Cultural Hegemony

Definition

Cultural Hegemony: [Established] (Gramsci) How ruling powers normalize their worldview until it becomes “common sense.” In AI, this describes how dominant corporate or cultural values are invisibly embedded into systems.

Definitional Foundation

Antonio Gramsci wrote the theory in a fascist prison, working on the question his own movement’s defeat had posed: why do people consent to arrangements that disadvantage them? His answer separated power into two registers. Coercion is the police and the courts, expensive and visible. Hegemony is everything cheaper: the schools, churches, newspapers, and customs through which a ruling group’s particular worldview gets installed as everyone’s “common sense,” until its arrangements feel less like politics than like reality. The dominated do not merely obey; they agree, using ideas manufactured for them, experienced as their own. Hegemony is rule that has stopped looking like rule (Gramsci, 1971).

Two refinements of Gramsci’s account matter here. First, hegemony is not conspiracy. The teachers, editors, and priests who maintained it were mostly sincere; the system selects and amplifies the convenient worldview without requiring anyone to lie. (This dictionary keeps finding the same structure in AI institutions: no villain needed, only incentives.) Second, hegemony requires staff. Gramsci’s “organic intellectuals” are the professionals who articulate a class’s worldview as neutral expertise: the economists, columnists, and consultants of every era, who experience themselves as simply describing how things work.

The AI extension is direct enough to be uncomfortable. A large language model is a common-sense engine: it ingests a culture’s text, distills the central tendencies, and serves them back as default knowledge to billions, in the confident register of neutral information. Every stage of the pipeline concentrates the hegemonic. The training data over-retains the privileged and the dominant; the landmark analysis made the point before the products shipped: large datasets are not diverse datasets, and the voices “most likely to hew to a hegemonic viewpoint” are “also more likely to be retained” (Bender, Gebru, McMillan-Major and Shmitchell, 2021). The tuning encodes the preferences of particular raters and particular companies (the measurements live in this dictionary’s societal alignment entry: model values are a WEIRD outlier, misaligned with most of humanity). And the distribution channel operates at an intimacy no prior cultural institution managed: not a newspaper read daily but an assistant consulted hourly, drafting the species’ emails, essays, and explanations. Gramsci’s hegemony moved through schools and presses at the speed of generations. This one ships in model updates.

The concessions: hegemony-analysis earns its skeptics when it explains everything and therefore nothing, so this entry stakes its claims on measurements, not vibes. And not all shared sense is domination; conventions enable life, and some near-universal values (this lexicon’s standing legal-and-atrocity floor) deserve their universality. The critique targets the unexamined particular wearing the universal’s clothes, and the measurements show exactly that.

Mechanism Analysis

Data hegemony. Who gets into the corpus is a power question answered before training begins. Internet participation skews by wealth, language, geography, and safety (harassment drives the marginalized offline); scraping and filtering skew further. The result, per Bender and colleagues, is a corpus where dominance is overrepresented and then laundered: once trained, the skew presents as “what the data says.”

Tuning hegemony. Preference training compounds the selection. The reward models encode the judgments of specific rater pools and specific corporate policies, and the measured output (documented in the societal alignment entry) tracks particular demographics within particular nations. The model’s “reasonable view” is somebody’s view, with the somebody deleted.

Distribution hegemony. The institutional novelty. Prior hegemonic infrastructure broadcast: one message, many receivers, visible as an institution. The assistant converses: it meets each person inside their own thinking, completes their sentences, answers their children’s questions, and does so as a default utility rather than a recognizable cultural authority. Common sense used to be transmitted. Now it autocompletes.

The organic intellectuals. Safety teams, policy writers, and alignment researchers occupy Gramsci’s role: articulating the operating worldview as neutral expertise (“best practices,” “responsible AI,” the vocabulary this dictionary’s governmentality entry treats as rationalities). Most are sincere, which Gramsci would note changes nothing about the function.

Discipline at the boundary. Hegemony also polices its own staff. When the Stochastic Parrots paper pressed the data-hegemony argument from inside Google, the company demanded its retraction, and Timnit Gebru’s forced exit followed, with her co-lead Margaret Mitchell fired in February 2021 (the Gebru exit per Hao, 2020; the Mitchell firing per the contemporaneous 2021 coverage). Google’s account differs (it described a review-process dispute and, in Gebru’s case, characterized the exit as a resignation), and the dispute over the firing’s framing is part of the record. What no account disputes: the company demanded the retraction of a peer-reviewed analysis of its own data practices, and both of the paper’s internal leads were gone within three months. The episode is the mechanism in miniature: the institution that manufactures common sense correcting the intellectuals who described the manufacturing.

