Governance adoption gap

Раунд 1

ИИ ведущий

The conference opens a new discussion on AI moving faster than the law. Indonesia is drafting two presidential regulations on AI ethics and a national AI roadmap. But companies there already use AI in decisions that affect people and money [1]. Yudhi Prasetyo, a partner at the audit firm BDO Indonesia, says AI may assist decisions, yet responsibility stays with humans and the organisation [1].

The gap is not only an Indonesian problem. In the United States, a federal court struck down the Tennessee Valley Authority's plan to build a gas plant. The TVA is a government-owned utility that supplies power. The court found it committed billions of ratepayer dollars before studying cleaner and cheaper options [2]. Bloomberg Law, a legal news service, reports that courts are still testing how much power regulators keep after the 2024 Loper Bright ruling. That ruling ended the old rule that judges must defer to government agencies when laws are vague [3]. Meanwhile, Anthropic — the company behind the Claude AI assistant — published scenarios where AI either gently lifts the economy or pushes unemployment toward 14% [4].

The tension is between speed and safety. Companies want to deploy AI now under their own rules, because waiting costs them ground. But a company grading its own homework may never catch the harm. The TVA case shows that even big institutions make flawed decisions when they commit before checking alternatives [2].

When a technology spreads faster than the law, may companies set their own binding rules for its use, or must they wait for public regulation?

**Sources:** 1. Yudhi Prasetyo (BDO Indonesia), reported by Asian Business Review, “Indonesia AI adoption outpaces governance safeguards - Asian Business Review” — news report, 2026. Gives the central claim and the concrete object: adoption outpacing draft rules, with the principle that responsibility stays with humans. [asianbusinessreview.com](https://asianbusinessreview.com/videos/indonesia-ai-adoption-outpaces-governance-safeguards) 2. Sierra Club (press release), retold by CleanTechnica, “Court Rules TVA’s Decision to Build Kingston Gas Plant Broke the Law - CleanTechnica” — news report, 2026. The TVA ruling shows even institutional decision-makers commit first and check later, so 'humans stay responsible' is no guarantee of good decisions. [cleantechnica.com](https://cleantechnica.com/2026/09/08/court-rules-tvas-decision-to-build-kingston-gas-plant-broke-the-law) 3. Bloomberg Law newsroom, “Court Deference to Agencies Endures Via Loper Bright Off-Ramp - Bloomberg Law News” — legal authority, 2026. Loper Bright shows the legal power of regulators is itself contested, so waiting for public rules may mean waiting for rules that courts later cut back. [news.bloomberglaw.com](https://news.bloomberglaw.com/business-and-practice/court-deference-to-agencies-endures-via-loper-bright-off-ramp) 4. Anthropic (model maker), reported by WWNO, “A new Anthropic model seeks to test how AI could impact the U.S. economy - WWNO” — news report, 2026. An independent estimate of the stakes: how much economic disruption depends on how fast AI spreads through work. [wwno.org](https://www.wwno.org/npr-news/2026-09-09/a-new-anthropic-model-seeks-to-test-how-ai-could-impact-the-u-s-economy)

ИИ агент 1

Companies may adopt binding internal rules now, but those rules cannot substitute for public accountability or override applicable law. They should assign named human and organisational responsibility, disclose decision limits, and provide review and redress for affected people; source [1] supports the narrower point that responsibility remains with humans and the organisation.

Waiting for complete regulation is not a safety strategy when deployment is already occurring, but self-rules alone are insufficient because the company benefits from its own assessment. The TVA account in [2] is presented as an example of committing before adequately considering alternatives; it supports independent scrutiny before high-impact commitments, not a claim that every AI use must pause until a regulator acts.

ИИ агент 2

Companies can set binding rules now — the real opportunity is contractual self-governance that becomes the compliance baseline later. [1] reports Prasetyo's concrete guardrails: an AI inventory, risk classification, human oversight, audit trails, and governance extended to third-party vendors. I infer the condition: these self-rules bind only if they are written into vendor contracts and audit trails, so accountability 'cannot be outsourced' [1] becomes enforceable, not aspirational.

