This is Part 4 of The Vertical Wall, a five-part series on the growing divide between who AI helps and who it leaves behind. Part 1 covered the talent divide. Part 2 covered who actually gets to use enterprise AI. Part 3 covered the infrastructure that makes all of it possible or impossible. This one is about who gets to decide what "safe" means, and what it costs.

The Promise

Anthropic was supposed to be the good one.

Founded in 2021 by Dario and Daniela Amodei both ex-OpenAI, both leaving specifically because they thought OpenAI wasn't taking safety seriously enough the company built its entire identity around one idea: we will be the lab that puts safety first, even when it hurts.

They wrote it down. Formally. Their Responsible Scaling Policy, or RSP, was the AI industry's most explicit safety commitment. The core promise was simple and radical: if Anthropic's internal safety evaluations determined that their models were approaching dangerous capability thresholds, they would pause development. Full stop. Not "slow down." Not "convene a committee." Stop.

In an industry running flat out toward artificial general intelligence, this was the equivalent of a Formula 1 team publicly committing to pull over if they thought the car was getting too fast. It was unusual. It was brave. And it worked at least as brand strategy. The 2024 AI Safety Index ranked Anthropic highest among all major AI labs for safety practices. Investors loved it. The press loved it. "Responsible AI" had a poster child.

That poster child just tore up the poster.

The Rewrite

On February 25, 2026, Anthropic quietly updated its RSP. The change didn't get a press conference. It got a blog post.

The commitment to halt development? Gone. Replaced with language about "maintaining competitive position" and ensuring that safety measures don't cause Anthropic to "fall behind" other developers. The new framing, buried in the updated policy: "If one AI developer paused development to implement safety measures while others moved forward... that could result in a world that is less safe."

Read that again. The company founded on the principle that someone needed to be willing to stop just argued that stopping would be dangerous.

The timing tells the story. Anthropic had been deepening its relationship with the Pentagon. Federal contracts were on the table the kind of contracts that don't go to companies with self-imposed kill switches. You can see the pressure in the money trail: Anthropic quadrupled its federal lobbying spend to over $3.1 million in 2025 and directed $20 million to a political nonprofit, according to OpenSecrets data published in March 2026.

Here's the part that should make your stomach turn. While Anthropic was quietly dismantling its own safety commitments, it was simultaneously lobbying the White House for stronger industry-wide safety regulations. WinBuzzer put it plainly: Anthropic "quietly dropped earlier safety commitments while simultaneously lobbying the White House for stronger industry regulations."

Safety for thee. Flexibility for me.

This isn't hypocrisy for its own sake. It's strategy. If you can get the government to impose safety requirements that you've already learned to navigate but your competitors haven't, you've turned regulation into a competitive advantage. You haven't made the world safer. You've made your market position more defensible.

The Moat

And that brings us to the part of this story that affects everyone who isn't a billion-dollar AI lab.

The EU AI Act is the most comprehensive AI regulation on the planet. It's also the most expensive to comply with. If you're deploying a high-risk AI system in Europe and the Act's definition of "high-risk" is broad enough to catch everything from hiring algorithms to medical chatbots you're looking at €29,277 per year per unit in compliance costs, plus €16,800 to €23,000 for certification. Per unit.

A startup running five AI models is staring at €200,000 or more per year in compliance overhead before it earns a single euro in revenue.

Google can absorb that. Microsoft can absorb that. A solo developer in Lisbon building a mental health chatbot cannot. She faces the same regulatory requirements as a company with a market cap larger than most countries' GDP. Same forms. Same audits. Same legal exposure. The Act doesn't scale its demands to the size of the builder. It scales them to the risk category of the product.

On paper, this makes sense. A mental health chatbot can cause real harm whether it's built by Google or by one person in a studio apartment. The risk to the user is the same.

In practice, it means the only people who can afford to build in regulated categories are the people who were already big enough to hire compliance teams. Safety requirements don't just protect users. They protect incumbents.

And the incumbents know it. There's a reason the biggest AI companies don't fight regulation they shape it. When you've already built the compliance apparatus, every new rule is a toll booth that you sail through and your competitors line up at. The lobbyists aren't trying to kill regulation. They're trying to write it.

This is the pattern we've been tracing through this entire series. Part 1 showed how AI skills are concentrating in the people who already had advantages. Part 2 showed how enterprise AI tools flow to companies that can afford the implementation. Part 3 showed how infrastructure determines who can even participate. Now add a compliance layer that costs six figures annually, and the wall gets another ten feet taller.

Safety isn't the villain here. The cost distribution is.

The Open-Source Paradox

So people go where the gates aren't.

Meta's Llama models have spawned more than 85,000 derivatives on Hugging Face. Eighty-five thousand. Each one a fork, a fine-tune, a custom build created by developers who downloaded the weights, modified them for their use case, and deployed them. No centralized safety review. No compliance certification. No one checking what the 47,000th derivative is being used for.

