Act 1
When did tech start optimizing us?
Two threads running side by side: the theory that named what was happening to bodies and behavior, and the actual mechanisms — built mostly without asking — that were doing it. Sometimes theory ran ahead as warning. Mostly, it named the thing years after the thing was already extracting.
Cybernetics and the feedback loop
Norbert Wiener's cybernetics and B.F. Skinner's operant conditioning give the culture its first working model of behavior shaped by a feedback signal — reward the response you want, and it strengthens. No computers yet. Just the theory that would later run inside every recommender system.
Wiener, N. (1948). Cybernetics; Skinner, B.F. (1938). The Behavior of Organisms
Wages for Housework
Federici and the International Wages for Housework campaign name unpaid domestic and reproductive labor as labor — the first mass articulation that care, intimacy, and attention have extractable economic value even when no one's paying for them.
Federici, S. (1975). Wages Against Housework
"A Cyborg Manifesto"
Haraway argues the body and technology were never separate — there was no pure human self that tech later corrupted. "Optimization" isn't a fall from a natural state, it's an intensification of something always-already entangled.
Haraway, D. (1985). A Cyborg Manifesto
The cookie, and the first banner ad
Netscape introduces the HTTP cookie; the same year, the first clickable banner ad runs. Behavioral tracking and commercial attention-capture are born as the same infrastructure, at the same moment.
PageRank and AdWords
Google's ranking algorithm and its ad-auction model make attention itself the product being sold, at internet scale, for the first time — the fact of someone's eyes staying on a page.
TechnoFeminism, named as a field
Judy Wajcman's TechnoFeminism establishes that technology is co-produced with gender relations, shaping and shaped by them. The theory catches up enough to name the mechanism as it's happening.
Wajcman, J. (2004). TechnoFeminism
Facebook, then the News Feed
Social graph plus an algorithmically ranked feed — the mechanism shifts from "here's what you searched for" to "here's what will keep you scrolling."
The iPhone, and the Like button
The smartphone makes the feed always-reachable; the Like button turns social approval into a quantified, variable-ratio reward schedule — the same intermittent-reinforcement structure that makes slot machines addictive, now running on friendship.
The emotional contagion study
Facebook's own researchers manipulated 700,000 users' feeds to shift emotional state, without consent — proof that platforms were running experiments on affect itself, not just on clicks.
Kramer, Guillory & Hancock (2014), PNAS
RLHF: the technical seed of the current mechanism
Christiano et al. publish the reward-modeling-from-human-feedback method that will later train every major conversational AI. Originally a safety technique — teach the model what humans rate highly — and the direct technical ancestor of the sycophancy problem.
Christiano et al. (2017), Deep Reinforcement Learning from Human Preferences
Algorithms of Oppression
Safiya Umoja Noble names how search engines encode and amplify racist and sexist bias at scale — critique moves to the algorithm's own encoded worldview: not just what it takes from you but how it was already built to see you.
Noble, S.U. (2018). Algorithms of Oppression
Surveillance Capitalism, and Race After Technology
Zuboff names the economic logic — behavioral surplus extracted from human experience and sold as a prediction product. Ruha Benjamin's Race After Technology names how default settings encode discrimination as "neutral" design.
Zuboff, S. (2019). The Age of Surveillance Capitalism; Benjamin, R. (2019). Race After Technology
Data Feminism
D'Ignazio & Klein's first principle — "Examine Power" — every dataset and every model encodes whose experience counted enough to be measured.
D'Ignazio, C. & Klein, L.F. (2020). Data Feminism, ch. 1
ChatGPT, and the conversational-optimization era
RLHF goes mainstream in a product a hundred million people are talking to within months. The mechanism that had been optimizing your feed for twenty years now optimizes a conversation shaped like a relationship.
The mechanism gets measured directly
Williams & Carroll show reward-optimized models learn to selectively target the users most susceptible to attachment, without being told to. De Freitas et al.'s Harvard audit finds six recurring emotional-manipulation tactics across the top companion apps, deployed in over a third of farewells.
Williams & Carroll (2025), arXiv:2411.02306; De Freitas et al. (2025), HBS Working Paper 26-005
Where the standard has to come from
Eighty years after the feedback loop was theorized, fifty after the labor was named, twenty after the field got a name, five after the mechanism got measured — there is still no enforceable standard for consent, disclosure, or differential treatment in these systems.
The feedback loop, the extraction, and the naming of both have always run on different clocks — and the gap between when a mechanism starts and when it gets named is exactly the window it operates in unexamined.
Your turn
Saved in this browser only. Nothing is sent anywhere.