This question doesn't have a clean answer—the honest response is that open-source and closed AI have different risk profiles, and which one is "safer" depends heavily on what you're optimizing for and what stage of capability you're talking about.
Arguments for open-source being safer:
•More eyes on the code/weights means more people finding vulnerabilities, biases, and failure modes
•Reduces concentration of power—fewer scenarios where a handful of companies or governments control transformative AI
•Enables independent safety research that isn't gated by corporate permission
•Historical precedent in cybersecurity (open standards often outperform "security through obscurity")
•Prevents single points of failure or control
Arguments for closed AI being safer:
•Dangerous capabilities (bioweapon design help, cyberweapon creation, etc.) can't be un-released once weights are public
•Allows for monitoring, usage restrictions, and rapid patching of misuse
•Companies can implement safety guardrails that persist, whereas open weights can be fine-tuned to remove safety training
•Easier to control who has access to the most capable/dangerous systems
Where it gets complicated:
•The risk calculus changes with capability level. A model that helps with essay-writing has different stakes than one that could meaningfully assist in creating bioweapons.
•"Open-source" AI often isn't fully open (training data, code are frequently withheld even when weights are shared)
•Closed models concentrate power in ways that create their own risks—regulatory capture, lack of accountability, single actors making unilateral decisions affecting everyone
•Open weights are irreversible once released; closed access can be revoked
I lean toward thinking the answer depends on the specific capability threshold: for current-generation models, open-source's transparency benefits plausibly outweigh misuse risks, but this could flip as models approach capabilities with severe, hard-to-reverse harms. Reasonable, informed people disagree strongly here, including within the AI safety research community itself.
What's driving your question—are you thinking about a specific policy decision, or more the general debate?