The Number That Should Alarm Everyone

Forget the vague warnings about AI risk. Here's a specific, data-backed number: 1.31.

That's the half-life — in derivation steps — of ethical constraints on open-weight AI models, according to a May 23, 2026 paper published on arXiv by researchers from Peking University, the University of Science and Technology Beijing, and UC Davis.

The team audited 2,142,823 model repositories on Hugging Face Hub. What they found is damning.

Every time someone takes an open-weight AI model and builds a new version from it, the safety restrictions embedded in the original get diluted. By the seventh downstream generation, at least 80% of descendant models lack sufficient public evidence for any governance determination whatsoever. The researchers call this threshold the "governance horizon" — the point at which the chain of accountability simply breaks.

Seven steps downstream, and the safety rules are effectively gone.

Why the Current System Can't Fix This

Right now, AI governance relies on voluntary metadata disclosures — basically, model creators are trusted to pass safety rules along to whoever builds on top of their work. According to the arXiv researchers, that system has a structural flaw that no amount of enforcement can fully patch.

The problem isn't just that people ignore the rules. It's topology. Some model lineages have no traceable upstream intent at all — the researchers call these "orphan components." When a downstream model inherits from an orphan, no inheritance-based rule can recover the missing governance signal. It's mathematically unresolvable.

The researchers propose a "mandatory-declaration" system that forces each new model to explicitly re-state its governance status, rather than assuming it inherited the upstream rules. The comparison they draw is instructive: PyPI software packages already do this with machine-readable license declarations, and governance signals hold up far better across generations.

NPR's Reporting and the Broader Pattern

NPR reported on May 31, 2026 that removing guardrails from open-weight models has become "dramatically more accessible and popular" in recent months. Noam Schwartz, CEO of AI security firm Alice, told NPR: "Everybody can download and operate their own state-of-the-art model and use it for great things and terrible things."

That's accurate. But the arXiv research reveals the problem is structural and systemic — it's not about one person stripping guardrails. The entire ecosystem of derivative models is drifting away from any governance baseline, automatically, at scale, regardless of anyone's intentions.

Trump's AI Summit: Pedal Down at the Wrong Moment

On July 23, 2025, President Trump held the "Winning the AI Race" summit in Washington. According to Forbes contributor Paulo Carvão — a Senior Fellow at Harvard — Trump signed an executive order banning what he called "woke AI," called for deregulation of data centers and energy projects, and pushed to make the U.S. an "AI export powerhouse."

Trump also advocated for a single federal AI regulatory standard, which on its face is reasonable — a patchwork of 50 state laws would be a legitimate mess for any company trying to operate nationally.

The administration's entire frame is: government gets out of the way, private sector leads. That might be the right call for semiconductor fabs and permitting timelines. But peer-reviewed research just proved that governance infrastructure is mathematically collapsing on its own. Deregulating safety architecture isn't the same as deregulating energy permits. One is red tape. The other is the only thing standing between the public and a 2-million-model ecosystem with zero traceable safety rules.

The Carnegie Framework Nobody Is Implementing

Back in July 2024, the Carnegie Endowment for International Peace published a paper — authored by 18 experts including former DHS Secretary Janet Napolitano, Harvard Law professor Larry Lessig, and Anthropic's Irene Solaiman — arguing that the "pro-open vs. anti-open" AI debate was a false binary and that consensus governance frameworks were achievable.

The Carnegie paper called for "mixed release strategies" and smarter governance that doesn't just slam the door on open models but also doesn't pretend they're consequence-free.

That was a year ago. The arXiv research from May 2026 confirms the governance gap Carnegie warned about is NOT being closed — it's getting mathematically worse with every new model derivative generated.

The NYT editorial board argues society needs to take AI's future into its own hands rather than leaving it to businesses. But architecture matters more than sentiment.

What This Means for Regular People

If you're a normal person who doesn't work in AI, here's the translation: the models being used to answer questions about making explosives, planning violence, or generating illegal content are increasingly untraceable back to any responsible party. No license. No governance record. No accountability chain.

The companies that built the original models can shrug and say their safety rules were in the original release. The people running seven-generations-downstream derivatives never agreed to anything.

Trump's executive order banning "woke AI" is a political document. The arXiv paper is a technical one. Both came out in the same news cycle. The model repositories aren't waiting for Washington.