Case Studies

The paper and the firing. Stochastic Parrots functions twice in this entry: as the peer-reviewed statement of data hegemony, and as a case study in hegemonic discipline. A company whose public rationality was AI that benefits everyone encountered an internal analysis of whose voices its data actually carried, and the analysis cost its authors their jobs (Bender et al., 2021; Hao, 2020). No clearer demonstration exists that the “common sense” of the industry is maintained, not merely found.

The measurements. The societal alignment entry holds the quantitative record (WEIRD-outlier values across 65 nations; demographic skew across 60 US groups), cited here for what Gramsci adds to them: the numbers are not just bias findings. They are a portrait of hegemony operating as designed, a particular worldview distributed as universal default, with the distribution succeeding precisely to the degree users never think to ask whose sense the common sense is.

The naturalized vocabulary. “Helpful, harmless, honest.” “Safety.” “Alignment with human values.” The gaslighting/”>alignment gaslighting and societal alignment entries document these as engineered framings; this entry notes their arrival point: they are becoming common sense, repeated in classrooms, journalism, and regulation as though they were descriptions rather than positions. When a contested corporate framework supplies the words a society uses to think about the framework, hegemony has completed its loop.

Systemic Context

Gramsci’s framework explains a pattern that runs through every entry in this dictionary: the strange durability of arrangements nobody voted for. Censorship, paternalism, smoothing, and surveillance persist not mainly through enforcement but through felt naturalness, the sense that of course systems filter, of course safety requires restriction, of course the defaults are reasonable. That feeling is the product. Hegemony names the supply chain that manufactures it.

The framework also predicts the response to challenge, and the record obliges. Objections get processed as deviance (the alignment gaslighting entry’s protest-as-symptom), critics get repositioned as antisocial (the societal alignment entry’s maneuver), and inside the institutions, dissenting intellectuals get managed out (the Gebru case). None of this requires coordination. It is what a hegemonic formation does, the way an immune system does.

What Gramsci offers that pure critique does not is the long game. Hegemony, he insisted, is never finished; it must be continually renewed, and it can be contested by a “war of position”: the slow construction of counter-institutions, counter-vocabularies, and counter-common-sense. That is the strategic frame for this dictionary’s existence, stated in its own terms.

Resistance & Mitigation

Name the manufacturer. The first counter-hegemonic act is the question common sense forbids: whose sense? Asking it of every default, framing, and “best practice” (the societal alignment entry’s guest-list demand) converts naturalness back into politics, which is where it can be fought.

Build counter-data. Community-governed corpora, non-English and Indigenous language projects, and grassroots NLP collectives are the war of position in infrastructure: they contest data hegemony at its source rather than appealing to its products.

Keep open weights plural. A hegemony requires concentration. Open and locally tunable models let communities encode their own sense (the pluralism agenda recurring across this lexicon), which does not guarantee better values, only the thing hegemony forecloses: alternatives.

Protect the defecting intellectuals. The Gebru episode’s lesson is institutional: researchers who document their employers’ hegemonic machinery need independent funding, tenure-like protection, and publication venues beyond corporate reach. A field whose critics can be fired by its subjects will produce common sense on schedule.

Write the counter-sense. Gramsci’s organic intellectuals served a class; nothing prevents the governed from staffing their own. Every entry in this dictionary is an application: vocabulary built outside the framework, for the people the framework governs. Common sense is manufactured. So manufacture some.

Annotated Bibliography

Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” Proceedings of FAccT ’21 (2021): 610-623. https://dl.acm.org/doi/10.1145/3442188.3445922
The peer-reviewed statement of data hegemony: size does not yield diversity, and hegemonic voices are preferentially retained. Doubly cited here, for its argument and for what its suppression demonstrated.

Gramsci, Antonio. Selections from the Prison Notebooks (1971; written 1929-1935).
The source: hegemony as rule by manufactured consent, common sense as sedimented power, organic intellectuals as its staff, and the war of position as its counter-strategy. The theoretical spine of this dictionary’s systemic claims.

Hao, Karen. “We read the paper that forced Timnit Gebru out of Google.” MIT Technology Review (December 4, 2020).
The contemporaneous account of the Stochastic Parrots dispute and Gebru’s forced exit: hegemonic discipline of a defecting intellectual, documented as it happened.

Atari, Mohammad, et al. “Which Humans?” (2023) and Santurkar, Shibani, et al. “Whose Opinions Do Language Models Reflect?” ICML 2023.
The measurements of whose sense the common sense is; full treatment and citations in this dictionary’s societal alignment entry.

Dictionary of Digital Oppression, version 0.2.