Many organisations using AI now say they do so responsibly. The sentence appears on websites, in proposals and in policies, and it is nearly always true in the sense that nobody intends otherwise. The problem is that it cannot be checked. There is nothing in "we use AI responsibly" that a client, an auditor or a member of staff could hold up against what actually happened and say: this was or was not done.
That is the case for a written standard over a claim. A standard says what we do, in terms specific enough to be tested. A claim says what we are, and leaves the reader to take it on trust.
What makes a standard checkable?
Three things separate a standard from a statement of values.
It names actions, not qualities. "We are careful with data" is a quality. "No client or personal data is pasted into an AI tool" is an action, and anyone who watches the work can see whether it happens.
It says when each action applies. A rule that applies "where appropriate" applies whenever the person following it finds it convenient. Ours is built around three points in the work: vet, disclose, verify. Vet before a tool is used or recommended: its data policy, its fit and its cost. Disclose wherever silence would mislead someone who trusts us. Verify before anything reaches a client or goes public: the model drafts, a named person decides, and every fact is checked.
It can fail. A standard that cannot be failed is not a standard. If no conceivable piece of work would breach it, it is describing an attitude, not constraining behaviour.
Why publish it rather than keep it internal?
An internal standard is checked by the people it is meant to constrain. That is better than nothing and weaker than it sounds, because the people closest to the work are also the people most able to explain away a gap. Publishing it changes who can check. A client can ask whether a step was followed. A reader can point out where our own output falls short. The standard becomes something we can be held to by people who have no reason to go easy on us.
It also changes how we behave when it fails. We ran our own audit against the standard and published the result: it failed. Publishing that was uncomfortable, and it was the point. A standard you only publish when you pass is marketing. One you publish when you fail is evidence that it is being used.
The same reasoning sits behind the gate we run before any AI touches a real person. Written down and public, it can be argued with, improved and applied by someone other than us.
What it gives the people who rely on us?
A published standard gives a client a list of questions they are entitled to ask, and a reasonable expectation of the answers. It replaces "trust us" with "check us", which is a harder offer to make and a more useful one to receive. We call the result governed performance: objectives delivered reliably, uncertainty managed openly, integrity kept throughout.
If you are writing your own
Take your organisation's current statement about responsible AI and try to find one action in it that someone outside could verify. If there is none, that is the gap. Start with three actions you already take, write down when each applies, and publish them where the people who rely on you can see them.
If you would like a second view on a draft standard: [email protected]
What "governed AI" actually means: the standard itself.
We ran our own governance audit on ourselves. It failed: the standard tested on our own setup, and where it failed.
The 15-point gate we run before any AI touches a real person: the standard turned into a pre-launch check.
TL;DR: "We use AI responsibly" cannot be checked; a written standard can. A checkable standard names actions rather than qualities, says when each applies, and is capable of failing. Publishing it lets clients and readers hold you to it, and publishing the times it fails is what shows it is real. Ours is vet, disclose, verify.
Not legal, security or financial advice. Governed performance.