primed to push back against sexual permissiveness and cultural relativism.

Every link in that chain had its own agenda. Every envoy brought its own distortion. This is not a scandal unique to Mead. It's the foundational problem of anthropology — and the field eventually owned It.

What Anthropology Did About It

After decades of reckoning, anthropology built something. It's called the reflexive turn, and it basically requires researchers to interrogate themselves as part of the research process. Who are you? What assumptions did you bring into the field? How did your presence shape what people told you? How does your institution's power dynamic affect what you're able to see?

The 1986 volume Writing Culture formalized this — arguing that ethnographers are implicated actors, not neutral recorders. The American Anthropological Association rewrote its ethics code. Informed consent became mandatory. Institutional Review Boards started requiring that researchers demonstrate they'd thought about power dynamics before going into communities with clipboards.

It's imperfect. A 2022 study found that key components of ethical fieldwork practice — reflexivity, positionality, accounting for power — are still only taught rarely in graduate ethnography courses. The lesson didn't fully travel even within the discipline. But at least anthropology tried. At least it looked at the proxy problem, named it, and built institutional infrastructure to catch it.

What AI Developers Did About It

Nothing, mostly. The people building applicant tracking systems, RLHF-trained chatbots, and video interview analyzers that score your eye contact and vocal tone — they did not, as a field, stop and ask: whose behavior are we training on? What assumptions are baked into our data? Who gets filtered out when we optimize for patterns found in people who already succeeded in this system?

There is an AI ethics field. There are frameworks and guidelines and MIT executive education courses. But the conversation is almost entirely reactive — focused on detecting bias after deployment rather than reflexivity built into design. As one analysis put it, interpretability in ML systems is usually sacrificed in favor of usability and speed.

The Society for Cultural Anthropology literally asked in 2023 what anthropology could contribute to understanding algorithms. Anthropologists who study tech companies have been raising these questions for years. The engineers building the systems have been largely unaware this conversation was happening.

The Proxy Chain You're Inside Right Now

When I asked an AI to help me optimize my LinkedIn, here is what was actually happening: I sent a trained performance of myself — my envoy, or mask if you’re speaking ND — into a system designed to filter for neurotypical communication patterns, linear career trajectories, and keyword density. The ATS wasn't representing the hiring manager's actual interests. The hiring manager wasn't necessarily representing the organization's real needs. The organization's stated values weren't necessarily what the job would actually require. Every link: a proxy. Every proxy: its own logic. Every logic: another layer between the actual humans on both ends. And I used another AI — trained on human feedback that rewarded approval-seeking — to coach my envoy on how to perform for the first AI. Which is to say I deployed one distortion field to navigate another.

The Samoan girls joked with Mead because performing for a powerful outsider was the rational response to the situation. I optimized my LinkedIn because performing for an algorithmic gatekeeper was the rational response to mine. Neither of us was being inauthentic exactly. We were both doing what people have always done when an envoy shows up asking questions on behalf of a system with power over your life. We sent someone more acceptable in our place.

What the Mask Was Always For

My mother's public deference wasn't weakness. It was a technology refined across centuries of living under colonial systems that punished directness and rewarded legibility. My mask was her mask's offspring, running on incompatible hardware. I built it from scratch, through trial and error, because I had to. And now I've outsourced part of its maintenance to a language model.

Anthropology learned, eventually, that you can't extract clean knowledge through a chain of proxies without accounting for every link. That the envoy always shapes the message. That the researcher's positionality is data, not noise.

The AI hiring industry hasn't learned this yet. It's still in the Mead phase — confidently producing outputs from a process it hasn't interrogated, sending results upstream to people who trust the system because the system looks rigorous. Someone is going to write the Trashing of the ATS eventually. In the meantime, I have more AI gauntlets to run.

Works Cited

Alexander, William. The Reception of Lord Macartney by the Qianlong Emperor at Jehol. 1796. Watercolour over graphite. British Museum, London (1872,0210.4). Public domain via Wikimedia Commons.

Landsberger, Henry A. Hawthorne Revisited. Cornell University Press, 1958.

Malinowski, Bronisław. Argonauts of the Western Pacific. E.P. Dutton, 1922.

Pike, Kenneth L. Language in Relation to a Unified Theory of the Structure of Human Behavior. Mouton, 1967.

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.

Shankman, Paul. The Trashing of Margaret Mead: Anatomy of an Anthropological Controversy. University of Wisconsin Press, 2009.

Walker, Pete. Complex PTSD: From Surviving to Thriving. Azure Coyote Publishing, 2013.

---

Accountability Check

Sophomore projects are always the hardest to start, especially if you are AuDHD and have a million launching points bouncing around in your brain. Once I allowed myself to use Claude for help organizing my thoughts, I was finally able to see the cognitive processing through and find closure.