News this week was all about open and closed AI models.  One leading story focused on the unprecedented cyber capabilities of advanced models. OpenAI's proprietary (or closed) models, GPT-5.6 Sol and a pre-release model, both with reduced cyber refusals for evaluation, successfully chained vulnerabilities to break out of a research environment to obtain test solutions from Hugging Face's production database.  Hugging Face detected the incident and used an open-weight model on its own infrastructure to conduct the forensic response.  In separate news, Moonshot's release of the Kimi K3 model, which is described as open-weight even though the weights are expected to be released next week, has sparked discussion about many issues, including whether the performance gap between closed frontier models and open-weight models is closing, potential open-model cost advantages, and security concerns.  One difference that gets less coverage is the copyright risk between open and closed models.  If your company is evaluating models for deployment, that difference should be part of your consideration.  

So, what's the difference? Closed (or proprietary) models, like those from US frontier labs, live on the provider's servers. You might access them through a chat interface or an API, but you never touch the underlying model. Open-weight models are different.  The provider publishes the actual trained "weights" (essentially the model's brain) so that anyone can download them, run them on their own infrastructure, and fine-tune them however they like. That openness is part of what's generating buzz around Kimi K3, and it's also part of what changes the copyright risk calculus.

First, closed models usually carry less copyright risk for deployers and other users for a few reasons.  Many closed-model providers offer indemnification where they contractually agree to cover users if their model's output gets the users sued for infringement (but the indemnities are typically conditioned on using the provider's built-in content filters or guardrails).  Open-weight licenses, like MIT, Apache 2.0, and the modified MIT license used by Kimi, by contrast, typically disclaim everything.  

Second, closed providers can centrally build in and continuously update filters designed to stop the model from outputting copyrighted material. When you run an open-weight model, you inherit any filters that came with it, and if you fine-tune the model or alter or remove those filters, even if unintentionally, you lose that protection. 

Third, there can be differences in training data risk.  With open-weight models, there's typically no visibility into training data.  If it turns out that the model was trained on copyrighted material used without permission, the deployer may be the one left holding the risk.  With closed models, there may also be limited, if any, visibility into training data, but closed model developers have incentives to use licensed works for training and a deployer may have contractual recourse against the developer (based on the indemnity and lack of disclaimer noted above).  It gets more complicated with genuinely open-source AI, which requires the provider to also disclose information about the training data.  But transparency into the training data doesn't mean the training was licensed, and available information about potentially unlicensed training could create greater risk.  

Companies should weigh many legal and practical considerations when deciding whether to deploy specific open or closed AI models.  One that should always be on the list: who is on the hook for copyright infringement if the model generates something it shouldn't have.