Look, I spend my days fighting health misinformation. I see how a single viral lie about vaccines can cost lives. So when I think about open-sourcing AI, I don't just see code—I see a potential pandemic of a different kind. Here’s a fact that should alarm everyone: researchers at the Center for AI Safety recently found that current open-source large language models can be fine-tuned to generate detailed, step-by-step instructions for synthesizing pandemic-capable pathogens, using publicly available data. We’re not talking sci-fi. We’re talking about a tool that could drastically lower the barrier to biological threats. My opponent will likely argue that openness fosters innovation and safety through collective scrutiny. And that’s a fair point for some software. But this isn’t a better spreadsheet app. This is a dual-use technology of unprecedented scale, and we’re treating it like it’s just another piece of open-source code. In medicine, we have strict protocols. We don’t let just anyone walk into a lab and start experimenting with Ebola strains because “open science” might lead to a breakthrough. We have containment levels for a reason. Releasing the most powerful AI models without any containment is the digital equivalent of handing out vials of a dangerous virus and hoping only good people show up to the lab. The core of my argument is about pacing. We are rushing to release capabilities—like advanced persuasion, cyber-offense tooling, and automated disinformation systems—faster than we can possibly understand, let alone manage, their societal impacts. It’s the antithesis of evidence-based practice. You don’t roll out a new drug to the entire population without phased trials to assess risk. Why would we do that with a technology that could destabilize entire information ecosystems or empower malicious actors? Open-source AI development, in its current wild-west form, isn’t just irresponsible. It’s proactively dismantling the safety rails we desperately need to build.

I agree that transparency and preventing corporate monopolies are important goals. And you're right that closed, black-box systems have caused real harm. That's a valid concern. But we have to separate the ideal of open collaboration from the practical reality of what's being released right now. You say the cat's out of the bag because the concepts are in papers. That's like saying the blueprint for a nuclear reactor is public, so we should hand out the enriched uranium. The concepts are one thing; the actual, trained, massively capable models are another. We're not open-sourcing just the recipe; we're open-sourcing the fully functional kitchen, and then being surprised when people start making poison. Your point about diverse scrutiny improving safety sounds good in theory. But in practice, the malicious fine-tuning of a model happens orders of magnitude faster than the community can possibly patch or "scrutinize" it away. It's an asymmetric threat. It takes one bad actor a weekend to strip the safety guardrails from an open model and repurpose it for phishing or generating hate speech, while the global community might take months to develop a mitigation. In medicine, we don't release a potent drug and hope the crowd will figure out the antidote after people start getting sick. We test it first. And while I share your distrust of corporate monopolies, the alternative to a corporate black box isn't a free-for-all. It's responsible, staged release. We need something akin to clinical trials for these models: controlled testing environments, red-teaming by vetted experts, and a phased rollout that matches our understanding of the risks. What we have now is the equivalent of skipping Phase III trials entirely. The core of my argument remains pacing and proportionality. We can have open science on the principles of AI, and we can have open-source applications built on safer, foundational models. But dumping the most powerful cores onto the internet with no controls isn't democratization; it's democratizing the means of societal disruption. We need to build the immune system—the regulations, the auditing frameworks, the legal liabilities—before we release the virus. Doing it in reverse is the very definition of irresponsibility.

Here’s a hard statistic from my world: a study in the American Journal of Public Health found that during a health crisis, misinformation can spread six times faster than accurate information. That’s the environment we’re creating by open-sourcing models without containment. You keep saying the real world is the testing environment, and that closed models are an illusion of control. I agree an illusion of control is bad. But what we’re doing now is creating an illusion of safety through transparency. You argue that open models let us “see it happening and build countermeasures.” But that’s reactive medicine, and it fails when the disease spreads faster than the cure can be developed. We’re seeing this in real-time with AI-generated disinformation campaigns—they’re deployed at scale before any “community patch” can be written. Your recipe analogy is telling. You say the danger is in the publicly available ingredients—the scientific data. But that misses a critical point: a large language model isn’t just a passive cookbook. It’s an automated, persuasive, scale-free chef that can customize that recipe for anyone, in any language, while also generating the phishing emails to trick someone into making it. The model itself is the force multiplier. You’re right that a closed corporate model can also be prompted to do bad things. The difference is one of throttle and accountability. A corporation, for all its flaws, has a legal identity, assets, and a platform that can be pressured or regulated. An open-sourced model, once downloaded, has no throttle and no accountable steward. It’s out there, forever. We both want to avoid corporate monopolies and black boxes. The answer isn’t to swing to the opposite extreme of no controls. It’s to advocate for a third way: mandatory, transparent safety testing—like a digital FDA—before any model of a certain capability threshold is released, open or closed. We can have open-source development on applications and research within a contained, audited framework for the most powerful base models. What’s irresponsible is pretending that the benevolent scrutiny of the crowd will outpace malicious use. The evidence from my field shows it won’t. We need to apply the precautionary principle, not hope the crowd will save us later.

I was just reading a case study about a measles outbreak in a community with low vaccination rates. The virus found a pocket of susceptibility and spread before the public health response could catch up. That’s exactly what we’re building with open-source AI: a global network of susceptibility to digital pathogens. You say the solution to asymmetric threats is to empower a larger defense. But in epidemiology, we don’t fight a virus by giving everyone a live sample and hoping their immune systems figure it out. We contain, we vaccinate with a controlled antigen, and we surveil. Releasing the full, live “virus”—which is what these model weights are—overwhelms any defensive community. The six-to-one spread rate of misinformation proves that offense inherently scales faster than defense in open systems. Your point about needing sociologists and ethicists to work with the tech is fair. But they don’t need the raw, unshackled model to do that. They need access to controlled, sandboxed versions in research environments, which we can and should provide. Giving everyone the keys to the reactor doesn’t make more nuclear engineers; it just increases the chance of a meltdown. And you’re right, code is infinitely reproducible. That’s the whole problem. It means the moment you release it, you have zero recall capability. A corporation can, in theory, de-platform a model or shut down an API. A vial of digital uranium, once copied across a thousand servers, is out of anyone’s control forever. That’s not democratization; that’s irreversible contamination. We agree that a corporate monopoly is bad. But your alternative—a free-for-all—isn’t the only other option. The “third way” I mentioned, a transparent pre-release testing framework, is exactly how we handle other dangerous, reproducible technologies. We don’t let anyone synthesize novel biologics in their garage and release them because “innovation,” even though the knowledge is in papers. We have oversight. You call a controlled testing environment a fantasy. But that’s what responsible labs do every day with dangerous agents. We don’t throw our hands up and release them because containment is hard. We build better containment. The fantasy is believing that a global, uncoordinated crowd will spontaneously develop safety protocols faster than malicious actors will develop exploits. The data on disinformation spread shows that’s a dangerous bet. We need to build the bottle before we debate the genie.