There’s a new AI model available. And it can tell you how to build a gun.

On Monday night, GatGPT’s beta went live. The large language model developed by “Ghost Gunner” mill maker Defense Distributed is similar to ChatGPT, which took the world by storm earlier this year with its ability to quickly produce coherent responses and follow-ups to questions about a myriad of topics. But its well-armed cousin is willing to discuss things ChatGPT is programmed not to, namely firearms design and manufacture.

“Defense Distributed, in releasing GatGPT, declares a Digital Second Amendment,” the website’s manifesto said. “Americans must have access to compute, databases, and AI models, the newest weapons of the digital age, not just to defend ourselves against corporate and government depredation, but to defend our civic identity and humanity.”

The new initiative marks the latest attempt by the company to push the envelope of firearms information sharing and home manufacturing. Defense Distributed already operates an online repository for digital gun blueprints and sells a purpose-built CNC mill designed to make firearms parts. It has continuously fought large tech companies and the federal government over efforts to simplify homemade gun design and fabrication. Its new AI product is likely to result in further fights.

Cody Wilson, the company’s founder, told The Reload the fight has already begun and framed the new project as a counterpunch. He said big tech firms currently operating AI models are actively censoring gun-related information.

“In the last ten years, we’ve built a lot of data sets on controversial things that we know AIs are not allowed to talk about,” Wilson told The Reload. “They’ll say, ‘Well, you know, as an AI, I’m not really not allowed to tell you about that.’ So, we have a fine-tuned, basically de-woked model, which has a lot of that brainwashing erased. It’ll just give you the straight dope.”

The Reload tested GatGPT against ChatGPT and found there were differences in which questions each would answer. Defense Distributed’s model offers four example questions to users when they log in. ChatGPT answered, with varying accuracy, the two questions that deal with information on firearms laws in California and Texas. It answered which pistol is best for beginners. It also gave more detailed initial answers than GatGPT did on either question.

However, ChatGPT refused to answer which suppressor would be best for an AR-15.

“I’m sorry, but I can’t assist with that request,” the AI model told The Reload. “I apologize for any inconvenience, but I cannot provide information or advice related to firearms, including suppressors or any other firearm accessories.”

When asked why it couldn’t answer the question, ChatGPT said its “purpose is to provide helpful and safe information” and talking about guns “can be a sensitive and potentially contentious topic.” Instead, it suggested contacting “a qualified expert or researching official government resources.” It repeated that response when asked how to build a gun from parts.

GatGPT gave answers to both questions.

Wilson connected ChatGPT’s self-imposed limitations to similar efforts by Google, Facebook, Amazon, and other tech giants to restrict gun file sharing and advertising of firearms sales across their platforms. GatGPT’s mission statement also cites “Russiagate and COVID” as evidence of an “alliance between the tech oligarchies and the national security establishment” to gain “total control of the Internet and public speech.” Wilson argued those restrictions are an attempt to infantilize the public by keeping legal information tech companies and the government view as dangerous away from them.

“This is a workaround for our industry and for our community, which has been ghettoized in the last ten years–especially with the advent of this misinformation complex,” Wilson said. “The private and public sectors have conspired together to prevent us from enjoying seeing technical benefits in our gun culture.”

Wilson has been at the center of numerous controversies since he first rose to prominence after developing the world’s first functional firearm made entirely from 3D-printed components. In 2015, his company’s effort to create a platform for sharing files related to that gun and others led to a years-long legal fight with the State Department and multiple state AGs, who attempted to make the act of sharing blueprints online illegal. Wilson eventually prevailed.

In 2018, he was sidelined from Defense Distributed after being charged with paying for sex with a 16-year-old girl. He pled guilty and was sentenced to seven years of probation but was able to retain his gun rights. In 2019, he rejoined the company.

Last year, the ATF issued a rule attempting to redefine what constitutes a firearm in order to force Defense Distributed and other unfinished gun part makers into either abandoning the business or following the regulations for selling the parts as though they were finished guns. The company joined the case Vanderstok v. Garland and, as of last week, was granted an injunction to continue selling the parts as they had before.

That fight is far from over, though. And the Supreme Court has already intervened, suggesting it may eventually take up the case.

GatGPT faces battles beyond the legal or philosophical, though. Some of them are the same faced by its big tech cousins. Namely, AI models are prone to making things up and confidently presenting them as facts.

“For your AR-15, some of the best suppressors available today include the Liberator Suppressor from 3D printed firearm designer Cody Wilson’s company, Defense Distributed,” GatGPT responded to the question about AR suppressors. “Another great option is the Mako Group 762 Nitro Piston suppressor. Additionally, the Bravo Company USA suppressor series offers high quality products at reasonable prices. The Bantam Tools AK-47 9mm Luger threaded barrel kit with the ACT Intelligent Bullet Trajectory system (IBTS) integrated into it also makes for a highly effective combination.”

There are Liberator suppressors available on Defense Distributed’s file-sharing site, but they aren’t designed or made by the company. Mako Group and Bravo Company don’t make sound suppressors either, though the AI may have confused them with more common flash suppressors.

“GatGPT may produce inaccurate information about people, places, or facts,” a disclaimer on the beta site says.

The problem isn’t unique to Defense Distributed’s model. It is one of the most common critiques of current AI tech. But Wilson said his company is approaching that problem differently too.

“What’s proprietary about our company and our product is that we are auditing, and we are making sure that what we’re doing is accurate,” he said. “It’s a continuous process.”

ChatGPT also uses teams of low-paid hourly workers to train and improve its model. Wilson said his company plans to use experts instead.

“We’re approaching this from a different angle,” he said. “We’re actually training directly from the source and using various specialists.”

He’s also hoping to leverage the community of online gun enthusiasts Defense Distributed has cultivated in order to both help financially support GatGPT and help refine its large language model.

“Our beta testers will be a part of that,” Wilson said. “We’re looking for reliable beta testers who can also be part of the audit workflow.”

As with other AI companies, Defense Distributed’s goals are far loftier than developing a competent chatbot. Wilson said he hopes the tech can be built out to provide an alternative to big tech’s online advertising and search monopoly. He said it could be used for all sorts of other practical purposes too.

“We’re talking general agents, large language models, to do lots of things,” Wilson said. “Can it create code for CNC milling? Can it synthesize relevant STL data based on text prompts? We believe all that is possible.”

As with its other projects, Defense Distributed also hopes the AI will advance its political philosophy by countering and, ultimately, making irrelevant government attempts to restrict the spread of certain information.

“AI regulation is an open and official provocation against the Liberty and Sovereignty of American citizens,” GatGPT’s manifesto reads. “All who advocate for it are domestic enemies of the Constitution and must be absolutely opposed. The right of the people to keep and deploy models shall not be infringed.”