The Cleanup Crew Was Always the Architecture Team

🌐 The AI Inflection Point (TAIIP) examines the AI era from the outside in, showing how infrastructure, policy, and power shape outcomes — when decisions made far upstream quietly determine what workers, communities, and institutions experience downstream. Curated monthly.

It goes like this.

A technology company builds something powerful and deploys it fast. Questions emerge about bias, about harm, about what happens when the system produces outcomes nobody explicitly approved. A team gets assembled to address those questions: an ethics team, a responsible AI team, a safety team, an alignment team. The team does real work. They identify risks. They write memos. They flag the gray zones that the deployment teams haven’t accounted for.

Then growth slows, or a competitor accelerates, or a new partnership creates pressure to ship faster. And the ethics team—the one doing the work of understanding where the system breaks—gets cut. Reorganized. Transferred into other functions where their authority dissolves into process compliance. Or eliminated outright.

The justification is always efficiency. The result is always the same: the people who understood where the system breaks are no longer in the room when the next version of the system gets built.

This is not a recent phenomenon. It’s a structural habit. And in 2026, as AI systems scale faster than the governance infrastructure designed to oversee them, that habit is producing consequences that can no longer be contained inside individual organizations.

This month in TCP, You Already Know Where It Breaks, I wrote about what this pattern costs at the individual level and what becomes possible when you stop accepting the misfiling. This essay is about what it costs at the institutional, systems, and policy level, and why the reclassification of architecture work as support work is not an incidental oversight but a structural problem with structural consequences.

Let’s look at what actually happened.

Microsoft’s Ethics & Society team was, at its peak, a 30-person department. Engineers, designers, philosophers — people whose job was to ensure that AI products were designed and deployed responsibly, and who identified risks before those risks became public crises. In October 2022, the team was cut to seven people in a reorganization. In March 2023, those remaining seven were eliminated — casualties of a broader layoff wave, timed precisely as Microsoft was accelerating its integration of OpenAI technology across its entire product suite.

The reported reason: pressure from senior leadership to get OpenAI models into customers’ hands faster. The ethics team, one source noted, was seen as pulling the reins — slowing things down by pointing out potential societal consequences and legal ramifications.

That framing is the tell. Pulling the reins is not how you describe an architecture function. It’s how you describe a support function that’s getting in the way. The reclassification happened in the language before it happened in the org chart.

Meta disbanded its Responsible Innovation team (RIT) around the same period. The team, which included engineers, ethicists, and civil rights specialists, was tasked with identifying and addressing potential harms to society from Meta’s products. The company stated that the team’s work would continue, but its members were redistributed to other roles to focus on specific issues rather than centralized oversight.

Twitter (X): Ethics Unit dissolved its Machine Learning, Ethics, Transparency, and Accountability (META) team, which was well-regarded for its work in ethical AI and algorithmic transparency. The META team had previously published research on algorithmic bias and launched initiatives like the “bias bounty” contest.

OpenAI’s Superalignment team, formed in 2023 to study long-term existential risks from advanced AI, was disbanded in May 2024. Its Mission Alignment team, which was formed in September 2024 to ensure AGI benefited all of humanity distributed its team’s 6–7 members to other roles.

The pattern is not subtle. When speed becomes the primary metric, the people who understand the costs of moving too fast get reclassified as obstacles — and then removed.

The pattern has now scaled beyond individual companies. The Trump administration entered office in 2025 having torn up Biden-era AI safety orders, pledging to strip away what it characterized as barriers to innovation. By May 2026, the same administration was scrambling to draft an executive order on AI safety — triggered by the emergence of models capable of finding security flaws at unprecedented speed. A senior White House official observed that cabinet members were suddenly using the words “safety” and “AI” in the same sentence, “which is not how the admin was talking about these issues even a few months ago.”

The governance vacuum doesn’t stay invisible forever. It becomes visible the moment the cost of ignoring it arrives faster than the infrastructure to manage it.

Share

Here’s what those teams were actually doing, stripped of the ethics branding.

They were mapping the terrain of systems that produce gray-zone outcomes — the probabilistic, distribution-based results that binary governance frameworks weren’t built to handle, as I mapped in The Governance Lag. They were building the decision frameworks that governance infrastructure requires but rarely has. They were identifying, in advance, the places where algorithmic outputs would create harm that technical performance metrics wouldn’t capture. They were, in structural terms, doing architecture work — designing the constraints and accountability structures that make complex systems governable.

That’s not support work. That’s not compliance work. That’s not the cleanup crew.

