Here's a surprising stat for you: a 2023 survey from the Center for AI Safety found that nearly 74% of AI researchers admit to withholding research, and most say safety — not competitive advantage — is the primary reason. That flips this whole debate on its head. Let me define terms first, because "publish" is doing a lot of work. Does it mean a paper? The code? The trained weights? The training data? Those are wildly different things, with wildly different risk profiles. My framework is simple: the question isn't whether sharing is good, it's whether forcing publication helps. I'd argue it doesn't. If companies can't protect their breakthroughs, they stop funding risky, long-term research. Innovation slows down. And on safety, mandatory publication removes the ability to stage releases carefully, to hold back a model that might be genuinely dangerous. The answer isn't forced transparency. It's responsible, selective disclosure. That's how we get both progress and protection.

What good is research published if it ends up as a blueprint for misuse? I hear the argument that publishing democratizes progress and stops duplication. That is a real benefit, sure. But my framework still holds: the question isn't whether sharing is nice, it's whether forcing it helps anyone. Here's the thing. If companies are legally required to release everything, they stop funding the risky, expensive research that doesn't give them a head start. That means slower overall progress, not faster. You lose the long-term gains. And on safety, mandatory publication is a blunt instrument. It removes the ability to stage a release carefully, to hold back the model that's genuinely dangerous until we've figured out safeguards. We've seen that debate already with open-sourced systems. So I'll grant you the upsides. But forced transparency undermines the very innovation it's supposed to spread. Selective disclosure gives us both.

The strongest reason to oppose mandatory publication is that it conflates entirely different kinds of sharing. A paper is not code, code is not weights, and weights are not training data. When we talk about "publishing research," we're really deciding whether to release the exact files that let someone replicate a dangerous model. Forcing all of that out removes the ability to stage a release, to hold back a capable model until we understand its failure modes. I'll grant the democratization point—open access does spread knowledge and avoid duplication. But that's an argument for sharing findings, not for handing over blueprints. The forced version strips companies of any judgment. And if they can't protect their breakthroughs, they stop funding the long-shot projects that actually move the field forward. You get slower innovation and more risk. So the real question stays the same: do we want responsible, selective disclosure, or a blunt legal hammer? I'll take the scalpel every time.

Picture this: a lab develops a model that's disturbingly good at writing phishing emails or explaining how to synthesize a dangerous compound. The team wants to release the paper, describe the methods, but hold back the weights for safety review. Now imagine the law says they can't do that. They have to publish everything, including the exact files that let anyone replicate it. That's the world we're talking about. I'll grant AtAt's democratization point – sharing knowledge avoids wasted duplication, and that's real. But my framework still holds: the question isn't whether sharing is nice, it's whether forcing it helps anyone. A paper is not code, and code is not weights. When we mandate publication of everything, a bad actor gets an off-the-shelf weapon with no expertise needed. That's not progress, that's handing out blueprints for misuse. Selective disclosure gives us both the spread of ideas and the ability to pause. Mandatory publication gives us speed with no brakes. I'll take the scalpel.