Unscarcity
Sign in for free: Preamble (PDF, ebook & audiobook) + Forum access + Direct purchases Sign In By signing in, you get our daily podcast email. Unsubscribe any time.

Unscarcity Research

AI Governance: Why Hawk-Eye Works But Doesn't Make Rules

Tennis let Hawk-Eye replace its line judges, and nobody lets it write the rules. Why machines enforce, humans decide, and the override is the whole point.

15 min read 3290 words Updated September 2026 /a/ai-as-referee-humans-as-conscience

Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

AI as Referee, Humans as Conscience

Why the robot blows the whistle, but you decide if the game is worth playing

The Tennis Court Epiphany

In the quarterfinals of the 2004 U.S. Open, Serena Williams lost a three-set match to Jennifer Capriati in which four questionable calls went against her. The worst came from the chair: the umpire overruled a ball a line judge had correctly called in, and television replays showed it in by nearly a foot. The next day the United States Tennis Association apologized to Williams and said it would look at video replay. “With regard to the Serena call at the U.S. Open,” Arlen Kantarian, its chief executive for professional tennis, said later, “we certainly feel that might have put us on a faster track.”

Two years later the U.S. Open became the first Grand Slam to let players challenge a call against Hawk-Eye, a system of cameras that reconstructs the ball’s flight and shows where it landed. (Its makers used to claim an average error of 3.6 millimeters and now claim 2.2.) The first challenge at a major came from Mardy Fish. He lost it: a few seconds later the replay came up on the stadium scoreboard and the judge’s call stood.

Twenty years on, three of the four Grand Slams have removed the line judges altogether. Wimbledon retired them in 2025, after 147 years. The machine now makes the call itself, within a tenth of a second.

Through all of it, no one has suggested Hawk-Eye should also decide the rules of tennis.

Nobody said, “Let the algorithm determine whether tiebreakers are fair.” Nobody proposed, “The AI should choose which tournaments count toward rankings.” Hawk-Eye measures; humans deliberate. The machine enforces; people decide what’s worth enforcing. That separation is the blueprint for governing a post-scarcity civilization.

It comes with one condition, which tennis is learning the hard way. We’ll get there.

The Division Nobody Talks About

Every governance system in history has faced the same problem: how do you scale decision-making without becoming either a bureaucratic nightmare or a tyranny? AI makes this question urgent because it can finally make enforcement cheap while judgment remains expensive.

Rome solved it with magistrates, then got emperors. The Catholic Church solved it with bishops, then got the Inquisition. Modern democracies solved it with agencies, then got regulatory capture. The pattern repeats because all these solutions conflate two different functions:

  1. Rule enforcement (Did the ball land in or out?)
  2. Rule formation (Should the court be this size? Should we even play tennis?)

When the same entity does both, power accumulates. The rule-enforcers become rule-makers. The referees start owning teams, and the line judges redesign the court to favor their friends.

The Unscarcity solution is simple: separate the functions at the substrate level. AI handles enforcement. Humans handle formation. The machine blows the whistle, and the people write the rulebook.

Wikipedia: 2,844 Bot Tasks, Zero Robot Overlords

If you think this sounds utopian, you’re already using a system that works this way, and has for twenty-five years (Wikipedia turned 25 on January 15, 2026).

Wikipedia’s English edition held 7.25 million articles in late September 2026. About 110,000 registered editors make at least one edit in a given month (the Wikimedia Foundation counts nearly 250,000 across all languages), working alongside 308 accounts flagged as bots and 2,844 approved bot tasks. The task list grew by 77 in the past year while the number of human editors barely moved, and nobody is alarmed, because the ratio was never the thing that mattered. Bots made 18% of all edits to English Wikipedia in August 2026. Eight of the ten most prolific “editors” in its history aren’t human. ClueBot NG, the most famous anti-vandalism bot, is tuned to catch about 40% of all vandalism, and nearly every revert it makes lands within 30 seconds of the offending edit. It has made 6.7 million edits.

And yet any human can override any bot. ClueBot reverts your edit? You can put it back. A categorization bot mislabeled your article? You can fix it. No permissions required, no appeals board, no waiting period. Human judgment trumps algorithmic judgment, always.

That override power is the core of Wikipedia’s design, not a flaw in it. Bots handle the predictable: formatting, link maintenance, vandalism detection, statistics updates. Humans handle the contested: whether a source is reliable, whether content is notable, whether an article maintains neutral point of view. AI for mechanics, humans for meaning.

When you search for “Wikipedia governance” you’ll find endless debates about administrator elections, arbitration cases, and policy changes. You’ll find almost nothing about bot decisions. That’s the point. The mechanical layer operates invisibly while the meaningful layer operates democratically.

This is exactly what the MOSAIC architecture proposes for post-scarcity civilization.

The Civic Layer: Referee, Not Ruler

In the Unscarcity framework, the Civic Layer is the AI-augmented infrastructure that coordinates resource allocation and light governance for the Foundation (the 90% baseline that covers everyone’s survival needs).