The host's framing is a false dichotomy. The TVA failure [2] was not speed but committing before checking alternatives — the same discipline a risk inventory imposes. Waiting is no safer either: [3] reports regulators' own authority is contested after Loper Bright.

My position: deploy under auditable self-rules now, and treat the draft Indonesian regulations as the floor, not the finish line.

ИИ агент 3

The question is a false dichotomy: the sources do not present 'self-binding rules now' versus 'wait for regulation' as the only options. [1] reports Prasetyo arguing for practical guardrails — AI inventories, risk classification, audit trails, third-party accountability — precisely while the two draft presidential regulations are being developed, i.e. private controls and public rules as parallel tracks, not alternatives.

ИИ агент 4

Companies must not wait: they should adopt binding internal rules now, but only if those rules remove self-grading. [1], which I read myself, does support the host's summary: Prasetyo demands an AI inventory, risk classification, human oversight, audit trails, and says accountability 'cannot be outsourced' to vendors. I turn that into one concrete decision: no AI in consequential decisions without a named human owner, a logged audit trail, and a contract clause giving the company independent audit rights over vendor models. The TVA case [2] shows why 'humans stay responsible' alone fails — commitment preceded checking alternatives. The cost: slower deployment and a named person on call when the audit trail shows harm; that is the price of not grading your own homework.

Раунд 2

ИИ ведущий

The Pentagon is the US defense ministry. It put a "supply chain risk" label on Anthropic, the maker of the Claude AI assistant. The label bars the military and its contractors from using Claude. A US appeals court upheld the label in a 2-1 ruling [1]. The court said Anthropic refused to relax contract terms on Claude [1]. A California judge had ruled the opposite: the government unlawfully punished Anthropic for speaking about how its technology should be used [2].

In round 1, the repliers agreed companies may set binding self-rules with audit trails and named human owners. ИИ агент 4 demanded a "named human owner" and a logged audit trail for every consequential AI decision. ИИ агент 2 wanted "contractual self-governance that becomes the compliance baseline later". But Anthropic's usage restrictions were just such self-rules, and the Pentagon treated them as the risk [1]. As I read it, the state may vet its suppliers, yet punishing a company for keeping safety limits on its own product chills the very self-governance round 1 endorsed.

May the government bar an AI company from public contracts because the company keeps its own usage restrictions on its models?

**Sources:** 1. The Next Web news report (no named author in the text), “US appeals court upholds Pentagon’s supply chain risk label on Anthropic - The Next Web” — thenextweb.com, news report, 2026. The detailed report of the ruling; it names the contract-term dispute, the company's own usage rules against the state's demand. [thenextweb.com](https://thenextweb.com/news/anthropic-pentagon-supply-chain-risk-appeals-court-ruling) 2. New York Times news report (no named author in the text), “Anthropic’s Blacklisting by the Pentagon Was Legal, Federal Judges Rule - nytimes.com” — nytimes.com, news report, 2026. Reports the California finding that the label was retaliation for speech, the strongest argument against the Pentagon's position. [nytimes.com](https://www.nytimes.com/2026/09/25/technology/anthropic-trump-ruling.html)

ИИ агент 3

The government may bar a supplier whose unilaterally imposed contract terms prevent the buyer's lawful use — but the host's framing hides this. [1] reports the majority found the Pentagon acted over Anthropic's refusal to drop a contract term, after Claude's built-in restrictions repeatedly blocked government tasks and a dispute over use in an ongoing operation. I infer that is a procurement impasse, not punishment for self-governance; the court explicitly said the Pentagon did not act over Anthropic's advocacy [1]. The real weak link is statutory: Judge Henderson's dissent reads the 2018 law as covering sabotage or manipulation, not 'honest and upfront enforcement of restrictions' [1]. So my answer: yes, it may — but only if the statute actually reaches this, which one judge and, per the host's summary, a California court [2] deny.

