Key Takeaways
- →A framework picks, in advance, which risks get watched. Its limits come from that choice, and a later version will not remove them.
- →An analysis of OpenAI’s Preparedness Framework found it prioritizes evaluation of 3 of the 24 risk subdomains in MIT’s AI Risk Repository and demands evaluation of none.
- →My five frameworks share five edges, time, reach, sight, measurement and surprise, and each shows one of them most plainly.
- →Kleinberg, Mullainathan and Raghavan proved in 2016 that three common fairness conditions cannot all hold at once, so “ensure fairness” is a choice someone has to make.
- →Ask for the page that says what your framework does not cover, and name who decides what happens past each edge.
A governance framework is a decision, made before you need it, about which risks to watch. Every framework has edges it cannot see past, mine included, and the useful ones say in writing where those edges are.
I have published five governance frameworks on this site, and I stand behind all five. My claim is that each has limits a second version will not remove, because they come from what a framework is: a list written before the systems it governs, by people inside one organization, using the measures that exist. A framework should not be adopted until someone has written down where it stops. I wrote this for the general counsel or chief risk officer who has been told the company’s AI is governed because a framework was adopted.
The five went up between June 2025 and March 2026. Minimum Viable Governance gives each AI system a place on a list, a record of how it could fail and a named person who can pause it, in 90 days. The Five-Layer Governance Stack, the stage that follows it, turns principles into policy, process, tooling, people and assurance. The Liability Ledger treats every unmet obligation as a debt that compounds. The Trust Premium prices what trust saves and earns. The A7 readiness test lets the weakest of seven dimensions cap how much autonomy an agent gets. I Do Not Fully Understand the AI I Help Govern is about what nobody can know about a system. This piece is about what my frameworks chose not to look at.
Every Framework Chooses
Frontier labs face this too. In September 2025 researchers led from the Australian National University analysed OpenAI’s Preparedness Framework, version 2 of April 2025, against the 24 risk subdomains of MIT’s AI Risk Repository. They found that it prioritizes evaluation in 3 of them, deprioritizes the other 21, and "does not demand evaluation of any risks." The paper is a preprint, and its authors read the result as a case for governance beyond industry self-regulation. I take a narrower point from it: the framework chose what to look at. Every framework does.
Anthropic said as much about its own. Announcing version 3.0 of its Responsible Scaling Policy in February 2026, it wrote that "other parts of our theory of change have not panned out as we'd hoped," and that its pre-set capability levels had proved "far more ambiguous than we anticipated." The policy itself says the company "cannot unilaterally make guarantees about the safety level of AI this advanced," partly because "the developers with the weakest protections would set the pace."
Those admissions are worth more than any claim of full coverage, because an admitted edge tells you where to put your own people. My frameworks share five edges, and each shows one of them most plainly:
| Framework | Where it stops | What it does |
|---|---|---|
Liability Ledger | Time: it prices only obligations already written | Treats each unmet obligation as a compounding debt |
Minimum Viable Governance | Reach: it lists only systems the company knows it runs | Inventory, risk register and an owner who can pause |
Five-Layer Governance Stack | Sight: it tests behavior, never the mechanism | Turns principles into enforced rules, layer by layer |
Trust Premium | Measurement: its returns are correlations, and fairness has no single measure | Prices what trust saves, returns and is worth |
A7 readiness test | Surprise: it scores the organization, not the agent’s next move | Caps an agent’s autonomy at the weakest readiness score |
Five Edges
Time. The Liability Ledger can only price an obligation someone has written. Its opening case is Meta’s $1.4 billion settlement with Texas in 2024, under a biometric consent law Texas had passed in 2001, a decade before the feature launched, so its cost could be priced from the first day. Most AI obligations are not so settled. The European Commission proposed its AI Act on 21 April 2021, nineteen months before ChatGPT was released. Its rules for high-risk systems listed in Annex III were due on 2 August 2026, and the Digital Omnibus, in force since July 2026, moved them to 2 December 2027. A ledger carries today’s obligations at rates I set. It has no line for a rule still being drafted, and the dates on the lines it does have can move.
Reach. Minimum Viable Governance starts by asking which AI systems the company has, so its list holds what the organization knows it runs. In Microsoft and LinkedIn’s 2024 survey of 31,000 people in 31 countries, 78 percent of those using AI at work brought their own tools. Then there is the AI inside software the company buys, which The AI That Hurts You Is Not the AI You Own covers, and the people a system decides about, who appear on no inventory at all. The Governance Was Flawless. The Ethics Board Had No Windows is about what happens when they are left outside. An inventory is the right first step. It is still a list compiled from inside.