It is, however, consistently filed as support work — positioned in org charts as a check on the real work of building, rather than as a constitutive part of what building responsibly requires. And when it gets filed that way, it becomes easy to defund. Easy to reorganize out of the room. Easy to frame as overhead when speed and competitive pressure intensify.

The reclassification has a demographic dimension worth naming directly. A 2022 ACM study1 found that women and Black respondents prioritize responsible AI values more highly than other groups — including more than AI practitioners overall.

Communities that have had more reason to understand what inequitable systems produce have developed a different relationship to the work of preventing harm. That orientation is itself a form of expertise. And the work it produces has been systematically undervalued in ways that mirror broader patterns of whose knowledge gets recognized as technical and whose gets filed as instinct, or soft skills, or simply “good with people.”

When those teams get cut, it’s not just governance capacity that disappears. It’s a specific kind of expertise — built over years, in conditions that weren’t designed to support it — that gets removed from the room precisely when the decisions it’s most suited to inform are being made.

The governance lag I mapped in March isn’t separate from this pattern of disbanding ethics and responsible AI teams. It’s a direct consequence of it.

Every team eliminated is a body of institutional knowledge that doesn’t transfer cleanly into the teams it gets distributed across. The frameworks are half-built, the risk assessments started but not finished, the pattern recognition developed over years of watching where systems break that doesn’t survive reorganization. It disperses.

And the organizations that dispersed it discover, twelve to eighteen months later, that they’ve lost the capacity to see risks coming before they arrive.

Organizations that eliminated their ethics and responsible AI functions are now paying external consultants to rebuild from scratch what they had, at lower cost and higher quality, already built internally. The governance frameworks now being required by regulators—the risk categorizations, the oversight plans, the accountability documentation—are exactly the work those teams were doing before they were reclassified as overhead.

This is what institutional knowledge loss looks like in practice: you don’t notice it immediately. You notice it when you need the capacity and discover it’s gone — and when the cost of reconstructing it is measured not just in consulting fees but in the harms that accumulated in the gap.

Here’s what the outside-in lens reveals that the organizational view misses.

The governance lag isn’t primarily a regulatory problem. It isn’t primarily a technology problem. It’s a classification problem. Institutions have systematically misfiled the work of understanding where complex systems break, labeling it ethics, labeling it responsibility, labeling it compliance in ways that position it as secondary to the real work of building.

That misclassification has consequences at every level of the stack.

At the organizational level, it means the people who understand gray-zone risks aren’t in the rooms where gray-zone decisions get made. At the industry level, it means governance standards get built by the people who benefit from minimal constraint rather than the people who understand what constraint actually prevents. At the regulatory level, it means frameworks get written without the input of practitioners who know where the systems break in ways that metrics don’t capture.

And at the community level — which is where the TAIIP lens always returns — it means the people who absorb the consequences of gray-zone AI outcomes have no mechanism to participate in the decisions that produce those outcomes. Because the teams that were building that mechanism got reclassified and removed.

The governance vacuum isn’t an accident of insufficient regulation or insufficient technology. It’s the predictable result of systematically removing the people who were doing governance work from the institutions that most needed them.

In March, I ended with a question: will we build better governance deliberately, or wait for enough crises to force it?

The answer, it turns out, depends on whether institutions can correct a classification error they haven’t yet fully named.

The work of understanding where AI systems break is not support work. It is not overhead. It is not the cleanup crew arriving after the damage is done.

It is architecture — the foundational design work that determines whether complex systems are governable at all.

And until institutions file it that way — in org charts, in budgets, in who gets invited to the rooms where consequential decisions get made — the governance lag will persist regardless of how many regulatory frameworks get written or how many ethics principles get published.

Correcting the classification requires more than protecting ethics teams from the next round of cuts. It requires changing what counts as technical expertise, who gets recognized as doing governance work, and where in the decision sequence that work gets positioned. Not after the system is built and something goes wrong. Before the blueprint is finalized.

The cleanup crew was always the architecture team. Institutions that recognize that now—and restructure accordingly—will build AI systems that are governable. The ones that don’t will keep paying for the same misfiling, in the same way, until the cost becomes undeniable.

That reckoning is already underway. The only question is whether institutions meet it deliberately or get dragged into it by the consequences they could have prevented.

In TCP this month: You Already Know Where It Breaks — what the reclassification pattern costs individuals and why now is the moment to move. Coming later this month for Premium member’s Deep Dive: Seeing the Terrain: How to Navigate the Architect Transition Without a Map.

Leave a comment