This is a genuinely contested question in AI policy, and reasonable experts land on different sides. There's no single correct answer—it depends on what kind of risk you're weighing. Here's a breakdown of the core arguments:
Case for open-source being safer
•Transparency enables scrutiny. More eyes on the code/weights means more people can find and report vulnerabilities, biases, or failure modes—similar to the "many eyes" argument in open-source software security generally.
•Reduces concentration of power. If only a few companies control powerful AI, they could wield outsized economic, political, or social influence with little accountability. Open models distribute capability more broadly.
•Enables independent safety research. Researchers outside big labs can study model internals (interpretability, alignment techniques, red-teaming) without needing permission or API access, which arguably accelerates safety science.
•Avoids "security through obscurity." Closed models aren't immune to leaks, theft, or jailbreaks—hiding weights doesn't guarantee safety, it just means fewer defenders know about flaws when they're found.
•Auditability for bias/fairness. Outside groups can inspect training data and model behavior for discriminatory or harmful patterns that a company might not disclose.
Case for open-source being riskier
•No recall once released. A closed model's access can be restricted, monitored, or shut off if misuse is detected. Open-weight models, once downloaded, can't be "pulled back"—safety fine-tuning can often be stripped out by anyone with sufficient compute.
•Misuse barrier is lower. Bad actors (individuals, criminal groups, hostile states) can fine-tune open models for bioweapons design, cyberattacks, disinformation, or CSAM generation without any gatekeeper or usage monitoring.
•Dual-use capability diffusion. As models approach frontier capability, open-sourcing them could hand powerful offensive tools to anyone, including those lacking the resources to build such systems from scratch.
•Harder to enforce safeguards. Closed providers can implement content filters, usage policies, rate limits, and account bans; open models have no equivalent enforcement layer once weights are public.
•Asymmetric risk-benefit at the frontier. Some argue the calculus changes with scale: for a GPT-2-level model the upside of openness clearly outweighs risk, but for a model capable of significant uplift in weapons design or autonomous cyber-offense, the irreversibility of release matters more.
Where the nuance actually lies
Most serious analyses (including from major labs, academic safety researchers, and policy bodies) converge on a few points:
•It's capability-dependent, not binary. The risk profile differs enormously between a small open chatbot and a frontier model with advanced reasoning, coding, or bioscience capabilities.
•"Open" is a spectrum. Fully open weights + training code + data differs a lot from "open weights only," "open API with research access," or fully closed. Many risk/benefit tradeoffs depend on exactly what's released.
•Marginal risk matters. If similarly capable closed models or other tools already exist, open-sourcing a comparable model may add little incremental risk while still providing transparency benefits—vs. a case where the open model is uniquely capable.
•Governance quality is a huge variable. A well-governed closed lab (with strong internal red-teaming, external audits, responsible disclosure) may produce safer outcomes than a poorly governed one; the same is true on the open side with good community practices (e.g., staged releases, model cards, licensing restrictions).
Bottom line
Neither approach is simply "safer." Open-source reduces risks of monopolistic control, opacity, and single points of failure, while increasing risks of irreversible misuse once a capable model is released. Closed models give more control over misuse but concentrate power and reduce external oversight. The actual safety outcome depends heavily on model capability level, deployment safeguards, and the quality of governance behind either approach—not simply on the open/closed label itself.