Four Mechanisms, One Outcome: A Correction to the Synthocracy Thesis
METHODS · WORKING PAPER — Governance. Programme: Admissibility & Evidence. Empirical and analytical; anchored to dated, citable law as of July 2026.
CORRECTION NOTICE. This paper corrects the central empirical claim of the Synthocracy Institute, as stated in What Is Synthocracy?, the Colorado working paper, the EU policy brief, and Signposts, Dial 6. Evidence from two jurisdictions we had not examined refutes the claim in the form we made it. What follows states what was wrong, what survives, and — for the first time — what would refute us.
We claimed that when AI accountability law meets reality, disclosure obligations hold and decision-accountability obligations slip. We anchored the claim in two cases: Colorado, which repealed its risk-based AI Act before it took effect, and the European Union, which deferred its high-risk obligations by sixteen months while transparency duties held their date.
Two further jurisdictions refute that claim as stated. South Korea enacted decision-accountability obligations and brought them into force on schedule. China built the most demanding pre-deployment gate for algorithms in the world, and it has operated for four years. Neither deferred. Neither repealed. A thesis that predicts deferral or repeal predicts wrongly in half of the cases now before us.
What survives is narrower, and — because it is narrower — it is falsifiable in a way the original was not. The pattern is not in the mechanism. It is in the outcome: across four jurisdictions and four entirely different legal instruments, the form of accountability spreads and its substance does not. In none of the four can a person subject to an automated decision reconstruct that decision and contest it. That is the claim we now make, and the rest of this paper is about how to kill it.
What we got wrong
The original thesis specified a mechanism. It said accountability obligations would be postponed or removed while disclosure obligations survived. That is a strong claim, which is to its credit, and it is false.
South Korea. The Framework Act on AI — the AI Basic Act — was passed in December 2024, promulgated in January 2025, and took effect on 22 January 2026, together with its Enforcement Decree. For high-impact AI, it requires operators to establish risk-management plans, provide explanations of how outputs are generated including the key criteria and an overview of training data, implement user-protection and complaint-handling mechanisms, ensure human oversight, and maintain documentation. Risk management — the very obligation Colorado stripped — is a core duty, in force, today.
China. The Administrative Provisions on Algorithm Recommendation, in force since 1 March 2022, require providers of algorithms with public-opinion attributes or social-mobilisation capability to file with the Cyberspace Administration. The CAC grants the filing and issues a number within thirty working days — and where materials are incomplete, refuses the filing and states reasons in writing. The number must then be displayed prominently in the service, with a link to the published information. The filing pack covers the algorithm’s mechanism, its application scenarios, a security self-assessment, data sources, training data, and identified risks with mitigations. Generative services with public-opinion attributes must pass a security assessment before being offered publicly. Over five thousand algorithms have been filed.
Read that description without the country attached and it is the Institute’s own method: a pre-deployment, record-based, per-route access decision with terminal outcomes, reasoned refusal, and a public status marker. Admissibility has a working implementation at national scale, and we had not written a word about it.
Both facts refute the thesis as we stated it. We state them here, in full strength, before saying anything in our own defence.
What survives
Now look past the instrument to what a person subject to one of these systems can actually do.
In Korea, the obligations that carry immediate operative force are the transparency duties — prior notification that AI is in use, and labelling of AI-generated content. The decision-accountability obligations are drafted in a register that will be familiar to anyone who has read Colorado’s surviving human-review right: operators must make efforts to assess impact on fundamental rights, and must explain outputs to the extent technically feasible. The government is operating a grace period of at least one year during which investigations and administrative fines are generally deferred, except in exceptional cases involving serious social harm. The maximum administrative fine is KRW 30 million — roughly USD 21,000. The ministry describes its approach as minimum regulation. Human oversight is one factor among several in the high-impact determination, and in some sectors not a factor at all. Security and national defence are carved out.