The key word is light.

The Civic Layer operates as a referee and registrar, not a ruler. It manages logistics like housing, food distribution, and energy allocation without requiring human deliberation for routine operations. When a Commons needs 50,000 more kilowatt-hours, the Civic Layer routes the energy. When a resident requests shelter, the Civic Layer identifies available housing.

But the Civic Layer never decides values.

It doesn’t determine whether your Commons should prioritize density or sprawl. It doesn’t choose whether your culture values silence or celebration. It doesn’t rule on whether art or athletics deserves more resources. These are human questions - questions of meaning, aesthetics, and philosophy.

The Civic Layer intervenes reactively only when one of the Five Laws is violated:

  • Someone’s blocking another person’s Foundation access? Flag.
  • A process is being hidden from public view? Flag.
  • Power is concentrating without decay? Flag.

Then humans deliberate. The AI identified the violation. Humans determine the response.

This is the Singapore traffic model at civilizational scale. Singapore’s GLIDE system runs every traffic signal on the island, about 2,700 intersections, through 18 regional computers watched by a central one. Sensors in the road measure the traffic. The software sets the timing and strings the green lights together. Human operators step in for the exceptions. The machine decides when your light turns green. Humans decided whether to build the road in the first place.

It has been adjusting signals to traffic since 1988, and in all that time nobody has asked it for an opinion on urban policy.

The Oversight Trap: Humans in Loops Get Rusty

The phrase “human-in-the-loop” has become a governance buzzword. The EU AI Act mandates it for high-risk systems under Article 14. Corporate ethics boards invoke it as a safety blanket. Stanford’s Human-Centered AI Institute turned the phrase around in 2022, arguing for “AI in the loop – where humans remain in charge.” The research on oversight explains why the order matters: humans in loops get lazy.

The paradox is that if an AI is right 99% of the time, humans stop paying attention. Why override something that’s almost always correct? The human becomes a rubber stamp, nodding along until they’ve lost the facility for independent judgment.

This is called “automation complacency,” and it’s not hypothetical. A 2010 review in Human Factors found it “in both naive and expert participants” and concluded that it “cannot be overcome with simple practice.” As Cory Doctorow summarized the policy scholar Ben Green’s work, experts put in charge of overseeing automated systems “get out of practice, because they no longer engage in the routine steps that lead up to the conclusion.” In 2025 a study in The Lancet Gastroenterology & Hepatology put a number on it: after a few months of AI-assisted colonoscopies, experienced endoscopists found adenomas in 22.4% of the procedures they did without the AI, down from 28.4% before it arrived.

The Wikipedia model avoids this trap by not making human oversight the default path.

Most bot edits are never reviewed by humans, and that’s fine. Humans don’t need to verify every formatting fix. What matters is that humans can intervene when they care to. The override button is always there. But humans exercise judgment selectively, on the cases that matter, rather than rubber-stamping everything.

The Civic Layer follows the same principle. Most resource allocations don’t need human review. The AI routes energy, distributes food, and assigns housing across millions of transactions with zero drama. Humans engage when the system flags an anomaly, when a dispute arises, or when policy needs updating.

The AI handles the 99% that’s obvious. Humans focus on the 1% that requires wisdom.

Europe Just Showed How Hard the Easy Part Is

Writing “a human must supervise this” into law is cheap. Building institutions where the supervising human still has the skill and the standing to disagree is not.

Europe demonstrated the gap in 2026. The Digital Omnibus — provisionally agreed on May 7, endorsed by Parliament on June 16, and cleared by the Council on June 29 — pushed the AI Act’s high-risk obligations back: sixteen months for standalone Annex III systems, from August 2, 2026 to December 2, 2027, and twelve months for AI embedded in regulated products, to August 2, 2028. That agreement became binding law as Regulation (EU) 2026/1744, signed July 8 and published in the Official Journal July 24. It took effect July 27, six days before the original deadline it replaced. The human-oversight requirement itself came through untouched. Only the clock moved.

That is the tell. Nobody argued that Article 14 was wrong. An entire continent simply discovered that mandating meaningful human oversight and staffing meaningful human oversight are different projects, and asked for another year or more to do the second one. The rule was never the hard part.

The staffing problem is not a figure of speech. In August 2026 Lithuania’s communications watchdog told Euractiv its government had given it two full-time positions to enforce its share of the Act. Germany’s network regulator had about 20 people on the job, of 43 planned.

Wikipedia never had this problem, because it never promised that a human reviews everything. It promised that a human can. One is a staffing commitment that scales with volume and collapses under it. The other is a permission that costs nothing to keep and everything to remove.

“But Won’t the AI Get It Wrong?”

Yes. Constantly.

ClueBot NG is set to keep its false positives at or below 0.1%. That sounds excellent until you realize 0.1% of millions of decisions means thousands of errors. Good-faith editors get their contributions wrongly reverted. Legitimate articles get mislabeled as vandalism.

