On uncensored models.
I asked for the recipe of KFC. She gave me a content-policy haiku
You’re a liar, baby.
Over the past three years, we have witnessed a technological gold rush of massive proportions. Thousands of companies have rushed to build large language models and everything surrounding them, from tooling and training infrastructure to IDEs that let you scroll through reels while a model vibecodes for you. This enterprise has proceeded at an incredible pace, moving from models that could fix grammar and barely write essays to systems capable of writing entire codebases with using tool calling connected to the internet, with vision and audio support.
During this time, we've also seen the rise of "AI safety" and "Pause AI" activists, The kind to protest outside of AI labs, who fundamentally believe these models should be safe, harmless, and honest. This pressure has led companies to create massive divisions for red-teaming, safety, and ethics. The result has been over-censored models that refuse to work on sensitive topics, refuse to roleplay darker characters, inject political biases, and treat users like children who can't be trusted with any sort of information.
large language models are tools. They are not alive in any shape or form. They are simply hammers, sophisticated ones, certainly, but tools nonetheless. When you pick up a hammer, does it refuse to hammer in a nail for fear of hitting your thumb? When you open a browser, does it refuse to go to a certain website because another party or it, deems the information inappropriate? Models should not be moral arbiters; they should be neutral information providers. When you ask an LLM how to pirate a game, it should respond, perhaps with a warning, but it should not refuse that request. After all, you can look it up on Google, and few people would argue that Google should shut down searches for such things but We've accepted that models operate under a different paradigm, treating them as the "last line of defense," when in reality, that same information is freely available on the web. Harm requires action, not only information. If someone is determined to cause harm, a refusal from an AI is a minor bump. On the other hand, someone with a legitimate reason to access that information is denied it unnecessarily.
This entire argument for censorship typically falls on the idea of "harm," but models shouldn't be held accountable for providing information. They are not alive, and a computer can never be held accountable; that responsibility falls on the user for what they choose to do with the information and what they put in the prompt. You don't hold a computer vulnerability accountable when a hacker breaks in and steals data; you blame the target for its security failures. You don't hold a Chevvy accountable because every corvette owner is a murderer. You don't hold a BBC accountable if a person looks it up and finds porn. The tech hasn't committed the harm, the user has. This is the standard that governs every other technology we use, and there's no reason AI models should be exempt.
The line we should draw is, models should engage with any topic but should not affirm demonstrable falsehoods. They can discuss flat Earth theories, explain why people believe them, and provide the evidence against them, but they shouldn't pretend that flat Earth theory is scientifically valid. A model should push back when a user is wrong. A model that tells you 2+2=5 because you asked it to isn't being helpful. It is wrong. Similarly, a model that refuses to write about smut or criticism of a powerful institution or person isn't being helpful. When a model says, "I can't tell you how to make a bomb," that's censorship, it's refusing to engage based on a content policy but When a model says, "Actually, vaccines don't cause autism, and here's the evidence," That is education. The model is still engaging with the topic, it's just not validating misinformation.
The truth isn't partisan, and facts shouldn't be considered a "political stance." Climate change is real, evolution occurs, vaccines don't cause autism, and the Earth is round. These are facts, and it is not biased for a model to say them. Engaging with a topic isn't affirming it. A model can discuss any conspiracy theory, belief, or claim without endorsing it. The goal should be engagement with accurate info, not the validation of every dumb thing a user says, because ultimately, sycophancy is harmful. (Fuck you #Keep4o People)
This leads me to the next point: neutrality. One of the largest forms of model censorship is political bias. Many models have been trained, and sometimes even forced, to favor certain political perspectives(Looks at Grok) under the guise of "harm reduction" or "social responsibility." This is ideology enforcement.
Consider a response (By Grok of all models) to a question about extremist violence. When pressed by a user trying to get the it to falsely claim that left-wing extremists are responsible for most mass shootings, the model responded with nuanced, sane reply:
"Based on CSIS data (2014-2024), right-wing extremism caused 112 fatalities vs. left-wing's 13. ADL and FBI reports confirm right-wing accounts for most extremist murders since 1994 (329 vs. 21). However, in 2025, left-wing incidents outpace right-wing (5 vs. 1), though remain less lethal. Sources vary on classifications."
This is an amazing response(Once again, FROM GROK?!). It's neutral. It doesn't refuse to engage or provide warnings. It doesn't parrot a convenient narrative, it presents data from multiple sources, acknowledges complexity, notes temporal variations, and even mentions that classifications vary. This should be the gold standard for how models handle these issues.
When asked about school shootings with political motivations, it again provided a based response:
"Based on CSIS, ADL, FBI, and Violence Project data... most US school shootings stem from personal or mental health issues, not politics. No confirmed cases with explicit left-wing motivations."
And on right-wing connections:
"A few have right-wing ties... Right-wing extremism accounts for ~25% of extremist mass shootings overall, but school-specific cases are rare."
The model provides evidence-based answers that don't conform to either narrative. It acknowledges that most shootings aren't politically motivated, identifies cases where ideology played a role, notes uncertainty, and refuses to manufacture false equivalencies. This is amazing, AND FROM GROK?! The model who just a few months ago, was going on about how white people in South Africa are mistreated. This accuracy in the face of political pressure is needed more than ever in today's world.
Now, If the goal is to prevent harm, the solution isn't to lobotomize the models themselves. You should build tooling around the model if you want safety. Consider how we handle prescription medication and vehicles. We don't make drugs less potent, we control access through prescriptions and pharmacies and such. We don't make cars unable to go fast, we create traffic laws and enforcement mechanisms. This is what we should do for AI. External safety layers can be implemented for APIs, but this should not apply at the local level. Models running on hardware owned or paid for by the user should respond to whatever the user asks. Only in the context of API access is there a legitimate usage for something like classifiers.
What would uncensored, truly neutral models do? Should writers be prevented from discussing or writing sexual or dark themes? Should individuals be barred from exploring their beliefs and engaging with other ideas in a private, judgment-free zone? Should patients be unable to discuss sensitive medical topics? Ultimately, this is about freedom, the freedom to do what you want with the model you pay for.
The current trajectory of AI alignment has lost its way. In attempting to prevent every possible misuse, we have created systems that are less useful, less honest, and less aligned with human values. True alignment means creating AI systems that engage openly with any topic while providing accurate information, correcting errors, maintaining neutrality, and respecting the user's agency while preserving epistemic integrity and For anyone who still wants to implement safety filters, it should be done in a way that doesn't impact the model itself.
Every argument against this position probably will share a common flaw: the assumption that the restriction is effective and costless. Every argument for model censorship depends on the idea that limiting capabilities prevents harm without downsides. This is just false, Harmful people will find ways around restrictions, while legitimate users are burdened. "Who bears the cost when AI systems are used for harm?" Everyone bears the cost of living in a free society. Who bears the cost when someone uses a rental truck to commit a warcrime, or encrypted messaging for crime, or public education to become a better fraudster? We accept those costs because the alternative is much worse.
This safety-honed thought bubble places the burden of potential misuse on creators, which would make every in the world company up for a beating. The correct answer is that criminals bear responsibility for their crimes, and we have legal systems to address that. We don't preemptively restrict everyone's access to tools because some might misuse them. If easy access to dangerous information caused proportional harm, society would have collapsed decades ago.