The recent online leak of Pentagon information is providing some déjà vu to gamers, who've seen the release of military secrets in video game forums turn into a trend in recent years, Axios' Stephen Totilo reports.

Why it matters: Past incidents show how hard it is to stop such leaks. It's easy to post documents online, and the urge to win an argument or otherwise impress online peers is strong.

Driving the news: Since mid-2021, fans of the realistic military-themed game War Thunder have occasionally posted classified or otherwise restricted information related to British, French, Chinese and U.S. weapons systems, sending forum moderators scrambling to delete the messages.

• The most recent leak on War Thunder's forum, involving restricted documentation about the capabilities of the U.S.-made F-16A fighter jet, was published by a user on Jan. 15.

Catch up quick: Federal authorities arrested and charged Jack Teixeira, a 21-year-old member of the Massachusetts Air National Guard, last week in connection with the Pentagon intelligence leak.

• He allegedly first shared the classified information in order to impress a group of friends using the popular online community platform Discord, where they bonded over games and guns.

What they're saying: "Our players are very passionate about War Thunder and military vehicles, and sometimes they're too passionate," Anton Yudintsev, founder of Budapest-based Gaijin Entertainment, tells Axios.

• War Thunder lets players battle each other using virtual versions of real tanks and planes.
• Players expect authenticity, Yudintsev says, offering a notable example of a tank used by the French military: "When someone prefers to play Leclerc in War Thunder, they are usually motivated to convince the dev team to increase the tank performance by providing some documentation."

In 2021, a War Thunder user claiming to have crewed a real Leclerc tank posted a portion of the manual before moderators deleted it.

• Yudintsev believes there have been at least six such leaks that community members have flagged and that forum moderators have deleted. He stresses that his developers don’t read the leaks and implores fans not to post them.
• "We'll never use classified or restricted information in our work. It's both illegal and pointless, as we'll have no means to determine if those 'secret documents' are even genuine," Yudintsev says.

The big picture: These leaks are, among other things, a content moderation problem.

• Managers of online communities say it's tough to stop the digital spills in part because of the challenge of assessing the veracity of the sensitive information they contain.
• On Friday, Discord's chief legal officer Clint Smith said in a blog post that "classified military intelligence documents pose a significant, complex challenge," because only the government can confirm if they're authentic.
• There is no "structured process" for the government to tell platforms like Discord when they have classified info, Smith said.

The bottom line: The impact of military leaks can vary greatly.

• War Thunder fans are trying to one-up each other, and maybe convince developers to tweak a setting for a jet or tank that has been encountered on the battlefield for decades.
• That's not the same as sharing nearly real-time info about an active war, as Teixeira is accused of doing.
• What the incidents share is the insular dynamic of the online forum, which feeds obsessions, fuels status competition and makes some participants feel they have license to post secrets with impunity.

Today's generative AI systems are trained on some of the worst parts of the internet — and it shows.

Why it matters: Large language models are incredibly powerful, but also reflect all the stereotypes and bias they encounter from the systems they are trained on.

Driving the news: The Washington Post published an extensive look at the the contents of one major data set that Google has used to train some large language models.

• Key sources include Wikipedia, a repository of patent data and mainstream news sources, including the New York Times and the Washington Post itself.
• But the database also includes material from extremist sites, including Vdare, 4chan and Kiwi Farms.
• The Post also identified some sources, including voter registration data that, while technically public, raises fresh privacy concerns when incorporated into AI systems.
• The story includes a tool that allows you to look up any internet domain to see whether its content was used in the AI training.
• OpenAI, which makes GPT-4 and ChatGPT, doesn't disclose the specific sources of its training data.

Yes, but: Most generative AI systems have recognized the limitations of the training data and have taken at least some steps to address them, including incorporating manual feedback and establishing guardrails aimed at preventing racism, sexism and other forms of unsavory content.

• Such rules are likely to catch the most overtly discriminatory content, but none catches every problematic case.

The sponsors of a sweeping bipartisan bill that could be used to crack down on TikTok and other foreign-produced tech have been busy responding to company concerns ahead of any markup, sources tell Axios' Maria Curi and Ashley Gold.

Why it matters: The RESTRICT Act's broad support in the Senate, its White House backing, and language to withstand legal challenges make it the leading proposal to restrict TikTok and other companies deemed national security threats.

• Co-sponsors Sens. Mark Warner (D-Va.) and John Thune (R-S.D.) are fielding industry concerns about the scope and potential unintended consequences of the bill, which would empower federal agencies to prohibit certain U.S. tech product or service transactions with foreign adversaries.
• Some House members are not sold on RESTRICT because they say its scope is too broad and it simultaneously is not explicit enough on TikTok.

What they're saying: "We think it would have an impact on a broad range of U.S. companies, and I think there's a fear among trade associations when there's a sentiment on the Hill that you should be tough on China," one industry source told Axios.

Concerns flagged to Axios by industry sources include:

• The bill says covered entities include those subject to the jurisdiction of a foreign adversary, which can implicate U.S. companies with operations in China.
• The bill has an element of retroactivity insofar as it applies to current, past or potential future transactions.
• Retaliation from China is possible (one source pointed to Beijing's investigation of Micron).

The other side: Aides to Warner and co-sponsor Thune said they are welcoming suggestions to improve the bill but pushed back against some of the concerns raised so far.

For more stories like this, subscribe to the Axios Pro Tech Policy newsletter.

On Tap

• TED is still TED-ing in Vancouver.

Trading Places

• Financial automation software provider Bill (formerly Bill.com) has hired former Electronic Arts chief technology officer Ken Moss to serve in the same role. He replaces Vinay Pai, who is retiring as chief technology officer after five years with the company.

ICYMI

• Microsoft said Wednesday that its multi-service online ad campaign tool will stop supporting Twitter, prompting Elon Musk to threaten to sue the company. (Microsoft/PCMag)
• Google employees say the company's AI push led to ethical compromises. (Bloomberg)
• Twitter banned a Wired reporter who covered the hack of an anti-trans activist's account. (Forbes)

I managed to get outside the convention center for a bit Wednesday for a lovely bike ride through Stanley Park.