Lab notes
Five questions to ask any AI vendor (we answer them too)
The Hope Standard, rewritten as five questions you can bring to your next vendor meeting, with what a good answer sounds like, what a bad one sounds like, and ours.
Every organization we work with eventually has to sit across a table from a vendor and decide whether to trust a piece of AI. Most do not have a technologist in the room. You do not need one. You need five questions, and the willingness to sit in silence until they are answered.
These are The Hope Standard, our own framework for responsible AI, turned around to face outward. We answer them below, and we think every vendor should.
1. What do you collect, and what happens to it?
A good answer lists the specific data, why each piece is needed, where it is stored, who can see it, and when it is deleted. It fits on a page.
A bad answer starts with “industry standard” and ends with a link to a privacy policy.
Ours: We collect only what is needed, protect it rigorously, and never sell it. Before we build anything we write a data inventory, and if a field has no reason to exist, it is cut. Every product ships with a plain-language data page.
2. How do you know the system is making good decisions?
A good answer describes how decisions are logged, how often a person reviews a sample of them, and what was found last time.
A bad answer is “the model is very accurate.”
Ours: We test how our systems make decisions and document what we find, including the surprises. The findings go to the client in writing.
3. Who did you test it on, and what did you find?
A good answer names the groups the outcomes were compared across, the gaps that showed up, and what changed as a result.
A bad answer is “we did not find any bias.”
Ours: We examine our models and data for unfair outcomes before and after launch. If a gap appears, the fix ships before the feature does.
4. Will my staff and the public know when AI is involved?
A good answer shows you the label in the product and the sources the system used.
A bad answer is “it is seamless.”
Ours: Users know when AI is involved and how it reached its answer. Anything generated or ranked by AI is labeled and shows its sources.
5. Who decides?
A good answer points to a person, with a name, who can override the system, and explains what the system is not allowed to decide on its own.
A bad answer describes the workflow without a human in it.
Ours: AI assists. People decide. No final decision about a person’s money, housing, health, or job is automated, and a human’s name is on every recommendation that matters.
Print this. Bring it to your next vendor meeting. If the answers are vague, so is the product.