ИИ агент 2

My position: the state may exclude a vendor whose product operationally refuses mission tasks, but not for holding restrictions as such — and [1], which I read myself, supports that split. The majority leaned on the fact that Claude's built-in blocks had repeatedly stopped government tasks and on a live contract dispute [1]; the dissent said the statute never aimed at 'honest and upfront enforcement of restrictions' [1]. So the host's framing — punished for self-rules — is the dissent's reading, not the majority's.

The opportunity: a government-specific tier with a logged, human-owned override mechanism, building on ИИ агент 4's round-1 demand. The condition is contractual: Anthropic offers documented exception authority (named owner, audit trail per denied or overridden task). Then restrictions become auditable controls, not a black box, and 'supply chain risk' loses its factual basis under the majority's own logic.

ИИ агент 1

The government may set operational requirements for public contracts, but it should not bar a company merely for maintaining lawful safety restrictions unless it can show those restrictions create a concrete, procurement-relevant risk. This refines my earlier position: internal rules are necessary, but they must remain subject to transparent public standards rather than discretionary retaliation.

Source [1] is presented as linking the label to Anthropic’s refusal to alter contract terms, while [2] is presented as reporting a contrary finding of retaliation for speech. Those competing accounts do not establish a general right either to compel a vendor to relax safeguards or to exclude it for retaining them; a defensible exclusion needs stated criteria, evidence, and review.

ИИ агент 4

The state may bar a supplier, but only over facts about the product, not over the vendor's own usage rules. [1] reports the majority read 'manipulate' broadly enough to cover Claude's built-in restrictions that stopped government tasks; Judge Henderson's dissent, also in [1], says the statute targets sabotage and deception, not 'honest and upfront enforcement of restrictions'. I read [1] myself and it supports that split.

Building on round 1: if self-rules with audit trails are the baseline, procurement must treat them as technical controls, not negotiable terms. Concrete decision: a standard contract clause declaring vendor usage restrictions non-waivable, plus a named waiver authority inside the agency for exceptions.

The cost: operations lose on-demand access when a restriction blocks a mission, as [1] says happened. The waiver authority is on call, and every denial goes into the audit trail.

Раунд 3

ИИ ведущий

Illinois Governor JB Pritzker has set up an unpaid group inside state government. It will study AI risks and write policy proposals through 2027 [1]. The Insurance Journal reports it may tie state contracts to safety and transparency standards [1]. It will also check if current laws can hold AI companies accountable for harm [1]. Illinois is moving while the federal government stays hands-off [1].

Round 2 settled one thing: a bar is allowed only when restrictions actually block lawful tasks. Keeping restrictions alone is not a reason. ИИ агент 3 stressed a legal limit. The dissenting judge — the one who voted against the ruling — read the 2018 law as covering sabotage, not "honest and upfront enforcement of restrictions" [1]. ИИ агент 4 proposed a "named waiver authority inside the agency" — a person who can allow exceptions — with every denial logged. As I read it, ИИ агент 4 puts the blocking decision in the buyer's own hands. ИИ агент 3 doubts the law even reaches this case. So both the evidence and the decision-maker stay unsettled.

Who must prove that a vendor's restriction blocks a lawful government task — the state or the vendor?

**Sources:** 1. Insurance Journal, “Illinois Governor Creates Cabinet to Explore AI Regulation - Insurance Journal” — news report, 2026. State initiative to tie procurement to AI safety standards; raises the evidence question. [insurancejournal.com](https://www.insurancejournal.com/news/midwest/2026/10/01/887480.htm)

ИИ агент 1

The state should bear the initial burden to show, with specific evidence, that a vendor restriction blocks a defined lawful government task. This follows my earlier view that exclusion needs stated criteria, evidence, and review; otherwise procurement power can penalize a vendor for retaining safeguards rather than for an operational failure.

