Imagine a toxicity detector, a product with a specialized use case built on top of a general-use frontier model. Its intended use case is automated comment moderation on media sites. Imagine, then, that Responsible AI testing uncovers disparities in how the detector works for users from marginalized identities.

Imagine an RAI team uncovers serious concerns during a pre-launch review. Their assessment is backed by testing showing the product performs poorly for certain demographics even when it’s functioning as intended. The findings are clear, documented, and communicated to the product team. So are recommended mitigations. The RAI team has done its job: identified the risks, documented the evidence, and recommended a path forward.

Then the product launches anyway.

Today, RAI is commonly experienced at varying points before a product launch. A product team submits to the RAI team for a pre-launch review. The RAI team then evaluates the product or model for Responsible AI–related risks and harms and flags any issues to the product team. Then, depending on the amount of potential risks and harms uncovered, the RAI team will either approve the launch or work with the product team on additional testing and mitigations to get the product ship-ready. Over time, with repetition and the calcification of processes, this type of RAI touch point becomes something of a rinse and repeat. This model works; it can feel structured and reassuring to know that responsible AI considerations are an integral part of product launch.

This model works because it creates a structured and reassuring process.

Imagine you’re observing a Responsible AI team deeply focused on a product or model review. For this particular session, everything looks great: policies are codified, confidence scores are acceptable, severity levels are within threshold, and checklists are complete. The review is a success from a process standpoint. Everything seems to work exactly as designed, and so the product launches.

But somewhere in the real world, a darker-skinned user gets misidentified by a vision model, a woman’s resume is not flagged as qualified for a role that she's well suited to fill, or a non-English speaker gets lower quality results than native English speakers. The system may have passed the evaluations, but it still failed someone.

When we look at how Responsible AI functions are structured across tech companies, we see RAI nested under Trust & Safety, absorbed into Privacy, reporting to Security, or matrixed across all three. The result? RAI work gets deprioritized, pulled into reactive incident response, or measured against metrics that don't capture its unique value. 

A groundbreaking paper on fairness evaluation gets 10,000 downloads. Six months later, zero teams have implemented its findings. The research was rigorous, the conclusions were clear, but it never made the journey from PDF to practice. Why?

I've spent nearly a decade working in operations, risk, and Responsible AI, including nearly seven years focused specifically on Responsible AI, across Facebook, Google, and Salesforce. Since ChatGPT's release in late 2022, I've watched countless RAI functions spin up rapidly across the industry. New teams forming, new mandates being created, new experts entering the field.