The Lesson of Fable: Own Your Governance Layer

There are moments in technology when the future does not arrive quietly.

It arrives with a loud press release. It arrives with a stock that jumps 19% on the first day of trading. And yet sometimes, it arrives with a shutdown.

Yesterday Anthropic announced that the U.S. government had issued an export-control directive requiring it to suspend access to Fable 5 and Mythos 5 for foreign nationals, which the company said forced it to disable those models for every customer in order to comply. Anthropic said the directive appeared to relate to a narrow potential jailbreak, and it disputed that such a finding justified recalling a commercial model at that scale. I do not know where that dispute will land, however, regardless of what transpires, I want to get us back to a basic, fundamental principle: we need to own our AI governance layer.

The shutdown is a stark reminder of what could happen when we don’t own or cannot control our AI governance layer.

In this blog, I define AI governance as the set of policies we want to apply as users of an AI system: rules about security, privacy, data retention, transparency, regulatory compliance, and the kinds of work the system is allowed to do for us, etc. It is the part that should be specific to our business and our risk, not generic to everyone.

State-of-the-art AI systems use general-purpose frontier models, and so is the governance baked into them. My concern is not that these models’ governance is weak. From what I have seen they are strong, and the companies building them take the work seriously. My concern is that the governance we depend on is decided by one or a few, for everybody, and handed to us as a finished thing we cannot change.

Let’s set aside the politics and the national-security headlines for a moment, and look at this through some examples of ordinary, day-to-day legal work:

• An AI tool that we build a contract-review workflow on now declines to redline indemnification clauses in a contract, because the provider’s safety team decides that the tool can no longer provide “legal advice”.

• A content filter that protects a consumer chatbot strips the profanity out of the depositions in a litigation.

• A rate limit change prompts a silent downgrade to a lower-quality model that results in case law hallucination.

While many of these backend changes could technically be adjusted or configured, the reality is that most users are not aware that a change occurred, let alone how to prevent or fix it. These are the consequences of a vendor’s policies applying to everyone at once. We are left in a position where we didn’t build the rules, and therefore we couldn’t be accountable for bad things when they happen. Hence the finger-pointing begins, and that is not a good place to be in for any one.

AI governance needs to sit at an independent layer, not beholden to a single provider. When I bring this up in conversations with other policy and legal leaders, I am often met with a blank stare. It took a while for me to realize that the current generation of professionals grew up in an era where building on someone else’s land is the only business model we have known.

To most of us, infrastructure is, by definition, closed. Social media is a great example of businesses building their audience on someone else’s land, and then watching the rules change underneath them. The app store taught the same lesson, where a few platforms decide what is allowed and what cut they must take. A generation has nearly forgotten what open infrastructure looked like, because nearly every service now sits behind one dominant provider and wears one brand name. We have come to assume that a capability and a company are the same thing, that to use the function we must accept the platform.

It does not have to be that way.

So, what does owning our governance layer look like in practice? It starts with open infrastructure.

Take the examples of open-protocol social media networks like Mastodon (built on ActivityPub) or Bluesky (built on AT Protocol). These emerged because creators realized that being locked into a closed network means their reach could be cut off overnight by the “algorithms”. The same trend to build alternative AI infrastructure is emerging today.

Driven by the desire to manage token costs, prevent vendor lock-in, and ensure uptime, innovative teams are building inference middleware and smart routing gateways that allow users to switch seamlessly across different open-source and proprietary models. The natural extension to cost saving for inference and gateway providers would be to build the governance layer that users can own.

To own our governance layer, we must move it outside the walls of any single provider. This means as the users, we have the autonomy not just to define policies, but to iterate on them, track their impact, and evaluate model performance against our internal standards. When we shift from “accepting defaults” to “defining rules,” we regain the agency to understand exactly how and why AI behaves the way it does. Ultimately, this is about accountability. We cannot be held accountable for outcomes we do not control, and we cannot control outcomes when we do not own the architecture that governs them.

I hope the new generation working with AI does not see concentration as the natural state. The future worth building is an ecosystem where solutions of all kinds can flourish, where small teams can build durable tools without betting their entire company on one single provider, and where governance resides with the people accountable for it.

We can achieve this by:

• Defining governance policies based on our unique risks, rather than adopting vendor defaults.

• Keeping controls in a layer we own, ensuring we understand them and can be accountable for complying with them.

• Designing for vendor-neutral interoperability from day one, making model swapping a routine event rather than a crisis.

Fable may return. Another model may surpass it, another shutdown may come, and another breakthrough will certainly arrive.

We don’t own the model. But we should own the governance layer built on top.

For more practical tips on AI governance and innovation, check out GenAI for the Legal Profession: Power User Edition, AI Strategy for Legal Leaders, Atticus AI Habits Workshop and my Fairly AI blogs.