Talking with NVIDIA’s Chief Security Officer, David Reber, I asked the question on behalf of everyone who will never read a model card: since AI relies on data and algorithms, how can normal, non-technology people make sure the data and algorithms don’t have biases — especially when decisions might be based on the AI’s suggestion?
Read the question again and notice what it’s really asking. Not “how do experts audit models?” — there are conferences for that. It’s asking what the rest of us are supposed to do. The loan applicant. The patient. The job candidate. The parent. People who will live inside decisions made by systems they cannot inspect, built by companies they cannot call.
The uncomfortable answer: right now, almost nothing. And that gap — between the people who build AI and the people who are judged by it — is where the trust crisis of the next decade is already loading.
Look at it through the SPICE lenses — Strategy, Politics, Innovation, Culture, Execution:
Strategy. Every AI strategy deck has a slide about trust, and almost none of them answer the mortal’s question: how would I know? The companies that win the next phase won’t be the ones with the most powerful models. They’ll be the ones whose customers can verify, in plain language, why a decision went the way it did — because “trust us, the model is good” is a strategy with an expiration date, and the date is the first public failure.
Politics. Bias survives because accountability is a hot potato. The vendor says the client configured it. The client says the vendor built it. The regulator says the framework is coming. Meanwhile the person who got the biased decision has no name to call, no appeal to file, no standard to point to. This isn’t an oversight in the system. For everyone holding the potato, ambiguity is cheaper than accountability — until the lawsuit, the headline, or the law prices it properly.
Innovation. The genuinely innovative answer in that conversation wasn’t a new auditing algorithm. It was transparency as a design primitive: show how the system was trained, what it knows, where its knowledge stops — and pair it with constant testing and a feedback loop a human can actually reach. The innovation the mortals need isn’t explainability research. It’s a working phone number for the algorithm: a visible way to say “this decision is wrong about me” and have a person obliged to answer.
Culture. Inside most organizations, questioning the model’s output is still read as not being a team player — the analyst who flags a suspicious pattern gets asked why they’re slowing the launch. So the bias that a junior employee spotted in week two ships in week six, and surfaces in year two as a scandal everyone calls unforeseeable. Culture decides whether “the model might be wrong about people like this” is a career-enhancing sentence or a career-limiting one. Right now, in most buildings, it’s the second one.
Execution. So here’s the machinery, sized for mortals and the leaders who serve them: every consequential AI decision ships with three things — a plain-language reason, a human appeal path with a response deadline, and a published error-and-appeal rate, reviewed on the same calendar as the financials. Leaders: demand those three from every vendor, in the contract, before signature. Citizens: when they’re absent, treat the system’s confidence the way you’d treat a stranger’s — politely, and with your hand on your wallet.
Today, we ask non-technical people to trust systems they cannot question, built by institutions that cannot be reached. Tomorrow, the institutions that survive will be the ones that made themselves answerable before they were made to.
Ask yourself: if an algorithm made the worst decision of your life tomorrow, who exactly would you call — and is that answer written down anywhere?
Sources
CXOTalk Ep. 775, “NVIDIA’s Chief Security Officer: Managing Cybersecurity and AI” — my questions on cybersecurity habits and on how non-technical people can guard against biased data and algorithms (transcript): https://youtubetotranscript.com/transcript?v=jyt0O_XG4FE
Leave a Reply