In China, the gate is real and the refusal is real. What passes through it is not. An examination of the public filings found the descriptions pitched at so high a level of generality as to be almost devoid of meaningful detail — the filing for Weibo’s trending-search feature describes it as combining search, discussion, and dissemination popularity multiplied by an interaction coefficient, which an observer knowing nothing about the algorithm could substantially have guessed. The regulator may lack the in-house technical expertise to interpret what it receives. And the citizen’s entitlement, at the end of it, is a registration number displayed on a page. Not a reconstruction. Not a contest. The gate confers standing on the state.
So: Colorado removed the substance. The EU postponed it. Korea enacted it in a form that does not bind. China built the gate and admitted records that do not reconstruct.
| Jurisdiction | Instrument | Mechanism | What holds | What does not |
|---|---|---|---|---|
| Colorado | SB 189 (2026) | Repeal before entry into force | Notice, adverse-outcome disclosure, correction | Duty of care, risk management, impact assessment |
| EU | AI Act + Digital Omnibus (2026) | Deferral, 16 months | Article 50 transparency, on schedule | Annex III high-risk obligations, to Dec 2027 |
| South Korea | AI Basic Act (in force Jan 2026) | Qualifier-hollowing + non-enforcement | Notification and labelling duties | “Efforts to assess”; “technically feasible”; grace period; USD 21k cap |
| China | Algorithm Provisions (in force Mar 2022) | Binding gate, uninformative record | Filing, refusal with reasons, public number | Reconstructable content; any citizen route to contest |
Four legal cultures with nothing in common — a deregulating US state, a rights-based supranational union, a developmental East Asian democracy, a party-state. Four distinct mechanisms. One outcome, stated operationally: no person subject to a decision by any of these systems can reconstruct that decision and challenge it.
The danger in this correction
We must now say the thing that makes this paper worth publishing rather than embarrassing.
A thesis that predicted deferral, met a case of enactment, and responded by broadening itself to cover enactment-with-qualifiers is performing a textbook ad hoc rescue. It has purchased agreement with the data at the price of its content. The original claim forbade something specific: it said accountability obligations would not take effect. The revised claim, if we are careless, forbids nothing — any legal outcome whatsoever can be described after the fact as form-without-substance, because “substance” can always be defined as whatever did not happen.
This is the failure mode of every research programme that survives too long. We would rather name it than fall into it.
So two commitments, and they cost us.
First: the four cases confer no support on the revised thesis. They generated it. A hypothesis fitted to a sample cannot be confirmed by that sample. Whatever the table above shows, it is not evidence for what we now claim; it is the material from which the claim was constructed. The revised thesis enters the world with zero confirmations, and we will not cite Colorado, the EU, Korea, or China in its support again until it has survived a test it could have failed.
Second: the outcome must be defined operationally, in advance, so that it cannot be redefined afterwards. We define it as follows. A jurisdiction preserves decision-accountability if a person subject to a consequential automated decision can, as a matter of enforceable right: (a) obtain a record from which that specific decision can be reconstructed by an independent party, and (b) obtain a reconsideration capable of reversing the outcome. Both limbs, enforceable, in fact — not on paper. Anything less than both is form without substance. Anything meeting both refutes us.
That definition is narrower than our rhetoric has been. It excludes things we have written as though they counted. It is meant to.
What would refute us
The Institute holds foresight to a falsifiability standard. It has not held its own empirical thesis to one. That ends here. Four dated predictions follow. Each is risky: each could fail, each is checkable by anyone, and any one of them failing counts against us.
1. European Union — deadline of 2 December 2027. The deferred Annex III high-risk obligations fall due. We predict that either the date moves again, or the obligations take effect and no substantive enforcement action grounded in a decision-accountability obligation — risk management, human oversight, fundamental-rights impact — is concluded within twelve months of that date. Refuted if: the date holds and such an action concludes by 2 December 2028.
2. South Korea — grace period ending, approximately January 2027. We predict that the first administrative measures under the AI Basic Act are grounded in transparency obligations — notification, labelling — rather than in failures of risk management, explanation, or human oversight. Refuted if: the first concluded enforcement action rests on a high-impact decision-accountability duty.