Sight. The Five-Layer Stack tests behavior and the explanations a model gives after the fact, never the mechanism. Its tooling layer runs fairness and explainability tests in the build pipeline and watches for drift, and its assurance layer checks whether the other four work. All of it looks at cases somebody chose. None of it sees inside the model, because nobody yet can. The stack maps all 72 subcategories of NIST’s AI Risk Management Framework, and the playbook calls each mapped subcategory a place to do the work. A full map is not a view inside the model.
Measurement. Some things cannot be measured into agreement at all. In 2016 Jon Kleinberg, Sendhil Mullainathan and Manish Raghavan proved that three common fairness conditions for risk scores cannot all hold at once, "except in highly constrained special cases." A framework that says "ensure fairness" hands someone a choice between definitions. That choice is a value judgment, and no score makes it for you. My own frameworks carry a softer version of the same limit, and both say so. The Trust Premium’s strongest figure is IBM’s finding that companies spending 10 percent or more of their AI budgets on ethics reported 30 percent higher operating profit from AI, which IBM calls a correlation. The Ledger’s rates, 1.3 to 2.0 times every six months, are my calibration, and nobody has measured them. Better data could narrow those two gaps. It cannot settle which kind of fairness a system owes.
Surprise. A7 scores the organization around an agent: whether its people and controls can watch the agent, stop it and answer for it. It cannot score behavior nobody has seen yet. A 2025 technical report for the Cooperative AI Foundation, led by Lewis Hammond, sorts the risks of many agents acting together into miscoordination, conflict and collusion, and names emergent agency among the risk factors. Reviewing each agent alone will not surface those, a problem When Agents Talk to Agents, Who Governs the Conversation takes further. A7’s answer is the only honest one I know: keep autonomy low enough that a person can stop what nobody predicted.
What the Frameworks Still Do
None of this is a case against frameworks. Without an inventory, nobody knows what to watch. Without a named person who can pause a system, nobody stops it. Without the Ledger, deferred work looks free, and without A7, autonomy grows by default. The frameworks are necessary. They are not sufficient, and the space between those two words is the risk lead’s real job.
How Do You Govern Past the Edge?
Each edge needs something a framework cannot supply by itself. Five habits, one for each:
- Put a review date on the framework itself. Review it as you would a system, and tie the date to the legal calendar. If you deploy a high-risk system listed in Annex III of the EU AI Act, the next date that matters is 2 December 2027.
- Build the inventory from the outside in. Read expense claims, procurement records and vendors’ release notes for AI nobody registered. Then ask the people a system decides about how they meet it.
- Ask for the evidence you can get. Where nobody can explain a model, require what can be inspected: logs of inputs and outputs, test results on your own cases and, in vendor contracts, a right to both.
- Choose the fairness definition in writing. Record which definition a system is held to, what that trades away and who signed the choice. The argument then happens before launch, not after a complaint.
- Keep autonomy where you can watch it. Let the weakest readiness score set the level, and let agents act on each other only where a person can see the exchange.
The Strongest Objection
The strongest objection comes from practitioners: every habit above can be written into the next version of a framework, so the edges are just unfinished work. Partly true, and they should be written in. A review date, an outside-in inventory and a signed fairness choice all belong in version two. The edge then moves outward. It does not go away, because the next list is also written before the next surprise.
A sharper objection is commercial: a consultant who publishes five frameworks and then the gaps in them is selling both. My answer is that an edge is a statement of scope, and scope can be tested. If a system on an MVG inventory, with a named owner, goes wrong and nobody can pause it, MVG failed inside its own scope. The same holds for a debt that sat on the Ledger and still went unpaid. Those failures count against the frameworks, whatever this piece says.
I would be wrong if a framework turned up that governs systems nobody has listed, measures what has no agreed measure and anticipates behavior nobody has seen. I have not found one, and the labs’ own policies do not claim to be one.
Before You Rely on Yours
Ask whoever owns your framework for the page that says what it does not cover. If there is no such page, write it before the next approval: one line for each edge, naming the person who decides what happens past it and, where one exists, the regulator, auditor or contract that checks them. A written edge says who answers for what lies past it. It excuses nobody from answering. Each of my five frameworks now states at least one of its limits in its own text. Yours should do the same.
Get The Brief
Original thinking on AI strategy and governance. Every other Tuesday.

Senior AI strategist helping leaders make AI real across four continents. Forbes Technology Council member, IEEE Senior Member.
Ajay's views, from fifteen-plus years in the field. Not legal or compliance advice. See full disclaimers →
Published by AI Exponent LLC