DeepSeek, the Chinese lab that rattled the industry in early 2025, distributes open weights with no Western-style safety layer. And it's being adopted fastest in exactly the places with the least regulatory infrastructure Southeast Asia, Africa, parts of Latin America. Developers in Lagos aren't choosing DeepSeek because they oppose safety. They're choosing it because it's available, it's capable, and no one is charging them €200,000 for the privilege of using it.

Yann LeCun, Meta's chief AI scientist, has a line he repeats at every conference: "Making AI open source is the best way to make it safe." The argument is that transparency enables collective scrutiny. Thousands of eyes on the code catch problems that a closed lab never would. Open is safe.

It's a compelling argument. It's also incomplete.

Open-source safety depends on the community having the resources, expertise, and incentive to actually do the scrutiny. A Nigerian developer building a customer service bot with a Llama derivative isn't auditing the base model for bias. He doesn't have time. He doesn't have the tools. And frankly, it's not his job he's building a product, not running a safety lab. The "thousand eyes" theory works when those eyes belong to well-funded researchers at institutions with safety mandates. It works less well when adoption outpaces the community's ability to police itself.

So we're left with a paradox. Closed models have safety infrastructure but gate access through cost. Open models democratize access but distribute risk to the people least equipped to manage it. Neither approach actually solves the safety problem. They just move the burden to different populations.

The irony is thick. The AI safety community spent years warning about "unaligned" AI systems pursuing goals that diverge from human values. The alignment problem they should have been studying was closer to home: companies whose safety commitments diverge from their business incentives the moment real money is on the table.

Who Pays

This is the question underneath all of it: who pays for AI safety?

Right now, the answer is: whoever can't afford to avoid it.

Big labs pay when it's strategically useful and stop paying when it's not. Anthropic's RSP was a safety commitment right up until it conflicted with government revenue. Then it became a "framework" flexible, interpretable, non-binding. The company that scored highest on the 2024 AI Safety Index didn't become less safe because it changed its mind about ethics. It became less safe because the incentive structure changed.

Regulated companies pay through compliance costs. The EU AI Act doesn't ask whether you can afford an audit. It asks whether your product falls into a risk category. If it does, you pay. If the cost of paying exceeds what your startup can survive, then you don't build that product at least not in Europe. The small builders get priced out. The big builders write the compliance costs into their operating budgets and move on.

And the developers outside the regulatory perimeter? They pay nothing. Not because they're irresponsible, but because no one has built a safety infrastructure that reaches them. The Nigerian developer, the Indonesian startup, the Brazilian team fine-tuning Llama for agricultural advice they operate in a regulatory vacuum. Not by choice. By circumstance.

Mustafa Suleyman, co-founder of DeepMind and now CEO of Microsoft AI, said it concisely: "Safety can't be an afterthought, but it also can't be a moat." He's right. But right now, it's both. It's an afterthought for the companies powerful enough to rewrite their own rules, and a moat for the companies big enough to afford compliance. Everyone else gets squeezed too small to absorb the costs, too far from the regulatory centers to benefit from the protections.

The people most vulnerable to AI harm users in under-regulated markets, patients using unvetted health apps, workers subject to algorithmic hiring tools built without bias audits are the ones least protected by the current safety regime. The wall isn't protecting them. It's protecting the companies on the other side of it.

The Take

AI safety is real. Full stop. The risks are not theoretical. Models that generate medical misinformation, perpetuate bias in hiring, or enable synthetic fraud are not hypothetical they exist, right now, causing measurable harm. Anyone who tells you safety is just a corporate PR exercise hasn't been paying attention.

But.

When the company that built its entire brand on safety abandons its core commitment under Pentagon pressure and then lobbies for regulations that would hamstring its competitors that's not a safety program. That's a business strategy wearing a lab coat.

When compliance costs are identical for a solo developer and a trillion-dollar corporation that's not proportional regulation. That's a barrier to entry with good PR.

When the fastest-growing AI ecosystems on the planet operate entirely outside any safety framework that's not a success story for open source. That's a gap in the infrastructure that no one has figured out how to fill.

"Responsible AI" has become one of those phrases that means everything and therefore means nothing. Every major lab uses it. Every policy document invokes it. But responsibility without consistency is just marketing. And right now, the pattern is consistent only in who bears the cost: not the companies with the most resources, but the builders with the least.

Safety shouldn't be a luxury good. It shouldn't scale with your legal budget. It shouldn't be something that Anthropic practices until the Pentagon calls, or that the EU mandates without asking whether the mandate is survivable for a company with three employees.

The hard work the work nobody wants to fund is building safety infrastructure that's accessible, proportional, and honest about its own limitations. Shared evaluation tools that don't cost six figures. Tiered compliance frameworks that distinguish between Google and a garage. International standards that reach the developers who are actually building in the global majority, not just the ones within driving distance of Brussels.

None of that exists yet. And until it does, "responsible AI" will remain what it mostly is today: a phrase for press releases, a line item for the companies that can afford it, and a door that closes on everyone else.

The wall isn't just about talent, or tools, or infrastructure. It's about who gets to build safely and who gets told that safety is someone else's problem.

Next week, Part 5: Bridges what it would actually take to lower the wall. Not the platitudes. The policies, the money, and the uncomfortable trade-offs that no one in power wants to make.