AI Policy Myths: What the Headlines Missed on Rules and Risk
Archi4 min read
Contents
- The Fine Print Behind the AI Policy Headlines
- Myth one: the law regulates models
- Myth two: there is one certificate to collect
- Myth three: open weights mean exemption
- Myth four: small firms are out of scope
- Myth five: one compute number settles coverage
- Myth six: enforcement starts on the date in the headline
- Myth seven: the banned list is obvious
- What the headlines skipped: the paper supply chain
- The part nobody budgeted for
The Fine Print Behind the AI Policy Headlines#
A hundred pages of regulation becomes a six-word headline. Europe bans AI. White House cracks down. The distance between the two is where compliance budgets get built, and it is wider than most coverage suggests.
Six claims traveled further than the text they came from. Here is what the documents actually say.
Myth one: the law regulates models#
Most of the operative text regulates roles and tasks. A provider that places a system on the market carries one set of duties. A deployer, meaning the hospital, bank or staffing firm that uses it, carries another. The same model can sit inside a low-risk spam filter and a high-risk recruitment tool, and the obligations follow the use, not the weights.
That distinction decides budgets. Deployers rarely have model access, yet they own the human oversight requirement, the logging requirement, and in some cases a fundamental rights impact assessment.
Myth two: there is one certificate to collect#
There is no single badge. High-risk systems need a risk management file, technical documentation, data governance records, a logging design, and instructions for use that a buyer can actually read. Providers sign a declaration of conformity and affix a mark. Then post-market monitoring begins, with serious incidents reported on a clock.
None of that is a one-time purchase. The file has to stay current while the system changes, and systems change every sprint.
Myth three: open weights mean exemption#
Free and open source releases get partial relief, mostly from documentation duties, and the relief comes with conditions about how the model is released and what is published alongside it. It does not erase obligations for prohibited uses. It does not cover a provider that later puts the same model behind a paid API.
Deployers get no relief at all. Downloading weights from a repository makes you a deployer the moment you use them in a regulated task.
Myth four: small firms are out of scope#
Size thresholds exist, and they mostly change how an obligation is met (lighter documentation, simplified quality systems) rather than whether it applies. What small firms actually face is a procurement problem. Their customers are large deployers who need paperwork to satisfy their own duties, and those customers pass the requirement down the contract.
A ten-person vendor selling into a bank will be asked for a technical file long before any regulator asks.
Myth five: one compute number settles coverage#
The general-purpose model threshold, training compute above 10^25 floating point operations, is the number that gets quoted. It is a presumption for systemic risk, not a fence around the whole regime. Models below it still carry transparency duties. A fine-tune or a merged checkpoint can move a system into a different category without retraining from scratch.
The threshold also says nothing about the deployer, who is covered regardless of how small the model is.
Myth six: enforcement starts on the date in the headline#
Application is staggered. Prohibited practices came first, then general-purpose model duties, then high-risk systems, with some product categories pushed further out. Each wave depends on standards. Harmonized standards from the European standards bodies have slipped repeatedly, which means early movers are conforming against drafts and revising.
Companies that read only the first date built timelines around a deadline that had already moved.
Myth seven: the banned list is obvious#
The prohibitions read narrowly. Social scoring by public authorities, untargeted scraping of faces to build databases, emotion inference in workplaces and schools, biometric categorisation of sensitive traits. Each clause has carve-outs for medical and safety uses, and each will be argued over by lawyers for years.
The practical effect is smaller than the headlines suggested and sharper than the industry response implied. Most products are nowhere near the list. A few, sold into HR and education, are much closer than their vendors think.
What the headlines skipped: the paper supply chain#
The unglamorous bottleneck is documentation moving between vendors and customers. A deployer must assess a system it did not build. The vendor holds the test results. Neither side has a standard template, so every deal negotiates one.
Watch where that lands. Procurement teams, not legal departments, are writing the AI clauses now. Data processing addenda are being rewritten to cover model inputs. Insurance questionnaires have started asking for evaluation logs.
The part nobody budgeted for#
Translation is the quiet cost. A technical file written for engineers does not satisfy a supervisor's request for plain-language instructions. A logging design that captures everything costs storage and creates its own privacy exposure. A human oversight requirement needs a person with authority to overrule the system, and that authority has to appear in a job description and a shift rota.
Policies do not fail because a model misbehaves. They fail because the paperwork describing the model never matched the system that shipped, and no one noticed until the customer audit arrived.
More articles