3. China — algorithm registry, rolling. We predict that filings will continue to be accepted at a level of generality from which no specific decision could be reconstructed, and that no individual route will be created by which a person may obtain the basis of a decision affecting them. Refuted if: the CAC requires filings sufficient for an independent party to reconstruct a specific automated decision, or creates such an individual right of access.
4. Colorado — Attorney General rulemaking, due 1 January 2027. We predict that “meaningful human review” will be defined such that an official’s concurrence with an automated determination, without recorded reasons and without demonstrated capacity to reach a different outcome, satisfies the requirement. Refuted if: the rules require recorded engagement or evidenced override capacity.
If three or four of these hold, the thesis has survived a real test. If two or more fail, the thesis is wrong and we will say so on this page.
The competing explanation, honestly held
There is a rival hypothesis that accounts for the same four cases, and it is not obviously worse than ours.
Capacity. Decision-accountability obligations are hard. Reconstructable records, functioning override, independent audit — these require standards, conformity infrastructure, technical expertise inside regulators, and trained officials. Disclosure obligations are easy: a label, a notice, a number on a page. On this account, disclosure holds because it can be delivered, and decision-accountability lags because it cannot yet. The EU deferred because the standards had not materialised. Korea qualified its duties because operators cannot yet meet them. The CAC accepts thin filings because it lacks the expertise to demand better. Colorado is the outlier, explained by ordinary politics.
We do not dismiss this. It explains a great deal, and our own EU brief conceded it. But the two hypotheses come apart on a single question, and it is the question our predictions are designed to force: where capacity exists, does substance follow?
The Chinese state is not short of leverage over its platform companies, nor of engineers. It could compel filings that reconstruct. It does not. The EU will, by December 2027, have had the standards infrastructure it said it lacked. Korea’s grace period is explicitly a period for building capacity. In each case there is a date after which the capacity excuse expires. If substance arrives when capacity arrives, capacity was the explanation and we were wrong. If capacity arrives and the records stay thin, the deadlines move again, and enforcement continues to land on labelling, then something other than capacity is selecting which obligations survive.
That is a real test, and we may lose it. We would rather lose it publicly than hold a thesis that cannot.
What this correction does not settle
Three things remain open, and one of them is uncomfortable.
First, four cases is not a sample. Four cases chosen because they are legible to an English-speaking researcher is not even a random four. Japan, India, Brazil, the United Kingdom, the Council of Europe’s framework convention — none of these has been examined by us, and we should not be believed about a global pattern until they have been.
Second, the Chinese case raises a question our method has not answered and must: admissibility is not inherently protective. A pre-deployment, record-based gate can confer standing on a citizen or on a state, and nothing in the structure of the gate determines which. The Institute has argued for gates. We have not argued for whose gate, and the difference is the whole of the matter.
Third, and most uncomfortably: this correction was prompted by evidence we went looking for. Nobody else forced it. That is to our credit and it is also the problem — a single-author institute checking its own thesis is not peer review, and this paper does not become peer review by saying so.
FAQ
Is the Institute saying its central thesis was wrong?
Yes, as stated. We claimed decision-accountability obligations get deferred or repealed while disclosure survives. South Korea enacted them on schedule and China has operated a binding pre-deployment gate since 2022. The mechanism claim is refuted. What survives is a claim about outcome: across four jurisdictions with four different mechanisms, no person subject to an automated decision can reconstruct and contest it.
Isn’t broadening the thesis to fit the data exactly the move you criticise in others?
It is the danger, and it is why this paper defines the outcome operationally in advance, refuses to count the four cases as evidence for the revised claim, and states four dated predictions that could refute it. A restatement without those three things would be an ad hoc rescue. With them, it is a hypothesis.
What single finding would most damage the thesis?
A jurisdiction where a person subject to a consequential automated decision can, as an enforceable right, obtain a record permitting independent reconstruction of that decision and obtain a reconsideration capable of reversing it. We know of none. One would be enough.
Doesn’t the capacity explanation account for everything you observe?
It accounts for much of it. The two explanations diverge on whether substance follows once capacity arrives. Our predictions are timed to the dates after which the capacity explanation expires. If substance arrives with capacity, we were wrong.
