TSP has read eleven AI governance pledges issued by public companies over the past year. The language is almost interchangeable. Committed to responsible AI. Committed to human oversight. Committed to transparency. Nine of the eleven drew favorable coverage in the business press within a week of release. Only two named, anywhere in the document, the specific person with the authority to halt the system they were describing.
That is not an editing oversight. It is the pattern itself.
A pledge costs a communications team an afternoon. Naming a person who can stop a revenue-generating system costs something real: a line of authority that did not exist before, a delay that someone will resent, a conversation about what happens to that person's standing the first time they use the authority and turn out to be wrong. Boards will approve the first cost without discussion. The second cost requires a decision, and decisions can be deferred in a way that drafting cannot.
The two companies that named a stop-authority holder did something else worth noticing. Neither treated the naming as the finish line. One required the named executive to report, in writing, every quarter on whether the authority had been exercised and why or why not. The other built a standing exception log that the board reviewed before any other agenda item, specifically so a quarter with zero exceptions could not be mistaken for a quarter with nothing worth watching. Both boards had clearly asked a harder question before publishing anything. Not what should we say we believe, but who in this building can actually make the system stop, and how will we know if they ever had to.
TOUCH STONE LAW: The Law of the Hollow Pledge A commitment that names no one who can act on it is not governance. It is a press release wearing governance's clothes, and the first system failure will tear the costume off in public.
TSP has a name for the pattern underneath the nine pledges that stopped at language. It is the Declarative Board Failure Pattern: a board declares a commitment, points to the declaration itself as evidence of diligence, and builds nothing underneath it. The pattern survives because declaration and governance produce the same press release. They do not produce the same Tuesday morning when a model does something nobody anticipated and someone has to decide, in real time, whether anyone has the standing to pull the plug.
Directors rarely discover this gap by reading the pledge. They discover it by asking one question in the room and watching how long the silence lasts. If your AI system produced a materially wrong decision tonight, who has the authority to take it offline before market open, and does that person need anyone else's approval first. A board that has actually built the architecture answers in one sentence, with a name in it. A board that has only declared the commitment produces a longer answer, usually beginning with "it would depend," which is the sound of authority that was never assigned.
This is not a technology question, and treating it as one is how boards excuse themselves from it. The AI system is incidental. The same silence would follow the same question asked about a rogue trading algorithm, a product recall, or a plant manager who will not stop a line he knows is unsafe. Companies that have already answered that question for one domain tend to answer it quickly for AI, because the muscle already exists. Companies that have never been forced to name a stop-authority holder for anything discover, usually during an actual incident, that the muscle was never built.
The directors who get this right do not treat the naming as a one-time event either. Authority that is granted once and never exercised, reported on, or revisited becomes ceremonial within a year, no different from the pledge it was meant to replace. The two companies above built the review into the calendar precisely because they understood that a stop-authority holder nobody has checked on is drifting back toward decoration.
What a board leaves behind is not the pledge it published. It is whether the next director, five years from now, can still name the person who holds the authority to stop a system that is moving faster than the room can follow, without opening a document to check. A pledge gets printed and framed. A governance architecture gets inherited and used. Only one of those outlasts the people who built it.