Responsible AI Part 4

Every conversation about RAFT’s fourth pillar runs into the same unmarked pothole: the word “transparency” is doing two different jobs, and the framework never separates them. Ask a company to be transparent about its AI use, and it might explain how a model reached a decision. Or it might hand over a record of every system that touched the underlying data. Those are two distinct obligations, not one, and treating them as interchangeable lets a company satisfy one while believing it has covered both. This article treats them separately, on purpose: explainable AI decision-making, and documented data touchpoints. Neither substitutes for the other. And that’s why we’re talking about responsible AI.

Explainability: Can You Show Your Work?

Transparency in the model itself, or in the software wrapped around it, means being able to say: I can tell how this machine made its decision. If I use an AI tool and don’t have it write a log of its decision-making process, I’ve failed transparency on explainable decision-making. Directing the tool to take notes at every stage, in a format I can actually audit, is what lets me point to exactly how the system reached its answer.

That distinction matters because explainability is a configuration choice, not an accident of the technology. A model can be made to show its reasoning at every step. A company has to direct it to do so, and keep the record somewhere auditable. Skip that step, and the failure belongs to the company, not the tool.

This is also where Part 1’s warning about rubber-stamping agentic workflows becomes concrete. A workflow that hands off task after task with no intermediate log cannot be explained after the fact, no matter how carefully a human reviewed the final output in the moment. Explainability has to be built in before the decision happens, not reconstructed from memory afterward.

The Ingredients Behind the Appliance

Documented data touchpoints are about the ingredients. If the model is the appliance, the data feeding it is the ingredients: what kind of data does a company have, and how biased is it? Everything is biased. The real question is what kind. Is the bias direct, against a protected class, or is it collinear? Collinear bias is one of the biggest problems in classical machine learning. A company might not discriminate against Black people directly. It might discriminate on household income instead. But household income correlates with ethnicity closely enough that discriminating on income means discriminating against a protected class, by definition.

The test for catching this is concrete: take a Black family and a Korean family at the same income level. The decision should come out identical. Take those same two families at different income levels, and the decision should track income, not ethnicity. If it doesn’t, that gap is exactly what this kind of transparency exists to catch. Explainability tells a company how the machine decided. Data-touchpoint documentation tells it what data fed the machine, and whether that diet was quietly discriminatory all along.

Answering that question in the abstract isn’t the same as having it documented. An auditor examining this kind of transparency wants granular detail: every system a piece of data touches, every person with access to it, and how well-protected it stays at each stop along the way. If a new hire logged in tomorrow, would they have immediate access to that data point? A company needs a written trail that answers that, not a confident guess reconstructed from memory when someone finally asks.

A Disclosure Statement Is a Practice, Not a Document

A public AI-use disclosure statement matters, but a single static statement isn’t enough. Disclosing AI use has to be continuous, not a one-time filing that a company writes once and forgets. Continuous disclosure reinforces the copyright claim on the human-originated work a company started with, and it respects a real, non-trivial segment of the audience that doesn’t want purely AI-generated content. Telling people when AI is in use, and how, earns their respect for the honesty, even from people who would rather a company not use AI at all.

That kind of disclosure creates a practical asset, not only an ethical one. A company that documents which source data fed which output builds a chain of evidence: proof that a given piece of content derives from its own material, rather than someone else’s. Enterprise systems like IBM’s WatsonX already build in this level of granularity, keeping a full trail from raw data through processed output, so the record exists if the system is ever subpoenaed. A company that can produce that trail retains the copyright on a fully derivative work, even though a machine helped summarize or structure it. Disclosure protects the audience. It also protects a company’s own claim to what it made.

How This Very Series Was Made

That standard applies here, to this series itself, and not only to the companies it describes. So here is the disclosure, in full, rather than a passing mention.

