News

Yann LeCun vs. AI Obscurantism: Why Open-Source Is the Only Defense

2026-07-01 • 6 min read

There is a scene that repeats. A company launches an AI model, declares it "too dangerous" to release publicly, and then charges for access. The script hasn't changed since 2019. Yann LeCun —VP and Chief AI Scientist at Meta, one of the three founding fathers of modern deep learning— has been saying for years that this script is propaganda, not science.

His latest public appearance, in a discussion with Steven Levy at VivaTech, was blunt: "If you block a tool because you think it's very dangerous, you are in medieval obscurantism."

This isn't a LinkedIn quote. It's a statement of principle about who should control AI and, by extension, who should control knowledge in the next century.

The Fear Narrative as a Commercial Strategy

The recent trigger is Claude Mythos, Anthropic's model presented in April as "the best ever created" and restricted to a closed circle. The company cited security risks. The US government added access restrictions by nationality. The message: this AI is so powerful that only a few can touch it.

LeCun summarizes it without embellishment: "Anthropic and a few others are a lobby that essentially wants AI not to be open source because they believe it is intrinsically dangerous technology."

The question is whether they truly believe it's dangerous or whether the danger narrative suits their business model. Close the code, control access, charge the subscription. The "it's dangerous" works as a barrier to entry disguised as ethics.

LeCun compares it to the Catholic Church calling Galileo a heretic: they give you the tool, they take it away, and they tell you it's for your own good.

AI as a Souped-Up Wikipedia

For LeCun, today's AI is not an omnipotent oracle or an existential threat. It is, in his own words, "a way to disseminate knowledge." Language models allow people to easily access repositories of human knowledge. That is, today, what AI does: democratize access to knowledge.

The analogy is deliberate. Wikipedia wasn't shut down because it could be used for disinformation. It stayed open because shared knowledge benefits more people than it harms. Open AI follows the same logic.

"If you buy a pen, you don't want the company that sells it to tell you what you can write with it," LeCun said. The metaphor is simple but the conclusion is radical: AI tools cannot depend on a handful of companies deciding what you can do with them.

When Regulation Protects the Incumbent

The European debate is where obscurantism translates into legislation. The EU AI Act, approved in 2024, debates how to regulate foundation models. Some countries push to make open-source models illegal or burden them with obligations only giants can meet.

Mistral's official position —the French startup that alongside Germany's Aleph Alpha defends open-source in Europe— was summarized by CEO Arthur Mensch on X:

Regulate products, don't regulate technology. Promote open source foundational models. At the very least, don't regulate them in ways that favor incumbents.

In other words: regulate the final product (a hiring system, a medical assistant, a credit scoring tool), but not the underlying model. Regulating the base technology is like regulating the combustion engine instead of the car. The result is that only those who already have an engine can compete.

LeCun warns what happens when you close the code: "When you do research in secret, you fall behind. The rest of the world will go open source and will overtake you." The example is right there: DeepSeek, the Chinese startup that shook markets with an open-source model matching the best at a fraction of what US giants invest. It was built on PyTorch (Meta) and Llama (Meta). On open code.

The Dependency Trap

Behind the technical debate there is a geopolitical one. If Europe bans or suffocates open-source AI, it condemns itself to depending on three or four US or Chinese companies for its entire cognitive infrastructure. LeCun was explicit:

"We cannot afford to have those systems come from a handful of companies from the West Coast of the US or China."

The dependency isn't just about cost. It's about sovereignty. If your AI runs on a remote API, your data leaves, your provider can cut the tap, change terms, raise prices, or decide your use case is no longer acceptable. OpenAI's history —founded as open-source and turned into a closed company— is the manual on how to betray that principle.

Open-source is not an ideological stance. It is a condition of sovereignty: the ability to audit, deploy, modify, and maintain without asking permission.

What LeCun Doesn't Say (and That Also Matters)

LeCun is right on the essentials, but there are nuances an honest post cannot omit.

Open-source models can also be dangerous. An unrestricted model can be used for disinformation at scale, social engineering, or offensive cybersecurity. The difference isn't that open-source is harmless; it's that the danger is manageable when the model is transparent, auditable, and modifiable by the community.

Product regulation is necessary. Mistral itself acknowledges this. An AI system that decides whether you get a loan or a medical diagnosis cannot operate without oversight, open-source or not.

Unequal competition is real. Meta funds Llama with its advertising empire's revenue. DeepSeek has the backing of the Chinese state. Pure garage open-source exists, but competing at the foundation level requires enormous resources. Open-source doesn't eliminate power concentration; it redistributes it.

Neurosint's Conclusion

At Neurosint we defend open-source not because it's cheap, but because it's the only architecture that preserves control. When we deploy a model on a client's infrastructure, that client can:

Audit what the model does.
Modify its behavior without asking a provider for permission.
Maintain the system if the original provider disappears.
Protect their data within their perimeter.

The fear narrative LeCun criticizes isn't just an academic debate. It's the argument used to justify closure, dependency, and lock-in. And as LeCun said, with the patience of someone who has spent decades hearing the same speech: that's not prudence. It's obscurantism.

Sources: