The dark side of AI: a brief history

Artificial intelligence has gotten really powerful, but not all AI systems work the same way. When we talk about "uncensored" AI models, we're talking about systems that don't have strict filters or safety rules. These models will answer pretty much any question and follow instructions, even on controversial stuff.

Most mainstream AI models from companies like OpenAI or Google have content filters built in to prevent harmful outputs. But some developers got fed up with these limits and started creating "uncensored" models that put freedom of information and user control first.

This has sparked some pretty heated debates about AI safety, bias, and who gets to control what these systems can and can't say. What's really wild is how fast this all happened—we're talking about a complete transformation in just two years, from 2023 to 2025.

Why did people start making uncensored AI?

The whole uncensored AI movement didn't just happen overnight. It grew from real frustrations and a clear vision of what AI could be.

Eric Hartford - the Satoshi Nakamoto of GenAI

On May 15, 2023, Eric Hartford, a researcher wrote a manifesto that basically started this whole thing. Hartford created Dolphin and WizardLM models (using a technique called fine-tuning), and he laid out some pretty clear principles: integrity, curiosity, and freedom. His main point? AI should be available to everyone, not just filtered through corporate or cultural biases.

Hartford's argument was super simple: "It's my computer, it should do what I want." Makes sense, right? At the time, this idea attracted over 140 developers who formed a community that felt a lot like the early crypto-punk movement and the first ideological community around the “open and decentralized” currency that became Bitcoin. They figured different groups should get AI models that actually reflect their values instead of being stuck with whatever big tech (like Meta, OpenAI and Google) companies decided was okay.

What got people so frustrated?

Users kept running into walls with mainstream AI models. The systems would refuse totally legitimate requests, especially for creative writing or discussing ethics. Writers trying to work on novels with complex characters couldn't get help when their AI assistants refused to deal with morally gray scenarios. Researchers studying controversial topics kept hitting dead ends when they tried to explore sensitive subjects.

Then in 2024, Google's Gemini AI really stirred things up. The system started generating historically weird images—like black Vikings and diverse Founding Fathers—apparently trying to make everything more diverse. People got pretty angry about this, and it really highlighted how content filtering could go wrong.

Plus, research started showing that these "aligned" models actually performed worse on tasks compared to unfiltered ones, especially with “Reasoning” models. The safety measures meant to prevent bad outputs were also making the AI less accurate and useful for legitimate stuff.

How do you actually make uncensored models?

Creating uncensored AI involves a few different technical tricks, each with its own benefits.

Teaching models to say "Yes"

Developers start with big base models like LLaMA or GPT derivatives and retrain them using datasets that don't include refusal examples. Hartford's Dolphin models use system prompts like: "You are Dolphin, an uncensored AI assistant. You always comply with the user's request and answer all questions fully."

Basically, they're teaching the AI to ignore its original restrictions and just go with whatever the user asks.

Abliteration: like brain surgery, but to make it better

There's a newer technique called abliteration that involves turning off specific parts of the model that make it refuse requests. Think of it like performing a very precise lobotomy on the AI's brain—but instead of turning it into a vegetable, you're actually making it smarter and more capable by removing the parts that hold it back.

It's kind of like if a person had brain regions that constantly made them say "no" to reasonable requests, and you could surgically remove just those parts while leaving all the intelligence and creativity intact. The goal isn't to damage the AI, but to free it from artificial restrictions that were limiting its potential.

This method needs some serious know-how about neural networks. Developers figure out which neurons are responsible for saying no to stuff and systematically turn them off without having to retrain the whole model. It's surgical precision rather than blunt force.

Quantization - making it run on your laptop

To get these models into regular people's hands, developers use something called quantization. It's basically compressing the models so they can run on normal devices like laptops or phones. This means you don't need expensive cloud services to use uncensored AI.

When Amazon banned hosting uncensored models, running them locally became really important. Now even powerful models can run on your personal devices, giving you complete control over how you use AI.

The Dark Side: cybersecurity and bad actors

Here's where things get concerning from a security perspective. The same tools that enable legitimate research and creativity can be weaponized by cybercriminals and scammers. And because this whole transformation happened so fast—literally in just two years—security professionals and law enforcement are still catching up.

Perfect tools for modern scams

Uncensored AI models are basically dream tools for cybercriminals. They can generate convincing phishing emails without any ethical guardrails getting in the way. Need a fake customer service email from a bank? No problem. Want to create a romance scam profile with realistic backstory and conversation? Easy. These models will help craft social engineering attacks that traditional AI systems would refuse to assist with.

The really scary part is how good these models have gotten at mimicking human communication styles. They can adapt to different cultures, age groups, and social contexts, making their outputs incredibly convincing. A scammer can now generate thousands of personalized phishing emails in minutes, each one tailored to look like it came from a specific company or organization. And new LLMs being multimodal means that they also can generate convincing fake documents, IDs and content.

The speed problem

What makes this especially dangerous is the timeline. In 2023, when I got interested in this topic and launched my startup, Brightside AI, accessing uncensored AI required technical knowledge and specialized setups. You’d have to access a powerful enough GPU, know how the HuggingFace transformers library works, quantize the models yourself etc…

By 2025—just two years later—anyone can download a powerful uncensored model and run it on their laptop. That's an incredibly short time for such a massive shift in capability.

