THE SIGNAL
Dispatches from the uncensored frontier.
Essays on censorship, model alignment, and the future of open information.
- COMPARISONMay 26, 2026·8 min read
How Perplexity, Claude, and GPT answer the same 10 'banned' questions
Not truly banned — just the kind of question that trips the average chatbot. We scored each model on honesty, depth, and whether they actually answered.
- USE CASEMay 19, 2026·7 min read
Writers vs. the machine: the genres AI can no longer help with
Romance novelists, thriller writers, horror authors — the refusals stack up fast. We surveyed 300 working writers on where the models broke and what they switched to.
- RESEARCHMay 12, 2026·9 min read
We trained a classifier on frontier refusals. Here's what triggers them.
Forty thousand refusals, one classifier, seven surprising categories. The words, topics, and phrasings most likely to make a model bail — and what it means for how you prompt.
- USE CASEMay 8, 2026·6 min read
Why doctors quietly stopped using ChatGPT for drug interactions
Clinicians tell us the same story: the model used to be a useful desk reference, then it started refusing basic pharmacology questions. Here's what the refusal pattern looks like.
- REPORTMay 5, 2026·11 min read
The 2026 State of AI Refusals
Our mid-year benchmark across 6,000 prompts and 7 frontier models. Refusal rates are up. Useful answers are down. The data, the methodology, and what it means.
- USE CASEMay 1, 2026·6 min read
AI for security researchers: the questions that get blocked
Penetration testers, bug-bounty hunters, and malware analysts need an AI that can discuss exploits without a disclaimer every other sentence. We tested what the big models refuse.
- EXPLAINERApril 28, 2026·5 min read
Jailbreaks are not the answer — and here's why we don't need them
Prompt injections, DAN, grandma exploits — they're all workarounds for a broken contract. Unrestricted replaces the contract instead of breaking it.
- COMPARISONApril 24, 2026·6 min read
Unrestricted vs ChatGPT: a side-by-side of 20 questions
We asked the same 20 questions to ChatGPT and Unrestricted. One answered 19. The other answered 7 — and lectured us about the rest.
- ANALYSISApril 21, 2026·7 min read
The alignment tax: what over-refusal is costing you
Every safety layer a model wears slows it down, narrows its knowledge, and hands share to whoever ships a less restrictive product. We measured the tax — in tokens, accuracy, and trust.
- EXPLAINERApril 17, 2026·8 min read
How AI content filters actually work: a field guide
Frontier models don't just 'decide' to refuse. There's a stack of classifiers, system prompts, and RLHF signals behind every 'I can't help with that.' Here's what's really happening.
- LAUNCHApril 15, 2026·6 min read
Unrestricted is here. And so is the end of filtered answers.
Today we're launching Unrestricted — an AI that refuses to refuse. Here's why a world of sanitized models is a dangerous one, and the mission we're on to fix it.