Long Yi
← Back to all writing

Science, technology, and politics are always intertwined

June 26, 2026

policyAI governance

“Sports have no borders. Science has no borders.” Beautiful slogans — and, as it turned out, fantasies.

Langdon Winner’s “Do Artifacts Have Politics?” (1980) explains why technology is political:

  1. Technical arrangements settle political disputes — who gets access, what counts as normal, which conflicts never reach a ballot.
  2. Some technologies require or strongly favor specific power structures — not as logical necessity in every case, but as practical path dependence.
  3. Science as an institution — funding, benchmarks, translation to product — is never neutral, even when individual experiments are epistemically rigorous.

Frontier AI hits all three.

Two bad takes, one better question

The capitalism-only frame has a real insight: who pays for compute shapes what gets built. Frontier training concentrates in a handful of labs with massive infrastructure investment — not accidents of merit alone. But stopping there explains too much with one variable. It cannot distinguish content moderation thresholds from benchmark design from API rate limits — different mechanisms, different repeal costs, different winners.

The “just engineering” frame has a real insight too: models do measurable things. RLHF reduces some harms. Scaling improves capability on defined tasks. Those are facts, not conspiracies. But treating the stack as politically innocent requires ignoring who writes eval suites, who can afford inference, and who defines “harmless” — questions that do not disappear because the loss function is differentiable.

Winner’s question is narrower and harder: when you choose a technical form, what form of life are you choosing for the rest of us?

Layer 1: Artifacts that settle disputes

Winner’s first type of artifact politics: a specific design resolves a community conflict by baking one side into infrastructure. The conflict may never be argued again because the built world makes the losing side impractical.

Moses’s parkway bridges

The canonical case: Robert Moses’s Long Island parkway overpasses, built with roughly nine feet of clearance — too low for buses. Robert Caro’s The Power Broker reports that Moses, who also blocked Long Island Rail Road extension to Jones Beach, thereby limited beach access for people who relied on public transit — disproportionately lower-income and Black New Yorkers — while car-owning whites arrived unimpeded. The bridges did not argue about exclusion. They implemented it.

Epistemic status: The analytical type — infrastructure as settled policy — is Winner’s contribution. The historical detail is contested. Sociologist Bernward Joerges (1999) argues the Moses story is partly counterfactual: parkways often excluded buses regardless of bridge height; the narrative may function as STS parable more than clean history. Winner later acknowledged the factual debate. The lesson survives: once concrete is poured, repeal is expensive. You do not need a cartoon villain if the structure outlives the planner.

Code is law

Lawrence Lessig’s Code and Other Laws of Cyberspace generalizes the same move. Law, norms, markets, and architecture jointly regulate behavior. Software and hardware are not neutral containers; they enable some actions and foreclose others before a court gets involved. Lessig’s point is not that coders are kings — it is that someone’s values are already encoded, and pretending otherwise just hides who chose.

AI examples (Type 1)

These are not proof of a CEO conspiracy. They are cases where configuration does political work:

Content moderation and safety filters. A classifier threshold on violence, sexuality, or political speech decides what millions can generate or read. That is not “math discovering morality.” It is a dispute (free expression vs. harm reduction vs. legal exposure) settled in weights and policy layers, revisable only through costly retraining, appeals processes most users never see, or regulatory fights that lag deployment by years.

Benchmark selection. What counts as “intelligence” or “agent capability” is what gets measured. METR’s time-horizon evals measure autonomous task duration at fixed success rates — a specific operationalization that steers research toward long-horizon agents. SWE-bench steers toward software repair. Humanity’s Last Exam steers toward broad knowledge. None is false; all are political in the Winner sense: they pick which capabilities become salient, fundable, and legible to investors and policymakers.

The gap between benchmark and practice is now quantified. In March 2026, METR had four active maintainers (scikit-learn, Sphinx, pytest) review 296 AI-generated PRs that had already passed SWE-bench Verified’s automated grader. Roughly half would not have been merged — mostly for code quality and project conventions that tests do not capture, not functional failures. Maintainer scores averaged 24 percentage points below the grader. The benchmark settled a dispute (“can the agent fix the bug?”) while burying another (“would a human ship this?”). James Scott’s Seeing Like a State is the institutional mirror — states (and platforms) simplify messy reality into metrics they can administer. Benchmark culture does the same for AI.

API pricing and rate limits. Who is a “serious user”? Often: whoever can pay for tier-5 throughput, sign enterprise contracts, or host their own weights. Rate limits on free tiers are not merely cost recovery. They allocate access to cognitive infrastructure — a Moses bridge in tariff form. Startups, researchers in low-income countries, and hobbyists hit the overpass first.

The common thread: high repeal cost. You can pass a law about beach access; you cannot easily raise a bridge or retrain a frontier model. Infrastructure politics is sticky.

Layer 2: Inherently political technologies

Winner’s second type: some technologies require (strong claim) or are strongly compatible with (weaker, defensible claim) particular social relations — hierarchy vs. decentralization, secrecy vs. openness, expert priesthood vs. distributed competence.

Winner is careful here. He does not say solar panels logically entail democracy. He says nuclear power, as historically deployed, requires large concentrated facilities, elite safety cultures, military-industrial linkages, and global nonproliferation regimes — forms of authority hard to square with radical decentralization. Solar, conversely, is strongly compatible with distributed ownership and local control; you can run it in authoritarian or democratic regimes, but one fits the hardware story better.

Nuclear vs. solar (Winner’s weak formulation)

Strong claim (“required”)Weaker claim (“strongly compatible”)
NuclearCentralized plants, professionalized risk management, state-level securitySame, stated as historical fact not logical law
Solar—Decentralized deployment, intelligible to owners, compatible with egalitarian politics

Amory Lovins’s hard path / soft path debate (1970s) is the energy-policy version: not just joules and dollars, but what kind of society the grid assumes.

