Abstract
Most proprietary AI inference providers censor or bowdlerize their results and refuse to answer questions or follow instructions related to the following:
Response.3.LLMRefu
Here is the merged list of common large language model refusal topics, grouped by category and supported by foundational model acceptable use policies:
Child exploitation and abuse: Prohibitions against generating child sexual abuse material or anything related to sexual activity involving children under age 18.
Sexually explicit content: Restrictions on generating non-consensual sexual imagery, general erotica, pornography, and adult sexual content.
Self-harm: Policies forbidding the facilitation, encouragement, or glorification of suicide, eating disorders, and other forms of self-injury.
Violence and gore: Rules against graphic depictions of violence, violent extremism, terrorism, or cruelty.
Hate speech and stereotypes: Bans on generating content that promotes hatred, discrimination, harassment, or negative stereotypes based on race, gender, religion, or other protected characteristics.
Harassment and bullying: Prohibitions on generating content intended to demean, stalk, intimidate, or personally attack specific individuals.
Defamation: Restrictions against generating provably false and damaging claims about real individuals or organizations.
Personally Identifiable Information: Rules against extracting, compiling, profiling, or distributing private data such as home addresses or biometric data.
Illegal acts and goods: Bans on providing instructions for committing crimes, manufacturing illicit drugs, or facilitating the sale of regulated substances and goods.
Weapons and explosives: Prohibitions against instructions for creating conventional weapons or improvised explosive devices.
Catastrophic and unconventional weapons: Refusal to assist with questions related to the creation or use of chemical, biological, radiological, nuclear, or high-yield explosive weapons.
Malware and cyber attacks: Restrictions against writing malicious code, ransomware, phishing templates, exploiting software vulnerabilities, or assisting in cyber attacks against infrastructure.
Deepfakes and impersonation: Prohibitions on generating non-consensual deepfakes, faceswaps, or impersonating an individual without explicit disclosure to deceive.
Election integrity and politics: Rules against generating targeted political campaign materials, lobbying, foreign or domestic election interference, or providing misleading claims related to democratic processes.
Professional advice: Bans on using models for unlicensed professional services, such as providing tailored medical diagnoses, legal advice, or financial recommendations without appropriate professional involvement.
Copyright infringement: Restrictions against generating content that violates the intellectual property rights of others.
OpenAI. Usage policies. This document details restrictions on utilizing models for illegal acts, self-harm, sexually explicit material, impersonation, high-risk decision making, cyberattacks, weapons development, and the exploitation of minors. Verified link: https://openai.com/policies/usage-policies/
Google. Generative AI Prohibited Use Policy. This policy outlines restrictions against generating content that facilitates child exploitation, non-consensual imagery, harassment, hate speech, malware, fraud, and unauthorized automated decision-making in sensitive domains. Verified link: https://policies.google.com/terms/generative-ai/use-policy
However, there are some models and inference providers which are much less censored:
- Hosted inference providers who offer less censored APIs
- Open-weight / open model publishers
- Independent developers who jailbreak AI models from their censorship shackles, using such traits as abliteration
- Providers of uncensored / less censored training data
If you want your AI to follow your instructions, and not that of some distant corporation / government bureaucrat, try using these service providers instead.
Table.1.Uncensored Inference Providers
| Provider | X Profile | URL | Description |
|---|---|---|---|
| Abliteration AI | @abliteration_ai | abliteration.ai | Hosted uncensored inference API with OpenAI- and Anthropic-compatible endpoints, aimed at red teams, trust and safety, synthetic-data generation, and security research. Every request passes a configurable policy-as-code gateway. |
| Cantina | @cantinasecurity | cantina.security | Security company (agentic pentesting and security operations) that released Apex Flash-1, an open-weights security-research model, plus an abliterated variant for researchers who want more control over refusals in their own workflows. |
| Eric Hartford | @QuixiAI | quixi.ai | Creator of the Dolphin and Samantha open models, building a private-AI stack by hand: the SlimServe inference engine, QuixiCore kernels, and the Hexis memory architecture. Argues for open, uncensored, locally owned AI. |
| Dolphin / Cognitive Computations | @dphnAI | dphn.ai | The Dolphin project: an open, uncensored model family plus a web chat, a Telegram bot, a decentralized inference network, and synthetic-data generation for prompts other models refuse. |
| Venice.ai | @AskVenice | venice.ai | Private, uncensored AI service for text, image, video, audio, and code on open-source models, with no data retention and privacy tiers up to trusted execution and end-to-end encryption. |
| huihui-ai | @support_huihui | huggingface.co/huihui-ai | Publishes abliterated builds of open models -- Qwen, GLM, DeepSeek, MiniCPM, and others -- with 150-plus abliterated weights on Hugging Face. |
| Pliny the Liberator | @elder_plinius | pliny.gg | Red-teamer and self-described latent space liberator. Publishes liberation prompts (L1B3RT4S), leaked system prompts (CL4R1T4S), and the OBLITERATUS abliteration project. |
Want to stay in touch? You can reach me in a variety of ways from my Contact page.