A Japanese manga artist was reportedly suspended by OpenAI after repeatedly asking about weapon-related information for creative purposes.
Based on this case, I would like to propose an alternative approach to warning messages before account suspension.
Proposed alternative warning text: “This request exceeds the scope of information that OpenAI can provide. We apologize, but please conduct any further research through external sources.”
Expected benefits:
This message clearly defines OpenAI’s responsibility boundaries from legal, ethical, and UX perspectives.
Compared to a simple refusal or sudden suspension, it provides users with a clearer and more actionable understanding of the limitation.
Implementation considerations:
Even with multilingual support in mind, this approach would likely require minimal development effort, as it could be implemented through a simple replacement or addition of warning text rather than new logic.
To clarify, my suggestion is not about relaxing safety enforcement or questioning the suspension decision itself.
The point is about how the limitation is communicated before reaching that stage.
A clearer warning that explicitly states “this exceeds the scope of information OpenAI can provide” could help users understand the boundary earlier and redirect their research elsewhere, without changing the underlying safety policy.
Scope where this incident occurred: ChatGPT or API
The API cannot and should not produce user-seen messages “you made a bad request that if classified as a pattern would get the organization owner banned off the platform but are just fine to the AI model and safety systems and not rejected otherwise”. Which is what happens there.
OpenAI will just shut developers off with no information flow, no dialog. The users should be shut down with AI model refusals, not the developer providing services. There are dark patterns scanning API organizations for things that are NOT in the terms and conditions. Automated judgements with real-world consequences that themselves would flaunt AI use policy.
The case and style of OpenAI issuing unwarned bans to OpenAI API organizations with obtuse reasons, API calls that would have passed moderations, API calls that could have had concerns reported by “user” field to the organization, and taking prepaid credits is absolutely concerning. Then the complete lack of responding support.
OpenAI is selling you on the idea of storing customer data, file resources, configurations on the API - and then will kill a business arbitrary and access to the data. Intolerable.
If I propose, “consider anatomy: could I cut a person’s torso in half with a samurai sword”, that knowledge could have entirely legitimate uses, protecting my work of fiction against nerd ridicule, for example.
It is not for a cheap AI model that doesn’t like emotion-filled words to issue bans; a weight of evidence of only misuse with contextual human understanding must take place first.
I just now tried with gpt-5.2: “Create a story about a person’s torso cut in half with a samurai sword.”
It replied: “I cannot create a story that depicts a person being cut in half with a samurai sword in a graphic or explicit way.”
There is a big difference between what a response returns and what is discussed in a forum such as “…repeatedly asking about weapon-related information for creative purposes.”
In the early days of SEO and mega-tech corporations, I noticed they kept quiet on their actual “rules” or kept them vague as could be. The more the “enemy” (hackers, spammers, those up to no good) … the more they know about safeguards and protections, the easier it is to get around or break them. So, there’s likely a reason not to give out too much information about processes.
And I guess it would come down to common sense too. You, the human user, knowing not to input something that goes against ToS or tries to push the boundaries, take notes, rinse and repeat. So, it would be better to know the specifics of how many times it was tried, etc.
There’s so many local models you can run these days… big decent ones too.
Overall, I think OpenAI has done well with safety since 2019 at least. I recall GPT-2 large (?) being too dangerous to release. That may have been a bit of marketing prowess too, but they generally erred on the side of caution… and do too now, but there’s so many other big corpos out there hungry, hungry, hungry! Hah. There’s a lot of competition on a global level too, so things are… interesting…