This series is a human-led podcast and a human-led framework. For this round, I recorded verbal answers to Claude’s questions on video, sitting down and talking through each pillar the way I would in any strategy session, screw-ups, tangents, and all. That’s not a flaw in the process. It’s evidence of it: the reasoning, the analogies, and the positions in these four articles are mine, worked out loud, on the record. Claude then assembled those recorded answers into the finished articles you’re reading, organizing, tightening, and structuring the language, but not inventing the substance. The reasoning is mine. The assembly is machine-done. That’s the disclosure: true human input, machine-assisted assembly, and I’m telling you exactly where that line falls, the same way I’d expect any company to tell you where theirs falls.

The Standard: Operate as if You’re Being Audited Tomorrow

Every pillar in this series needs a standard a reader can actually apply, not only a value to admire. For transparency, Katie set that standard years before we ever wrote it down as RAFT:

“I always operate as if we’re going to be audited tomorrow. I like to have those very clear trails.”

The instrument for meeting that standard already exists. It’s the YAML checklist introduced in Part 2 on Accountability, built from RAFT plus a company’s own values, employee handbook, and terms of service. Use that same checklist to audit a company’s efforts across all four pillars: Respect, Accountability, Fairness, and Transparency. Produce the checklist, show the work behind it, and a company is in good shape.

The standard has teeth because real verification frameworks already back it up. A company bound by ISO 27001, SOC 2, or HIPAA compliance has to hold its AI vendors to those same standards. A vendor that can’t demonstrate equivalent rigor is a gap in the company’s own compliance, not a separate problem to set aside. Playing out the audit scenario now means asking concrete questions before anyone forces the issue. What needs to be documented and disclosed? How visible are those disclosures to partners and customers? Are the underlying contracts current? Answering those questions now costs far less than reconstructing years of AI usage after a regulator or an opposing attorney forces the issue.

When Honesty Itself Becomes the Problem

Part 3 raised a similar edge case for Fairness: a company that is openly and consistently biased is technically meeting the say-what-you-do, do-what-you-say test, however repugnant its values. The same logic applies to Transparency, in an even blunter form.

Bluntly honest but harmful transparency is acceptable only within a narrow frame. A company that discloses harmful values and lives them consistently is meeting the say-what-you-do, do-what-you-say standard, even if the values themselves are repugnant. That doesn’t buy a free pass: customers retain the responsibility to take their business elsewhere, except in genuine monopoly situations where that choice doesn’t really exist. What’s actually unethical, and what this pillar is really guarding against, is claiming good values while doing something else entirely.

Two Threads Worth Pulling Again

Two connections from earlier in this series apply directly here.

Data privacy’s transparency angle is straightforward: a company that says it keeps customer data completely safe, and doesn’t, faces more than an ethics problem. It’s likely exposed legally too.

And the “whose values” question from Part 1 resolves the same way here as it did there. Whose values populate a responsible-AI policy depends on who works at a company and who its customers are. A team doesn’t have to be uniform in its makeup; diversity is a good thing. But a company’s values and its customers’ values need to be in alignment. When they’re not, problems surface eventually.

Closing the Loop: Four Pillars, One Test

Four parts, four pillars, and one test running underneath all of them. Fairness is the test you run. Respect is the value the test protects. Accountability is who owns the test results. Transparency is the disclosure of those results.

Part 1 opened this series with the reason any of it matters now: models have grown capable enough to amplify whatever harm is built into how a company uses them. A cordless drill can only do so much damage in careless hands; a Milwaukee Hole Hawg bores a hole through a car. That capability gap is why “responsible AI” can’t stay a vague, feel-good phrase. It has to mean something specific enough to test, to own, and to disclose. Fairness, Respect, Accountability, and Transparency are how a company makes that phrase mean something: four pillars that only function as one connected system.

“Governance is like therapy and dentistry. Those first catch-up appointments are going to be rough, but once you get into a regular rhythm of maintenance, it gets a lot easier.”

Need help with your marketing AI and analytics?