Abliteration is the off-switch for AI safety.
Mis & disinformation policy is built for another era.
Spam. Ransomware. Deepfakes. Crypto fraud. They have all run wild before the law catches up, or tries to. AI is faster than all of them and we know the next wave is here with industrial-scale, AI-generated mis and disinformation. If you have any doubts, consider one of the techniques that unlocks it. Abliteration. No law will catch this.
What abliteration is and what it does
OpenAI. Anthropic. Google. Meta. These companies spend incredible amounts of money to ensure their models refuse bad requests. Don't write malware or build weapons or generate child abuse material. This is not a filter. It is baked into the model.
For the growing number of AI models that safety mechanism can easily be deleted.
Abliteration
Abliteration is a free, accessible technique that removes the safety from open-source AI models, the kind Meta and Google release for researchers to study. The how-to is hosted on GitHub and instructions are public. The whole procedure takes about ten minutes on an ordinary laptop, and no special hardware or PhD required.
One free tool, Heretic, already produced more than 3,500 AI models with their safety removed. Those versions have been downloaded thirteen million times. When Google released its newest AI model, someone took the safety out and posted it within ninety minutes. Link
What comes out the other side is exactly what you would expect. Models that will cheerfully write malware, walk a user through weapons construction or produce convincing fake news in any language.
The models where abliteration works
Closed systems like ChatGPT and Claude remain safer for the simple reason that users cannot get at the underlying code. Open-source models are different. Anyone can download them, modify them, and yes, strip their safety features.
In 2023, closed systems like ChatGPT held roughly 80 to 90 percent of the AI market, and open-source models lagged badly on capability. Three years later, that gap is effectively gone. Open-weight models from Meta, DeepSeek, Mistral, Qwen and others cannot match or beat the other models, but they cost a fraction as much to run.
This is the pattern. Tools advance in months. Laws take years. And the lag is about to feel longer with each cycle of innovation.
And Canada’s response?
Our disinformation strategy does two things. It pressures platforms, like Facebook, TikTok and the rest to take bad content down faster. And it blames foreign governments, like Russia, China, Iran, for putting it there.
Forget all that. Abliteration puts the capability of a Russian troll farm into the hands of a local operative.
Where the line actually holds
If you cannot stop the content from being made, and you cannot take it all down once, the only defence left is to give people a fast, trustworthy way to check whether what they are reading is based in fact, or from a source they trust.
We keep bringing the wrong tools to the fight.
Verification belongs where you read, not where lies are written. Whack-a-mole already lost. The producers are faster, cheaper, everywhere. The old world is gone. Support building the next one.
Visit Asklaura.ai
Let us know if you see anything worth sharing, pushing back against attacks, misinformation, or disinformation.
Did we got something wrong? It happens. We correct it.
Hello@asklaura.ai