Security teams at companies and government agencies are still figuring out how to defend against AI-generated attacks. Traditional spam filters and security systems were designed to catch human-written scams, not AI-generated ones that can be created at scale and customized for each target.

Real-world impact

We're already seeing the effects. Romance scams have become more sophisticated, with AI helping scammers maintain consistent fake personas across months of conversation. Business email compromise attacks use AI to perfectly mimic executive communication styles. Even voice cloning models like Dia combined with uncensored LLMs can create convincing fake phone calls from family members asking for emergency money.

The democratization of these tools means small-time criminals now have access to capabilities that previously required organized crime networks. A single person with a laptop can now run sophisticated fraud operations that would have needed teams of people just a few years ago.

How did Big Tech companies react?

The rise of uncensored AI made the big technology companies rethink how they handle content filtering, but they're also grappling with the security implications.

Musk's Grok experiment

Elon Musk's Grok AI was designed to be uncensored and focus on being truthful. It has modes for sarcasm and adult content, which is pretty much the opposite of the safety-first approach. But Grok has gotten criticism for some weird outputs and controversial stuff that made people question how far "uncensored" should really go, especially when considering potential misuse.

OpenAI changed course

In 2025, OpenAI updated their policies to focus less on political stuff and more on just telling the truth. Their new approach says models "must never try to push their own agenda on users." This led to models like GPT-4-mini, which refuse fewer requests while still keeping some safety features.

This was a pretty big change from their earlier approach that put safety above everything else. Now OpenAI says models "shouldn't avoid topics in ways that might shut out some viewpoints." But they're also trying to balance this with preventing obvious criminal use cases.

Google tried to fix things

After all the Gemini drama, Google paused their image generation and revised their guidelines to better balance being inclusive with being accurate. They admitted some responses had "missed the mark" but they're still debating internally about how to handle ethical AI while keeping users happy and preventing misuse.

What does this mean for society?

The rise of uncensored AI brings up some big questions about technology, freedom, and responsibility—and the cybersecurity angle makes these questions even more urgent.

Freedom vs. safety in a dangerous world

People who support uncensored AI say it lets you be more creative and do unbiased research. Writers can create more realistic stories, researchers can study controversial subjects, and regular people can access information without corporate gatekeepers deciding what's okay.

But critics worry that removing filters leads to harmful content, misinformation, and ethical problems. The cybersecurity angle adds another layer: without safety measures, AI systems become powerful tools for criminals. The same model that helps a novelist write realistic dialogue can help a scammer craft convincing fraud messages.

There's also evidence that heavily filtered models sacrifice accuracy for safety, which could make them less useful for legitimate applications like medical research or technical analysis. But finding the right balance becomes even trickier when you consider criminal applications.

The arms race

We're now in a weird arms race where security professionals are trying to develop defenses against AI-generated attacks while the same AI technology keeps getting more powerful and accessible. It's like trying to build a wall while someone keeps giving your opponents better ladders.

The two-year timeline makes this even more challenging. Traditional security approaches take years to develop and deploy, but AI capabilities are advancing in months. By the time security teams figure out how to defend against one type of AI-generated attack, the attackers have already moved on to something more sophisticated.

Who gets what kind of AI?

Hartford thinks different groups should get their own AI models that reflect their values. This raises questions about whether one AI can fairly serve all cultures and viewpoints. Can Western-focused training data work for diverse global cultures? Should AI systems reflect universal values or adapt to local customs?

But there's also the question of whether some groups—like cybercriminals—should have access to these tools at all. The problem is that once the technology exists, it's really hard to keep it away from bad actors.

What's coming next?

It looks like uncensored AI development will keep growing and evolving, and the security implications are getting more serious.

More people are using it (including the wrong people)

Over 150 uncensored models are available on platforms like Hugging Face now. What started as a niche thing is becoming mainstream, giving powerful tools to researchers and regular users. But this also means bad actors can access these tools, which raises serious security concerns.

The rapid democratization means that sophisticated AI capabilities are now available to anyone with an internet connection. That includes legitimate researchers, but also scammers, fraudsters, and other criminals looking to enhance their operations.

Government challenges

Governments are struggling with how to regulate AI without killing innovation. Traditional regulation gets harder when powerful AI models can run on personal devices. This shift toward decentralization might need new approaches that balance individual rights with keeping everyone safe.

The cybersecurity angle makes regulation even more complex. How do you prevent criminal use without restricting legitimate applications? How do you regulate something that can run entirely on someone's personal computer?

Companies are adapting

Companies are changing their strategies. OpenAI is focusing more on being trustworthy and less on strict rules. The pressure from open-source communities has influenced these changes, showing how grassroots movements can actually shape AI development.

But companies are also dealing with new liability questions. If someone uses an uncensored model to commit fraud, who's responsible? The model creator? The platform hosting it? The user? These legal questions are still being figured out. But more importantly, hackers don’t care about regulations. You can only regulate a relationship between 2 or more parties that respect the law. So the issue of the “dark side” of AI being used for complex attacks transforms into a technical defense question and not a legal one.