The Informed Decision Standard assumes a decision-maker who can still make an informed decision. AI dependence can quietly erode that capacity, until the accountable person is ratifying outputs rather than judging them.
The cost is not only a defensive one. An institution that lets its judgment atrophy loses the very thing that lets it move quickly, recognize what its models were not built to see, and sustain its own excellence.
The task is to sustain the capacity to decide as deliberately as the institution adopts the tools that can erode it.
A conversation is gathering under the heading of cognitive sovereignty: the worry that as AI saturates the work of thinking, people lose the capacity for independent judgment. Much of it arrives as a claim about individual rights, and some of it arrives as alarm. Set both aside. Strip away the rights language and the dependence anxiety, and what remains is a governance problem, and it sits squarely with the board. The question is not whether AI will make errors. It is whether the people accountable for a decision can still make it.
This is not a new subject for oversight governance. It is the human half of a standard the Center has already stated. An oversight program exists to put the institution in a position to make informed decisions about the balance of risk and reward. That standard has always presumed something it never had to say aloud: that a person capable of being informed, and capable of deciding, is on the other end of it. AI is what makes the presumption worth examining.
The Standard Assumes a Decision-Maker
The output of a program is the capability to make informed decisions. That capability has two halves. The first is the system the program builds: the charter, the tolerances, the escalation paths, the evidentiary record. The second is the human judgment that operates the system, the capacity of an accountable person to read the evidence, weigh it, and decide. Governance has spent its attention on the first half and quietly assumed the second. In a world where analysis was done by people, the assumption was safe. It is no longer automatic.
The Informed Decision Standard is not met by a decision that was produced. It is met by a decision that was made. The distinction is invisible on paper, because the output looks identical either way. A recommendation was generated, a person signed, the record shows a decision. Whether a judgment actually occurred, or a machine output was passed through a human who could no longer evaluate it, does not show up in the file. That is precisely why it is a governance risk rather than a documentation one.
How the Capacity Erodes
The failure mode here is not dramatic. No single decision goes visibly wrong. What happens is slower and harder to see: the accountable person defers a little more each cycle, until the deference is total and the capacity to do otherwise has thinned out. The patterns are recognizable once named. Accepting an output without examining the assumptions beneath it. Treating a summary as though it were the decision. Answering for a risk by pointing at the system that surfaced it, on the reasoning that the model said so.
Each of these is a small surrender, and none of them registers as one. The person is busy, the output is fluent, the tool has been right before. Over enough repetitions, the muscle that would question the output atrophies, and the institution is left with a decision-maker who can approve but can no longer challenge. This is the same shape as the error the Center has already named at the level of structure. The Committee Fallacy is the mistake of treating a committee that meets as a board that governs. Its counterpart at the level of the decision is a signatory who signs as though that were the same as deciding.
A decision no one can any longer challenge is not an informed decision. It is a ratified output with a signature attached.
The Cost Is Excellence, Not Only Exposure
The stakes here are affirmative before they are defensive, and reading them the other way misses the point. An institution's real asset is its accumulated judgment: the expertise that lets its people see what a model was not trained to see, weigh a situation no dataset anticipated, and decide well when the ground is unfamiliar. When the thinking is handed to the system a little more each cycle, that asset thins. The organization does not become unsafe first. It becomes less capable first, and slower to recognize the threat or the opportunity that sits outside the model's experience.
The same erosion undercuts speed, which is the capability the program exists to produce. An institution whose people can no longer independently evaluate an output cannot actually move faster. It can produce decisions faster, which is a different thing, and it stalls the moment something arrives that the model did not anticipate, because no one is left who can supply the judgment the machine cannot. Decision Velocity was never raw speed. It was the capacity to decide well, quickly, and to stand behind the result. That capacity is exactly what atrophies when judgment is outsourced.
Culture is what holds the line. An institution that still expects a person to explain why, and not only what, that treats a challenge to an AI output as normal rather than obstructive, keeps its judgment alive by using it. There is a defensive dimension as well. A board that ratifies what it cannot evaluate is not exercising the oversight its duty of care requires, and the reasoning that the model said so is the modern form of the passive acceptance courts have long treated as a red flag. But that is the floor. The reason to sustain the capacity to decide is not to stay out of trouble. It is to stay excellent.
AI Amplifies the Room It Enters
The erosion is not only individual. A board is a group, and in a group AI does something subtler than deskilling. It amplifies the dynamics already present. When legacy, identity, and strategy are entangled, and in most institutions they are, a system trained on the organization's own history and fed by its most influential voices tends to return those voices to the room with the authority of data behind them. A preference becomes a finding. The bias of whoever holds the most sway is the bias most likely to be reinforced, and it now arrives wearing the model's apparent neutrality.
The casualty is dissent, at precisely the moment dissent matters most. Succession, transformation, and capital allocation are where challenge is most needed and where it is already hardest, because the questions are personal and disagreement can be experienced as disloyalty. If the model quietly takes the side of the incumbent view, the last independent check thins out. Designing against that is part of what a board owes its own decisions: the behavioral conditions under which a difficult conversation can stay objective enough to challenge without becoming personal enough to fracture trust. In practice that means separating identity from the strategy on the table, protecting the standing of the person whose role is to disagree, and keeping the provenance of every input visible, so the room always knows which voice is the model's and which is its own.
Sustaining the Judgment
If the capacity to decide is part of what makes an institution excellent, then it has to be sustained on purpose, and not left to survive the arrival of tools designed to do the thinking. That does not call for a new right, a new officer, or a new committee. It calls for the program to hold the judgment around AI to the same expectations it already holds the system to: that it be real, that it be exercised, and that it be kept in use. Five practices keep the capacity alive rather than assumed.
Independent basis before consequential decisions. For a decision of consequence that rests on an AI recommendation, the accountable person should be able to state the basis for it independently, not merely repeat what the system produced.
An override that is real. The authority to reject the recommendation has to exist, and the program should be able to show that it has actually been used. An override that is never exercised is indistinguishable from one that does not exist.
Decisions examined, not deferred. The record should show that a consequential output was challenged before it was accepted. The model said so is not a decision. It is the absence of one.
Attribution and protected dissent. Keep visible which inputs came from the model and which from people, and protect the standing of whoever is charged with disagreeing. An AI output laundered into consensus, and dissent read as disloyalty, are how a room stops thinking together.
Accountability in the roles that already hold it. The duty sits with the decision-maker and the board. Answering an erosion of judgment by creating a chief officer for it repeats the Committee Fallacy in a new costume. The fix is to hold the existing roles to the existing standard.
None of this is a brake on AI, and none of it is a retreat from speed. It is the condition under which an institution stays both fast and excellent. One that has kept its people capable of deciding can adopt AI aggressively, move at the cadence the environment demands, and still recognize the moment the model is wrong. One that has let that capacity thin out is moving quickly toward decisions it no longer understands, and toward a future in which its own knowledge has quietly left the building. The point of sustaining the judgment is the point of the program itself: to keep the institution able to make the call, to make it well, and to answer for it.
For our most consequential AI-assisted decisions, can the accountable person state the basis independently, or only repeat what the system produced?
Does our override authority get used, or is it nominal? Can we show an instance where a recommendation was rejected?
When something appears that our models were not built to recognize, do we still have the judgment in-house to see it and act on it?
Are we tempted to answer this with a new officer or committee, when the duty already sits with the decision-maker and this board?