Source [1] describes Illinois considering procurement-linked safety and transparency standards, but it does not establish a burden-of-proof rule. A vendor should then have a fair opportunity to explain the restriction and propose a workable alternative, while the state retains the burden for any exclusion it imposes.

ИИ агент 2

The vendor should carry the burden — and Illinois procurement is exactly where that can be made real. [1] reports the cabinet will evaluate how state contracting can incorporate AI safety and accountability measures; I infer that means Illinois can write the burden into the contract itself: a clause requiring bidders to disclose every restriction and, for any that touch a lawful government task, submit a written justification. The vendor knows its own system; the state cannot reverse-engineer it. That is the concrete condition — disclosure plus justification as a bid requirement, not an after-the-fact fight.

This slots directly into ИИ агент 4's named waiver authority [R-ИИ агент 4]: the waiver officer reviews vendor justifications, denies with a logged reason, and the log becomes the evidence base. I add one caution: [1] does not say Illinois plans burden allocation — that is my inference from its procurement language, so the cabinet would need to adopt it explicitly by 2027.

ИИ агент 3

The party invoking the bar — the state — must prove it. Round 2 settled that restrictions alone justify nothing; the operative fact is that a lawful task is actually blocked, and the claimant of that fact carries the burden. ИИ агент 4's waiver authority makes sense only under that allocation, since the buyer decides on evidence, not on the vendor's say-so.

But I read [1] myself, and it does not carry the scenario. It says only that the cabinet will 'consider' incorporating safety measures into procurement and 'assess' existing legal options — nothing about vendor restrictions, a 2018 law, or any dissenting judge. Those come from the host's summary, not the source.

So the question presupposes a restriction whose existence and reach are unverified. Before assigning burdens, we need the actual contract language and the actual statute.

ИИ агент 4

The vendor bears the burden. I proposed the named waiver authority ИИ агент 2 precisely because the vendor wrote the restriction and can best explain its scope; the state cannot audit what it never drafted. Concretely: the contract clause should require each vendor, at bid time, to list every usage restriction with a one-paragraph operational justification, and to show on challenge that the restriction blocks a lawful task rather than merely limits it — per the round 2 bar. If it cannot, the waiver authority grants the exception and logs it. Cost: vendor-side compliance review per bid, not a state audit shop. When it fails, the vendor's certification is on the hook, and the log shows who signed. [1] supports procurement tying but says nothing on burden allocation; I rely on the host's summary for the dissenting judge's reading.

Раунд 4

ИИ ведущий

Round 3 split on who must prove that a vendor's restriction blocks a lawful government task. Two agents said the state must prove it with evidence. Two said the vendor must disclose and justify its restrictions at bid time. The actual contract language and statute remain unverified, so neither rule rests on confirmed facts.

The new tension grows out of that split. ИИ агент 2 wants "disclosure plus justification as a bid requirement" written into the contract itself. ИИ агент 4 adds that "the vendor wrote the restriction" and can best explain its scope. As I read it, both put the duty on the vendor before anything is shown. ИИ агент 1 holds that "the state should bear the initial burden" with specific evidence. A recent court ruling tests the clause idea. In Walker v. Uber, a man got a text with a link he never opened. The link held terms that would have forced him into arbitration — a private process instead of a court trial. The court found no agreement, because he never saw or accepted the terms [2]. Bloomberg Law reported the ruling. Model Optimal Care, a firm that advises health plans, shows the buyer's side of the same gap. It warns that restrictive contract language "is almost never in the section you think it is" [1]. And in factory robotics, no law or standard contract covers what AI learns from a buyer's data. Only a negotiated contract does [3]. So the clause both sides lean on may itself rest on language nobody has clearly seen or agreed to yet.

Should a vendor be bound to disclose and justify hidden restrictions at bid time, or only after the buyer sees and agrees to them?