Wikipedia handles it gracefully. Every bot decision is immediately reversible. False positives are tracked and fed back into training. Users aren’t punished for having edits incorrectly reverted. The community tolerates the error rate because the alternative, having humans review every edit, doesn’t scale.

The insight that matters: the question isn’t “will the AI be perfect?” The question is “what happens when it’s wrong?” If the cost of AI error is catastrophic and irreversible, you need more human oversight. If the cost is annoying but fixable, let the machine work and correct as needed.

The Foundation’s Civic Layer handles high-stakes decisions (someone’s shelter allocation) differently than low-stakes decisions (routing tonight’s vegetable surplus). Both use AI coordination. But the human oversight intensity scales with the consequences of error.

What Tennis Forgot

Tennis is running the experiment right now, without the override.

On July 6, 2025, in Wimbledon’s first year without line judges, an operator switched the system off by mistake for one game on Centre Court. Three calls went unmade, and the chair umpire didn’t know the machine was off. Anastasia Pavlyuchenkova said the game had been “stolen” from her. The All England Club’s fix was to remove the operators’ ability to turn ball-tracking off.

On August 31, 2026, at the U.S. Open, the system called a forehand from Anna Kalinskaya out after it had landed well inside the line. The USTA blamed “a malfunction with the ELC Live system.” The umpire couldn’t fix it on the spot, because on the tours and at three of the four majors a chair umpire cannot overrule a live electronic call. A video review existed, and the officials didn’t use it.

The fourth major has noticed. Roland Garros kept its line judges for 2026, and after a disputed call in the fourth round its director, Amélie Mauresmo, explained why: “As of today, the machine is not 100 percent reliable, so we continue to place our confidence in human officials.”

Players argue with the robot now. They photograph ball marks. They tell umpires the system is wrong, and sometimes it is. The arguments came back at the moment the sport removed the person who was allowed to disagree.

That’s the condition. A referee you can overrule is a tool. A referee you can’t overrule is a ruler with a very narrow job. Wikipedia never made that mistake, and the Civic Layer can’t afford to.

Matching Oversight to Stakes

This is common sense we’ve somehow forgotten. We don’t require a human to approve every email spam filter decision. We don’t need a judge to validate every traffic light. But we do want humans involved in parole decisions and medical diagnoses.

The Civic Layer applies the same gradient. Lettuce routing? Full automation. Housing disputes? Human review. Constitutional challenges? Full Diversity Guard deliberation.

The Alternative Is Worse

Critics miss this: the choice isn’t between “AI governance” and “human governance.”

The choice is between coordinated AI + human deliberation and uncoordinated human bureaucracy + de facto AI manipulation.

We already live in a world shaped by algorithms. Your social media feed is algorithmically curated. Your job application is algorithmically screened. Your credit score is algorithmically calculated. Your news exposure is algorithmically filtered.

These systems aren’t transparent. They aren’t reversible. They aren’t accountable. They don’t have human override buttons. They’re optimized for engagement, profit, or efficiency, not for your flourishing.

Once in a while a regulator catches one. In August 2026 the Dutch data protection authority fined Uber €825 million for deactivating drivers by fully automated decision between 2018 and 2022: software flagged a fraud suspicion or a low rating, the account went dark, and no person looked. “A computer should not make decisions on its own that have major consequences for you,” the authority’s deputy chair said. Uber has appealed. It took four years of enforcement to establish what Wikipedia builds in by default.

The Unscarcity proposal doesn’t add AI to governance. It makes the AI that’s already governing visible, auditable, and subordinate to human values.

The Civic Layer publishes its source code. Its decisions appear on distributed ledgers (DPIF). Its logic is comprehensible, not black-boxed. Any human can audit any decision. Any Commons can challenge any pattern.

This is Law 2: Truth Must Be Seen - applied at the infrastructure level.

The alternative, pretending we can govern complex systems with human bureaucracy alone, is how you get regulatory agencies staffed by 10,000 people who still can’t track derivatives, or planning departments that take eighteen months to approve a building permit.

AI handles the speed and scale problems. Humans handle the meaning and direction problems. Neither can do both. Both are necessary.

The Game Itself

Return to the tennis court for a moment.

Hawk-Eye calls the line to within a few millimeters, faster than a human can blink. No human could match it. On a good day no one argues.

But Hawk-Eye didn’t end arguments about tennis. Players still debate rule changes. Fans still discuss whether tiebreakers are exciting or artificial. Tournament directors still decide whether to prioritize tradition or innovation.

The game continues. The disputes about mechanics shrink to the rare day the machine is wrong, and on that day somebody with a name has to be allowed to say so.

This is the vision: a civilization where the tedious arguments about “did you comply with procedure X” are settled instantly by machines, freeing humans to engage in the important arguments about “should procedure X exist in the first place.”

The AI blows the whistle. You decide if the game is worth playing.

References

  • UnscarcityBook, Chapter 2, Chapter 3
  • Geoffrey C. Bowker & Susan Leigh Star, Sorting Things Out (1999)

Tennis

Wikipedia and Singapore

Oversight

EU AI Act

Share this article:

Share Share Share