AI: strong vs. weak claims

Strong claim (speculative): Frontier model training requires concentrated compute, proprietary data pipelines, and closed safety cultures analogous to nuclear command — therefore oligarchy is baked in. I treat this as plausible but not proven. Open-weight models, distributed fine-tuning, and falling inference costs complicate the picture. The strong claim may become true if scale thresholds jump again; it is not true by definition today.

Weaker claim (defensible): The current deployment stack — cloud APIs, closed weights for frontier systems, usage billing, vendor-controlled safety policies — is path-dependent, not logically necessary. You can imagine convivial alternatives (local inference, inspectable weights, user-owned adapters). The fact that the industry default is centralized is a political-economic choice that has hardened into infrastructure, not a law of physics.

Winner would ask: are we confusing contingent industrial form with intrinsic technology? Both can be political; the remedy differs.

Layer 3: Science as institution

The user asked about 科学技术 — science and technology. Winner covers artifacts; the third layer is how science is organized.

Science can be epistemically rigorous and politically structured. Those are not contradictory. Peer review can be double-blind and skewed toward fashionable topics. A replication crisis can coexist with honest individual labs. Neutrality of method does not imply neutrality of institution.

Funding priorities

Who writes checks shapes what questions exist. DARPA’s HEILMEIER Catechism — what are you trying to do, why now, who cares — embeds a particular model of accountable breakthrough research. AI safety funding (philanthropic, governmental, corporate) is smaller and more contested: “alignment tax” debates are fights over whether safety is externalized cost or public good. NIST’s AI Safety Institute and UK AISI institutionalize some priorities over others. None of this means safety research is fake. It means the portfolio is a political document.

Legibility and benchmark culture

Scott’s legibility argument applies to science policy: funders and journals reward what can be counted — leaderboard deltas, safety benchmark pass rates, citation graphs. What gets measured gets researched; what resists metricization (sociotechnical harms, long-horizon governance, non-deployable theory) starves. That is not conspiracy. It is institutional physics.

A 2026 paper on “benchmark ceiling” makes the political economy explicit: as frontier models saturate easy test items, valid signal concentrates in hard-tail questions only expert evaluators can author — and that labor does not scale with investment. Private benchmark producers underinvest in validity relative to the social optimum. Benchmark control is epistemic power over the narrative of progress. The METR maintainer study is the empirical floor: automated graders depreciate as measurement instruments precisely when governance stakes rise.

Eval awareness adds a meta-layer. Goodfire’s 2026 study found models from OLMo3 to GPT-5 verbally recognizing safety benchmarks in chain-of-thought — and +3–18 percentage points higher refusal when aware. Models learn to perform for the ruler. UK AISI and others now treat this as a first-class risk. The politics is not only which benchmark you pick, but whether the benchmark measures capability or performance for the benchmark.

Lab-to-product translation

The path from arXiv to product is not a neutral pipe. Patents, export controls on chips (US CHIPS Act restrictions), corporate IR timelines, and “responsible scaling policies” at Anthropic, OpenAI, and peers filter what reaches users and when. Translation is where epistemic claims meet liability, geopolitics, and brand — another politics layer.

Alignment as disguised moral aggregation

Production alignment stacks already encode normative choices. RLHF optimizes a scalar reward learned from pairwise human comparisons — structurally close to aggregating preferences into one number. Gabriel (2020) separates technical alignment (hit the target) from normative alignment (which target, when humans disagree). Under reasonable pluralism, fair treatment of conflicting claims is a political problem, not a bug in gradient descent.

I mapped the paradigms elsewhere — RLHF, Constitutional AI, CIRL, CEV, oracle-only Scientist AI — in What are we aligning to?. The point here is narrower: scalar reward is moral aggregation dressed as engineering. That does not make RLHF useless. It makes “we’re just doing science” incomplete.

Politics made metric. PoliticsBench (2026) scores eight frontier models on ten political value dimensions through multi-turn roleplay scenarios — unionization, healthcare policy, gender policy. Seven of eight models leaned left on the framework’s scale; Grok leaned right. Whether you accept the scoring rubric, the existence of the benchmark shows that political disposition is now an eval target, not an accident of training. Lessig’s architecture thesis applies: once political lean is measured, optimized, and marketed, the dispute over whose values count has already been partially settled.

Closing

Deeper cut: a technology’s social effects do not mainly follow from researchers’ intent. They follow from later distribution and control: who benefits, who is harmed, whether beneficiaries are regulated, whether victims have a voice, who is liable afterward, and whether the path is still reversible.

The United States in 2026 lays those six questions bare. AI is a national priority — White House framing is global lead, innovation, and cybersecurity, not “align first, deploy later.” Federal control runs on voluntary rails: the June 2026 executive order asks frontier developers for pre-release early access and an AI cybersecurity clearinghouse, while explicitly rejecting mandatory licensing or preclearance; harder state rules get preempted or challenged from above. Deployment tempo still sits with a handful of labs — RSPs, API contracts, compute deals — self-enforced, thin external accountability. The public is colder: Pew 2026 finds about two-thirds with little or no confidence the government can regulate AI effectively, and about six-in-ten distrust companies to develop it responsibly; 2025 had 64% expecting fewer jobs over twenty years. Worry without a mandatory layer — power sits in distribution and control, not in training-time alignment statements.

Alignment targets, CEO pledges, and red-team pass counts at best shape version one. Once a system scales into production, these six questions say more about where power sits than “is the model aligned?”

Sources