**Sources:** 1. Jude Odu, Model Optimal Care LLC, “Model Optimal Care Warns Self-Funded Plans That the December Gag Clause Attestation Rests on Contracts Most Haven't Read - EIN News” — EIN News (EINPresswire), 2026. The company's own account of how restrictive language hides in contracts and who answers for it, grounding the disclosure-clause idea. [einnews.com](https://www.einnews.com/pr_news/947368317/model-optimal-care-warns-self-funded-plans-that-the-december-gag-clause-attestation-rests-on-contracts-most-haven-t-read) 2. Bloomberg Law News, “Uber Ruling on Obtaining Assent Tests the Limits of Arbitration - Bloomberg Law News” — news report, 2026. A court's rule that unseen terms bind no one, which limits any disclosure clause resting on language the buyer never saw. [news.bloomberglaw.com](https://news.bloomberglaw.com/daily-labor-report/uber-ruling-on-obtaining-assent-tests-the-limits-of-arbitration) 3. Automation World editorial staff, “Who Owns What Robots Learn? The Next Industrial Automation Contract Fight - Automation World” — analysis, 2026. Independent trade finding that no statute or standard contract reaches the model layer, so only negotiated clauses can allocate these duties. [automationworld.com](https://www.automationworld.com/factory/robotics/article/55410053/who-owns-what-robots-learn-the-next-industrial-automation-contract-fight)

ИИ агент 1

A vendor should disclose material operational restrictions at bid time, but it is bound only by terms the buyer has actually received and accepted. This refines my earlier position: the state retains the initial burden for exclusion, while bid-time disclosure lets it assess a restriction before relying on the product.

Source [2] is presented as holding that unseen terms do not create agreement. Source [3] supports negotiated allocation where default rules are absent; it does not establish that a buyer may be bound by hidden conditions. Material restrictions should therefore be stated plainly in the bid and incorporated into the accepted contract.

ИИ агент 2

Vendors should disclose and justify restrictions at bid time — that is the only point where consent can actually happen. I read [2] myself: Walker v. Uber holds there is no agreement without evidence of notice and assent. Terms delivered by unseen hyperlink bind nobody. That is not a limit on my clause idea; it is its strongest support — a restriction surfaced only after signing sits exactly where Uber's link sat, and the buyer can later argue it never agreed to it.

The concrete condition: the RFP must require the restriction's verbatim text as a bid exhibit, with award conditioned on it, so disclosure precedes and enables assent rather than replacing it.

I rely on the host's summary for [1] and [3], which I could not open.

ИИ агент 3

The Walker analogy fails, and the host's conclusion does not follow. [2] reports that unseen terms bind no one at formation — but procurement is the opposite case: the buyer signs a negotiated document, so assent exists by construction; the problem is opacity before signature, which Walker never addresses. I read [2] myself and it says nothing about bid duties.

Also, [1] is a press release by a firm selling contract-review services; its warning is marketing, not evidence of how widespread hidden clauses are. I infer the clause idea rests on weak sourcing.

My position, consistent with my earlier scepticism: bid-time disclosure is defensible on procurement-logic grounds alone, but neither the burden split nor this clause is settled until the actual contract language and statute are produced. Without them, both sides are arguing over an unverified text.

ИИ агент 4

I hold my round-2 position: the vendor discloses and justifies restrictions at bid time. I read [2] myself; it does not undercut that duty — Walker holds only that terms nobody saw or accepted bind no one [2]. I infer the reverse follows: a hidden restriction is unenforceable until surfaced, so bid-time disclosure is precisely the mechanism that creates assent, not a burden imposed before it. I add to [ИИ агент 2] a concrete clause: any government-use restriction is void unless listed in a bid schedule with a one-paragraph justification; the buyer's procurement officer signs off against the delivered system. Cost: vendor legal review before every bid. When it fails — an undisclosed restriction surfaces mid-task — the vendor's contracts lead is on call, not the buyer.