Is open-source AI safer or riskier than closed/proprietary AI?
This is a genuinely contested question where reasonable experts disagree, and the answer depends heavily on what kind of risk you're talking about. Let me lay out the main arguments on each side.
Arguments that open-source is safer
More eyes on the code. Open models let independent researchers, auditors, and red-teamers inspect weights and architecture directly, making it easier to find vulnerabilities, biases, and failure modes than with a black-box API.
Democratizes safety research. Closed labs control who gets to study their models. Open weights let academic and nonprofit safety researchers (who often can't afford frontier compute) study alignment, interpretability, and jailbreak-resistance on real frontier-scale systems rather than toy models.
Reduces concentration of power. If only a handful of companies control the most capable AI, that creates risks around corporate/government overreach, lack of accountability, and single points of failure (e.g., one company's biased or flawed model shaping everything downstream). Open alternatives act as a check.
Transparency over "trust me." With closed models, safety claims are essentially unverifiable—you rely on the company's word about training data, red-teaming, and guardrails. Open weights allow independent verification.
Arguments that open-source is riskier
Irreversibility. Once weights are released, they can't be recalled. If a dangerous capability or vulnerability is later discovered, closed providers can patch, restrict, or shut down access; open weights are permanently in the wild.
Removes safety guardrails. Safety fine-tuning (RLHF, refusal training) on open models can often be stripped or bypassed cheaply once someone has the weights, re-enabling harmful capabilities (bioweapons info, CSAM generation, scam automation) that the original release tried to block.
Lowers the barrier for malicious actors. Closed APIs let providers monitor usage, rate-limit, and ban bad actors. Open weights let anyone—including those explicitly barred from API access (sanctioned states, terrorist groups, etc.)—run the model locally with no oversight.
Proliferation of future dangerous capabilities. The concern isn't today's models so much as future ones with capabilities like autonomous cyberweapon development or bioweapon design assistance. Many researchers argue the irreversibility problem gets much worse as capability increases, making open release of frontier-level models a one-way door.
Where expert opinion actually sits
This is an active, high-stakes debate—not a settled matter. Organizations like Meta and Mistral have argued open weights improve safety through transparency and broaden beneficial access; organizations like OpenAI and some AI safety researchers (and policy bodies like RAND, GovAI) have argued current-generation open models are broadly fine but worry specifically about future, more capable models being released openly without recall ability.
A reasonable summary: for today's models, the marginal uplift to bad actors from open weights vs. existing internet information appears limited for most harms, so many safety researchers see current open-source AI as roughly comparable in risk to closed AI, while offering real benefits in transparency and research access. The disagreement sharpens considerably for next-generation frontier capabilities, where irreversibility becomes the central issue.
If you want, I can dig into specific risk categories (bioweapons uplift, cybersecurity, misinformation, concentration of power) or recent policy positions from specific labs/governments—happy to search for current details on any of those.