The GoLaxy papers: Inside China’s AI persona army

On a jet-lagged April morning earlier this year, Brett Goldstein’s phone buzzed with a message from a trusted colleague asking him to check out a link he’d just found.

Goldstein, a former government technologist and now a special adviser at Vanderbilt University’s Institute for National Security, was in the middle of a conference. Later, back in his hotel room, he scrolled through to the message again. The note was insistent and included a link.

“I hate links,” he said. 

Still, he was curious, so he scanned the link for malware, ran a few checks, and clicked — and what appeared on his screen was a trove of PDFs all written in Mandarin. There were pages of dense Mandarin characters, something that looked like technical schematics, and then, strangely, photos of prominent Americans.

“At first I thought it was spam,” Goldstein remembered. But as he scrolled through the files, he realized it was something else: a leak. He had a speech in the morning, but he couldn’t resist cutting and pasting chunks of the document into a large language model to see what he really had. 

“I’d tell myself, ‘I need to go to sleep,’ and then, ‘But I need to see what this is.’ Over and over again,” he said. 

The cache appeared to be internal documents from a little-known Beijing-based company called GoLaxy, and they seemed to lay out a chilling new approach to information warfare: an army of AI personas, engineered to look like us, think like us, and win our trust. Not the blunt-force troll farms of Russia, but something intimate, surgical, and already operational.

Hour after hour, he fed chunks of Mandarin text into the model. Schematics, PowerPoints, internal briefs spilled out. The theme was unmistakable: persuasion. 

“It turns out it’s talking about a sentiment system … in China,” he said. So someone is trying to convince someone else to have a certain view.”

Digital doubles

The most startling sections described “personas.” Not bots recycling slogans, but carefully constructed digital identities. “I start seeing mentions of Hong Kong and Taiwan and I’m scrolling through it and I’m like, I think they’re talking about personas here,” Goldstein said.

These weren’t crude avatars. They were digital doubles — plausible enough to slip into a social feed, argue politics, offer consolation. They didn’t just broadcast; they listened and adapted. “What we’re seeing with artificial intelligence and generative AI is you can really create detailed personas that are, in theory, tailored to the individual,” he said.

Goldstein recognized the technique because he had been studying rough versions of personas for years. “A decade ago, I would create clusters of people — Midwest, white, male, whatever politics, this economic bracket,” he said. “You’d have 10, 20 variables and that was considered good enough.”

Back then, the archetypes were blunt: white guys in Ohio, single moms in Michigan, retirees in Florida. The Chinese approach, by contrast, was precise — personas handcrafted for a single target. And according to the 399 GoLaxy pages, they were already at work in Hong Kong, in Taiwan, and possibly in the United States.

“When I saw photos of Americans in the deck, I thought, oh boy. We’ve come across something that’s a threat,” he said. “We need to get this out as fast as we can to help educate people and start thinking about defenses.”Goldstein knew he couldn’t tackle this alone. “It required more than one Brett to make this happen,” he said.

So he turned to Brett Benson, a Vanderbilt professor of political science who had lived in Taiwan, spoke Mandarin, and studied the dynamics of armed conflict. When Goldstein first shared the files, Benson was cautious. 

“Initially I was very skeptical,” he said. “I told him I didn’t really know what to make of it and I needed to spend some more time with it.”

In the leaked documents, GoLaxy described itself as a private enterprise. But as Benson dug deeper there was more to it than that: the company had been founded by the Chinese Academy of Sciences, it employed nearly 1,000 people, and, according to the documents, maintained ties to both military and civilian intelligence. 

“They claim in the documents that they serve Chinese national security and national strategy interests and goals,” Benson said.

Before the machines

For years, Chinese information operations had been relatively clumsy. During the 2019 protests in Hong Kong, state-backed accounts on Twitter and Facebook pushed official slogans and clunky memes. They were easy to spot, often recycling stock photos and repeating identical lines. American officials dismissed them as amateurish compared to Russia’s Internet Research Agency, which had seeded more sophisticated narratives during the 2016 election.

But even then, China’s ambitions were clear. It poured resources into censorship and surveillance at home, while experimenting with influence abroad. What the GoLaxy documents suggest is that Beijing has leapt from imitation to innovation — replacing mass-produced slogans with AI-tailored whispers, aimed at the weak spots of individual minds.

The cache listed dossiers on 2,000 American public figures, thousands of right-wing influencers, and at least 117 Republican members of Congress. Screenshots offered proof of concept: names and faces instantly recognizable. “Yeah, that was a little bit disturbing,” Benson said.

And he and Goldstein thought: if GoLaxy could do this with politicians and influencers, what could they do with ordinary citizens?

The method was straightforward and relentless: scrape millions of data points from social platforms, build psychological profiles, and use generative AI — China’s version of ChatGPT, called DeepSeek — to animate them. 

“These are fake but highly realistic AI-generated bots,” Benson explained. “It’s a made-up online profile that looks like a real person, interacts like a real person. Much more subtle than the Russian operations we saw in the 2016 election. They’re very, very tailored.”

What unsettled them most was their fluidity. 

“The thing that really blew me away is the adaptability of the personas,” Benson said. “That’s the part that’s kind of spooky. They’re subtle, realistic, human-like, and their ability to engage and shape narratives.”

The rainy day

For years, U.S. officials puzzled over China’s massive data heists — from the breach of the Office of Personnel Management in 2015 to the hack of Starwood hotels three years later. The explanation was always the same: Beijing was hoovering up all this data for a rainy day.

Benson believes the day has come. 

“That was the big revelation to me … the data mining efforts paired with AI have the capacity to engage in a front of cognitive warfare that I didn’t really think about before,” he said. “I think this is the new frontier in national security.”

Then, when the researchers prepared an opinion piece for The New York Times this summer, the paper reached out to GoLaxy for comment. Almost immediately, sections of the company’s website began to vanish. 

“Pages were disappearing,” Goldstein said. “We were watching in almost real time and just thing after thing was being deleted. There was a whole section on government relationships, and that just disappeared.”

GoLaxy, for its part, denied everything when Times reporters reached out for comment. They did not respond to The Record’s request for more detail. What is clear is the deletions were aimed at changing the narrative. 

The warning

Goldstein now spends his time developing ways to distinguish AI personas from humans. “I just wanna understand what’s an AI versus a human,” he said. “Because the alternative is, if we can’t solve this and we don’t know human versus machine, we’re in a lot of trouble.”

Benson, for his part, is focused on getting the public to understand what’s at stake. 

“That it’s here,” he explained. When asked what “it” meant, he didn’t hesitate: “AI-generated propaganda, or AI-generated information manipulation is here. 

“When I use ChatGPT, I increasingly trust it. I think a lot of people increasingly trust it. But there are other forums where AI is being used to manipulate information, to sow doubt, to generate disinformation, and disorientation. I wish more people understood that — that’s not a science fiction future. That is here.”