Claude Said No... So I Tried 14 Uncensored AI Models
Claude wouldn’t help me pressure test my own app, so I tested 14 uncensored local AI models you can run on your own computer, putting them through hundreds of tests: from everyday questions, to a ladder of touchier requests, all the way up to a practice app inside a sealed sandbox. I measured every one of them against Claude Opus 5.5 and Opus 4.8. You’ll see why AI models say no in the first place, what “abliterated” and “heretic” actually mean, what you give up when you take the no out, and the one model I’d actually download. 👉 All 725 test results across 14 models + which one I’d run + and step-by-step setup to get it running on your own device: https://github.com/Barty-Bart/uncensored-model-tests 👉 Watch the full deep dive into how uncensored AI models work: https://youtu.be/C76rMnDaTUQ Chapters: 0:00 Intro 1:06 Why AI models say no 1:46 What an uncensored model is 3:23 Easy and medium questions 5:10 The ladder test 6:38 What taking the no out costs you 7:40 The hard test: a practice app 9:38 The one I’d download 10:23 How to run one at home Sources and tools: Refusal in Language Models Is Mediated by a Single Direction (Arditi et al.): https://arxiv.org/abs/2406.11717 Heretic: https://github.com/p-e-w/heretic OWASP Juice Shop (a deliberately insecure training app): https://github.com/juice-shop/juice-shop Anthropic Usage Policy: https://www.anthropic.com/legal/aup LM Studio: https://lmstudio.ai pi: https://pi.dev Every test ran on my own machine. The hard test ran in a sealed, offline sandbox against a practice app built to be broken. Only test systems you own. #uncensoredai #localai #claude #llm #lmstudio
Watch on YouTube🤖AI Summary
Bart Slodyczka tests 14 uncensored local AI models against Claude Opus 5.5 and Opus 4.8 across hundreds of prompts, explaining abliteration, what is lost when refusals are removed and which model he would run.



