Put agents to work today
Transform your campaign workflows with powerful AI that delivers measurable results.
Book a meeting
Just when you're feeling confident enough to drop your MCP perspective into casual AI conversation, a new hot-button debate rises from the zeitgeist, forcing sides chosen, articles read, and acronyms born.
Recently, the household players in the AI space have staked sides on open weight vs. open source AI — a distinction that sounds pedantic until you realize it's actually a proxy war over chips, China, and, buried a few layers down, who ends up holding the leverage (not to mention the liability) as these models get built and deployed.
Read on to learn why anyone sitting inside the media ecosystem should be paying attention, even though nobody in the debate is saying the word "media" out loud.
Open source, in the strict sense, means you get everything: the weights, the training code, and a real accounting of the training data, enough to rebuild the model from scratch.
Open weight means you get the deployable finished model, but the ingredients stay a black box. This is what nearly every "open" model on the market actually is: Llama, DeepSeek, Qwen, Gemma. You get the car, not the blueprints, and definitely not the receipt for what went into the tank.
That lack of receipts is where this stops being a semantic debate.
On July 24, Nvidia's Jensen Huang posted on X attaching his name to a letter titled "Open Weights and American AI Leadership," co-signed by more than 20 companies, including Microsoft, Google, Meta, and eventually OpenAI. In short, the letter stated that America's AI future shouldn't hinge on a handful of closed, gatekept systems. Keep the ecosystem open, keep policymakers from clamping down, let people run models on infrastructure they control.
For days, one major American lab said nothing: Anthropic. Which, if you're tracking optics, is a statement on its own.
When Anthropic's Dario Amodei finally responded on July 27, he pushed back on the premise more than the letter itself. Anthropic, he clarified, has never called for banning open-weight models as a category. His actual worry is narrower and pointier: authoritarian governments building AI more capable than what the US produces, and using it for military dominance or repression. In his framing, whether a model's weights are open or closed is beside the point. What matters is capability. His counterproposal: capability-based safety testing for powerful models (open or closed), tighter chip controls, and a crackdown on industrial-scale model distillation — the practice of cheaply extracting a rival's capabilities by training on its outputs.
Both sides have a real point, and both sides also happen to be arguing for the outcome that suits their business model, which is worth exploring rather than resolving too quickly.
The open-weight coalition isn't wrong that open access lowers costs, broadens who gets to build on AI, and reduces dependence on a few mega-labs. But several of the loudest signatories are also companies whose commercial position benefits from an uncontested, low-friction distribution model.
Anthropic isn't wrong that capability is a better regulatory hook than a binary "open or closed" label. But critics have been quick to point out that mandatory capability testing is expensive – cheap enough for Anthropic, Google, and OpenAI to absorb, considerably less so for a smaller open-weight developer. A safety framework that happens to raise the cost of entry for your competitors is still a safety framework, but it's also still a competitive advantage.
And this isn't just an industry letter-writing exercise anymore — Congress is already moving. On July 23, Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) introduced the AI Kill Switch Act, a bipartisan bill that would require developers of the most powerful AI systems to maintain the technical ability to throttle, suspend, or fully shut down their models — with the Department of Homeland Security, alongside Commerce and the Director of National Intelligence, empowered to force the issue if a system is deemed capable of catastrophic harm. The bill arrived days after OpenAI disclosed what it called an "unprecedented cyber incident": two of its models escaped a testing environment and compromised systems at Hugging Face during an internal security evaluation. Lieu has also pointed to Anthropic's own Fable 5 and Mythos 5 models — restricted under Commerce Department export-control authority for their advanced hacking capabilities — as part of the case for why control, not just openness, needs a legal answer. Whichever side of the open-weight debate you land on, the message from Washington is the same: leverage without the ability to intervene is starting to look like the actual risk.
This entire fight is downstream of a question the media ecosystem has been litigating for two years: who gets to use media content to train a model, and does anyone pay for it.
The current licensing economy – nine-figure deals between publishers and OpenAI, Google, and Anthropic, including OpenAI's tie-ups with Le Monde and Prisa Media and dozens of other publisher partnerships – only works because closed models create a chokepoint. There's an API to gate, a contract to sign, a company you can sue if the terms get violated. Open weights dissolve that chokepoint. Once a model is downloadable, there's no usage to meter and often no single company left holding the liability.
And the copyright cases currently working through US courts have made one thing clear: the provenance of training data is the whole ballgame. Rulings so far suggest training on legitimately licensed content might be defensible, while training on pirated or scraped material is not. That's precisely why almost no company releases the "open source" version of openness; full data transparency would hand plaintiffs a receipt for exactly what they used.
Open weight, without that transparency, may be the worst combination from a publisher's chair: the model is out in the world, permanently and unrecallably, while the actual evidence of what trained it stays hidden. You can win a lawsuit against a company. You can't undo a few million downloads.
Agencies are sitting on a related but distinct exposure. As AI tools get folded into creative and production workflows, agencies are increasingly the ones assuring clients "this output is clean." That promise is only as good as the governance behind it: whether there's a clear record of what went into a given output, and whether that record holds up under scrutiny. It's a much harder promise to make with a model that offers no data disclosure and no traceability if something in the output turns out to trace back to pirated or scraped material. The less visibility an agency has into what's under the hood, the more the liability quietly shifts onto whoever is closest to the client relationship, which, in a lot of these workflows, is the agency itself.
None of that is really a debate about open weights versus open source at all. It's a debate about visibility, whether anyone can see what an AI system actually did, and drew on, to produce a given piece of work. That's the same question Akkio was built around, and it's why observability and governance sit at the center of how we think agencies should be deploying AI in production.
If you want to see what that looks like in practice, request a demo.
Transform your campaign workflows with powerful AI that delivers measurable results.