01 · Behaviour follows what is rewarded
Before any rule is broken, competitive pressure, deadlines, funding and performance targets have already decided what is rational. A team rewarded for speed will ship before it evaluates, whatever its principles say. Start any analysis by asking what each actor is paid, praised or punished for.
interpretationConfidence: mediumContestability: medium
Many governance failures begin before any rule is broken: competitive pressure, funding, procurement, liability exposure and performance metrics make some behavior rational regardless of stated principles.
Our reading: The framework memo's position. The UN report supplies the race-to-the-bottom concern at the international level; NIST names incentive structures and culture as conditions for effective risk management (section 1.2.4).
Inspect evidence · 2 passages
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
Governing AI for Humanity: Final Report (opens in a new tab)Executive summary, para. xvi (p. 8)
Governance regimes are said to need global span to avert AI arms races or a race to the bottom on safety and rights, to detect and respond to incidents spanning multiple jurisdictions, and to share AI’s benefits.
Scope note: An advisory body report representing a majority consensus of its members in their personal capacity. It is not a UN member-state decision. Page locators refer to the PDF's printed page numbers.
Passage checked 2026-09-06 · un-hlab-race-to-bottom
02 · Purchasing terms carry safeguards into the market
A buyer cannot write a law, but it can write a contract. Monitoring rights, data-use limits and exit terms change what a vendor must accept to win business. Procurement is an incentive instrument: it works through what sellers want, not through what they are forbidden.
source-assertionConfidence: highContestability: low
US federal AI contracts must give agencies the right to keep monitoring system performance and should limit vendor lock-in, so purchasing terms carry safeguards into the market.
Inspect evidence · 1 passage
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
M-25-22: Driving Efficient Acquisition of Artificial Intelligence in Government (opens in a new tab)Section 4(d), ongoing testing and monitoring; Section 4(c)(ii), vendor lock-in
AI contracts must let the buying agency keep monitoring and evaluating the system’s performance, and should include terms such as knowledge transfer that reduce the risk of vendor lock-in.
Scope note: Rescinds and replaces M-24-18. Applies to federal executive agencies; excludes national security systems.
Passage checked 2026-10-09 · omb-m-25-22-monitoring
03 · Liability changes what gets written down
When a producer may have to pay for harm, it starts keeping the records that would let it show a product was not defective. Liability does not set design rules. It moves the cost of failure, and that moves testing, documentation and release decisions upstream.
synthesisConfidence: highContestability: low
From December 2026 the EU treats software, including AI systems, as a product for no-fault liability, which changes what producers must be able to show when harm occurs.
Our reading: The second clause is an editorial reading of the liability incentive; the directive itself states the scope.
Inspect evidence · 1 passage
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
Directive (EU) 2024/2853 on liability for defective products (opens in a new tab)Article 4(1) definition of product; Recital 13; Article 2(1)
Software counts as a product for no-fault liability, and the recitals name AI systems as software. The rules apply to products placed on the market after 9 December 2026.
Scope note: In force since December 2024; applies to products placed on the market after 9 December 2026, when it replaces Directive 85/374/EEC.
Passage checked 2026-10-09 · pld-software-product
04 · A term is only as strong as the buyer's capacity
A monitoring clause means little if the agency has no one able to read the logs, no budget to test, and no credible option to switch vendors. Ask whether the institution holding the contract can use the rights it wrote into it. This is where incentives meet institutions.
source-assertionConfidence: highContestability: low
Representation and coordination in governance are not enough; commitments require implementation and capacity to produce tangible outcomes.
Inspect evidence · 1 passage
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
Governing AI for Humanity: Final Report (opens in a new tab)Executive summary, para. xviii (p. 8)
Representation and coordination are said not to be enough: accountability requires implementation so that commitments translate into tangible outcomes, including capacity development, much of it at national and regional levels.
Scope note: An advisory body report representing a majority consensus of its members in their personal capacity. It is not a UN member-state decision. Page locators refer to the PDF's printed page numbers.
Passage checked 2026-09-06 · un-hlab-implementation
05 · Worked example · fictional
A city buys a benefits-triage tool
Fictional exercise. A mid-sized city plans to buy an AI tool that ranks benefit applications for manual review. The vendor offers a low price if the city accepts its standard terms. The procurement team has six weeks and no data scientist. No jurisdiction or legal classification is assumed.
1. What is each party rewarded for?
The vendor is rewarded for closing the sale and limiting obligations. The procurement team is rewarded for meeting the deadline and the budget. Nobody in the deal is yet rewarded for finding errors that harm applicants.
2. Which terms would change the vendor's incentives?
Rights to monitor performance on the city's own applicants, limits on reusing city data, and a way to exit without losing records all make careful behaviour part of winning the contract.
3. Who pays if the ranking is wrong?
If wrongly deprioritised applicants have no claim against anyone, nobody in the chain has a reason to measure that harm. Ask what liability, if any, reaches the vendor or the city.
4. Can the city use the rights it buys?
A monitoring right needs someone to monitor. Without evaluation capacity, the term is a promise on paper. Budget for the capacity, or narrow the deployment until it exists.
5. What happens at renewal?
After three years of use, switching costs are high and the vendor knows it. Terms negotiated now decide whether renewal is still a choice later.
The cheapest contract buys the vendor's incentives along with its software. Terms on monitoring, data and exit move part of the safeguard into the market, but only if the city has the capacity to use them and the option to walk away.
What this lesson does not settle
This draft lesson has not yet been reviewed by the framework's author or an outside reviewer. It uses one US federal procurement memo and one EU liability directive as examples; neither describes procurement or liability law elsewhere, and neither is a legal analysis. Evidence on how well procurement terms change vendor behaviour in practice is thin and needs further research.