The AI safety narrative has become a branding exercise. Every frontier lab — the ones racing to ship the most capable models as fast as possible — tells you that safety is their top priority. They publish responsible scaling policies. They hire safety researchers. They write blog posts about alignment.
Then they ship the model anyway. On the timeline that keeps them competitive. Because the alternative is watching a rival eat their market share while they wait for eval results to come in.
This is not a conspiracy theory. It's an observable pattern.
What "Safety First" Actually Looks Like in Practice
Watch the actions, not the announcements.
Capability releases have consistently outpaced safety evaluations across the industry. Internal teams tasked with evaluating risks before deployment have, at multiple organisations, found themselves racing to finish assessments before shipping deadlines — not setting those deadlines. Safety researchers have publicly resigned citing misalignment between stated values and operational decisions. "Responsible scaling policies" are documents companies write themselves, interpret themselves, and enforce on themselves.
There are no binding external audits. There is no independent body with actual authority to say "you cannot ship this." The only check on a lab's safety claims is the lab itself.
That is not safety infrastructure. That is safety theatre with a press release attached.
The Steelman — Because It's Real
Before you write this off as another "big tech bad" take, here's what's true: frontier labs do employ serious safety researchers doing genuine, important work. Alignment is a real unsolved problem and real resources are going into it. Red-teaming evaluations catch real issues that get fixed before deployment. The people working on this are not frauds.
The argument here is not "AI labs are evil and lying through their teeth." The argument is more specific: the marketing language has dramatically outpaced the actual commitment level. When "safety is our top priority" is the public-facing message and "we're shipping next month" is the internal reality, the gap between those two things is the problem.
Credibility is built on the margin between what you claim and what you do. Right now, that margin is wide enough to drive a data centre through.
Why the Race Dynamics Make This Worse
Here's the structural issue. If safety genuinely slows you down — and real safety work does, because you find problems and have to fix them — then any lab that commits harder to safety than its competitors is at a competitive disadvantage.
The market right now rewards capability and speed. It does not reward caution in any meaningful way. Users choose the most capable model. Investors fund the fastest-moving labs. Talent gravitates toward the organisations building the most impressive things.
In that environment, "safety first" as a sincere operational commitment rather than a PR position requires extraordinary institutional will. The labs that have it should be celebrated. The ones whose actions suggest otherwise should be called out.
We are not doing enough of the second thing.
What Would Actually Help
Criticism without a direction is just venting. Here's what meaningful progress looks like.
Binding external audits. Third-party evaluations with actual authority — not voluntary commitments to "engage with evaluators," but mandatory pre-deployment assessments by independent bodies with teeth. If a lab won't agree to this, ask why.
Public release of eval results. If you ran safety evaluations before deploying a model, publish them. Show the benchmarks. Show the red-team findings (appropriately redacted for dual-use concerns). Right now these live in internal documents we never see. Transparency creates accountability.
Funding safety research at labs that aren't building frontier models. Alignment research shouldn't only happen at the organisations with the most commercial incentive to ship.
Standards that apply to everyone. Right now, each lab sets its own safety bar. An industry-wide standard — even a modest one — would prevent the race to the bottom where the least safety-conscious lab sets the pace everyone else has to match to stay competitive.
None of these are radical proposals. They're the kind of thing that exists in every other high-stakes industry. Aviation doesn't let airlines audit their own maintenance. Pharmaceuticals don't let drug companies decide when their own products are safe.
AI is moving faster than either of those industries did. The governance gap is getting wider, not narrower.
The Cost of Letting It Slide
If this is just marketing language — if "safety first" is a phrase that sounds good in investor materials without reflecting genuine operational priorities — then the cost is credibility. For the entire field. The more times a lab says "safety is our top priority" and then makes decisions that suggest it isn't, the harder it becomes to take anyone's safety claims seriously.
That matters because at some point, public trust in AI systems is going to be important. Really important. The decisions being made now about how to talk about safety versus how to practise it are building — or spending — that trust.
We're letting the spending happen without noticing the account balance dropping.