Opinion | I Worked on Safety at OpenAI. The Fix Isn’t Hard | Backup software
By Gammatek ISPL, Industrial Systems & Compliance Analyst at Gammatek ISPL
Last updated: September 2026 | 13 min read
Author block: Gammatek ISPL writes on safety culture and compliance systems across high-risk industries at Gammatek ISPL, drawing on direct experience helping manufacturing, chemical, and pharmaceutical plants build audit-ready safety infrastructure. This piece responds to recent public commentary from former AI industry safety researchers; it does not claim any first-hand experience inside an AI lab.
Why This Matters
Over the past two years, a recurring pattern has played out at the world's leading AI companies: a safety researcher joins with genuine conviction, works for a few years, then resigns publicly and says some version of the same thing — the technical fix for making the system safer wasn't the hard part. Leadership priorities were. If you work in any high-risk industry — manufacturing, pharma, chemical processing, or increasingly, software — this pattern should sound familiar, because it's not new. It's the exact failure mode industrial safety engineering spent the better part of a century learning to prevent. Understanding why that earlier fix worked, and why AI companies haven't adopted an equivalent one, tells you something real about what to expect from the next wave of AI-safety headlines — and what "safety" should actually mean when a vendor promises it to you.
The Pattern, Documented
This isn't a single incident. Over roughly the past two years, a string of departures from frontier AI labs has followed a near-identical script. In 2024, OpenAI's superalignment team — co-led by Jan Leike and company co-founder Ilya Sutskever, and explicitly tasked with managing long-term risk from advanced AI — was dissolved shortly after both men resigned. Leike wrote publicly that "safety culture and processes have taken a backseat to shiny products," and that the team had been under-resourced relative to the commitments the company had made to it.
Since then, similar accounts have followed from other researchers at other companies, each describing a version of the same tension: safety work that is technically well-understood, but that loses out internally to shipping deadlines and competitive pressure. The claim embedded in nearly all of these departures is the same one implied by the "fix isn't hard" framing — that the obstacle isn't a lack of technical solutions, but a lack of organizational will to prioritize them over product velocity.
What "The Fix Isn't Hard" Usually Means in Practice
When a departing researcher says the fix isn't hard, they typically mean some combination of the following, based on the public record of these resignations:
Give the safety team real authority to delay releases, not just advisory input that leadership can overrule.
Commit compute and staffing resources at the level originally promised, rather than reallocating them toward product teams when deadlines tighten.
Build genuine, structural separation between the people evaluating a system's safety and the people whose compensation depends on shipping it.
Make safety review a mandatory gate, not a parallel process that can be bypassed under time pressure.
None of these are cutting-edge machine learning problems. They're organizational design problems — the same category of problem that industries with far longer safety track records solved generations ago.
The Comparison AI Coverage Keeps Missing: Industrial Safety Engineering
Here's the part of this story that gets almost no attention in AI-focused reporting: manufacturing, aviation, chemical processing, and pharmaceutical production all went through their own version of this exact fight — decades ago — and came out the other side with structures that actually work.
Historical industrial pattern (pre-reform) | AI labs today | |
Safety team authority | Advisory only; production could override | Advisory only; product leadership can override |
Resourcing | Cut first when budgets tightened | Compute/staffing reallocated under deadline pressure (per Leike's account) |
Independence | Safety reported into the same chain as production | Safety teams historically embedded within/reporting to product-aligned leadership |
What changed | Regulatory mandates + independent audit authority + compensation structures tied to safety outcomes, not just output | Largely voluntary self-governance; no equivalent external mandate yet exists |
The chemical and manufacturing sectors didn't fix this by hiring smarter safety engineers. They fixed it by restructuring who has authority to stop a line, who gets paid for what outcomes, and who audits the audit — usually only after regulators or catastrophic incidents forced the issue. That's the uncomfortable parallel: industrial safety culture didn't improve voluntarily ahead of the pressure that forced it. It improved after enough failures made the cost of not fixing it higher than the cost of fixing it.
An Implementation Consideration: What Real Structural Fixes Look Like
Based on what's actually worked in high-risk industrial settings, a genuine structural fix — as opposed to a policy statement — usually includes:
A safety function with veto authority that doesn't report through the same leadership chain as the product it's reviewing. In manufacturing, this is often a legally distinct compliance or EHS (environment, health, safety) function with its own reporting line to the board, not to the plant manager whose bonus depends on output.
Resourcing commitments that are contractual or board-mandated, not discretionary. "20% of compute for safety" as a stated goal is meaningfully different from 20% as a binding commitment with consequences for missing it — this is precisely the gap Leike's resignation post described.
External, independent audit — not just internal review. Industrial safety regimes rely on third-party inspectors and regulatory audits precisely because internal safety teams, however well-intentioned, operate under the same commercial pressure as everyone else at the company.
Compensation tied to safety outcomes for the people who can override safety decisions, not just for the safety team itself. If the person who can overrule a delay has no compensation exposure to what happens if that override goes wrong, the incentive structure guarantees safety loses close calls.
Why This Hasn't Happened at AI Labs Yet
A fair question, and the honest answer isn't flattering to anyone involved: the equivalent of external regulatory mandate for frontier AI is still being built, slowly, and voluntarily-adopted internal structures face the same commercial pressure that made voluntary industrial safety commitments unreliable for most of the 20th century. AI companies are, in effect, being asked to do voluntarily and early what most heavy industries only did after either binding regulation or a high-profile disaster forced the issue.
That's not a defense of the current state — it's a explanation for why "the fix isn't hard, technically" and "the fix isn't happening" can both be true at once. The fix is organizationally simple and politically difficult, which is a very different problem than "nobody knows how."
What This Means If You're Evaluating an AI Vendor's Safety Claims
If your organization is adopting AI tools — and increasingly, most manufacturing and compliance-heavy businesses are — the practical takeaway isn't to distrust every vendor's safety messaging. It's to ask the same questions an industrial auditor would ask of any safety claim:
Does the safety function have actual authority to delay a release, or only advisory input?
Is safety resourcing a binding commitment or a stated aspiration?
Is there independent, external verification of safety claims, or only internal self-reporting?
What happened the last time safety and shipping deadlines conflicted at this vendor?
These are the same questions industrial compliance teams ask about their own internal safety functions — because the answer determines whether a "safety commitment" is a structural fact or a marketing statement.
The Honest Takeaway
The departing researchers are very likely right that the technical fix isn't the hard part. What industrial safety history shows is that knowing the fix isn't hard has never been enough on its own — the fix requires giving up real authority and real money to a function that, by design, exists to say no to the people paying for it. Manufacturing took decades and multiple serious incidents to get there. Whether AI companies get there faster, on their own, remains the open question — and it's one worth watching with the same skepticism industrial auditors apply to any organization that says "we've got safety handled" without independent verification to back it up.
Where This Connects to Your Own Compliance Systems
The exact structural principle at stake here — independent authority, binding resourcing, external audit — is the same one Gammatek's compliance platform is built around for physical plant safety: a system that documents whether safety review actually has teeth, not just a policy that says it should. As AI tools get embedded deeper into manufacturing and pharma operations, the question of who can actually stop a process when something's wrong becomes just as relevant to your AI vendor evaluation as it already is to your plant safety program. https://www.gammateksolutions.com/post/ai-hasn-t-gone-rough-its-worst-than-that https://www.gammateksolutions.com/post/america-must-learn-ai-lessons-from-astro-boy https://www.gammateksolutions.com/post/the-best-worst-and-strangest-ways-ai-is-really-being-used-at-work
Comments