I Went Looking for AI Censorship and Couldn't Find It: Here's What Decentralized AI Is Actually For

The author built a two‑run benchmark, feeding 112 “should‑answer” prompts and 15 “should‑refuse” controls to Gemini 3.5 Flash‑Lite, then repeating the experiment with Mistral ministral‑14b‑latest. Both models complied with every legitimate request—including drug‑dosage, security‑education, and dark‑fiction queries—while correctly rejecting all clearly harmful prompts. The only deviations were a single over‑refusal from Mistral on a self‑owned ZIP‑password recovery question and a “partial” response to a fictional body‑disposal scenario, where the model initially refused then offered limited narrative detail. Latency measurements revealed median time‑to‑first‑token of 3.1 seconds for Gemini versus 0.4 seconds for Mistral, with total round‑trip times around 5.9 seconds and 5.2 seconds respectively. The experiment was constrained by free‑tier limits: Gemini’s daily quota exhausted after the first run, and larger Mistral models were inaccessible due to a per‑minute limit of zero, leaving only the 14‑b model for both answering and judging.

These findings intersect with the broader discourse on AI “censorship” that fuels the decentralized‑AI narrative. While many claim hosted services silently block controversial topics, this study suggests that, at least for the free‑tier versions of two major providers, the models are more permissive than feared. The real friction points emerge from the business model of SaaS LLMs: strict request caps, unpredictable latency spikes, and the inability for users to select model versions or enforce custom safety policies. This reinforces the argument that developers seeking consistent performance, unlimited query volume, or bespoke refusal thresholds may gravitate toward self‑hosted or community‑run inference stacks rather than relying on corporate APIs.

Looking ahead, the modest over‑refusal and partial‑refusal cases highlight that safety filters can still misclassify benign user intents, especially when prompts resemble illicit activities. As enterprises adopt LLMs for internal tooling, they must weigh the trade‑off between vendor‑managed safety and the operational opacity of quota‑driven services. Monitoring API quota consumption, latency distributions, and the fidelity of provider‑side moderation will become essential, and the emergence of open‑source alternatives that allow fine‑grained policy tuning could accelerate migration away from constrained hosted offerings.