Signed! The opposition of 600 employees was useless—Google has "authorized" the AI safety switch to the Pentagon.

Signed! The opposition of 600 employees was useless—Google has "authorized" the AI safety switch to the Pentagon.

Google has signed an agreement with the U.S. Department of Defense allowing the Pentagon to use Google AI for classified work, delivered amid strong internal opposition from employees. The security restriction clauses in the agreement are ambiguously worded and have been criticized by legal experts as not legally binding, raising deep concerns about the boundaries of AI in military applications.

According to The Information, sources reveal the agreement allows the Pentagon to use Google AI for "any lawful government purpose," wording similar to controversial phrases seen in negotiations between other AI firms and the Department of Defense. The agreement also requires Google to help adjust its AI safety settings and filters at the government’s request, with terms considered more lax than those in the deal OpenAI reached with the Pentagon in February.

Just before the agreement was signed, over 600 Google employees jointly wrote to CEO Sundar Pichai, urging him to reject the deal, arguing that declining classified work is the only way to ensure Google AI is not abused. The signing marks a major shift in Google's stance on military AI collaboration and prompts the market to reevaluate the commercial and reputational risks of tech giants’ deep integration with defense departments.

Employee Opposition Failed to Stop the Agreement

Resistance within Google was evident prior to the signing. More than 600 employees joined together on Monday to write to Pichai, asking the company to reject the agreement and advocating that only by completely avoiding classified work can Google fundamentally prevent its AI from misuse.

This scene brings to mind history from eight years ago. In 2018, Google withdrew from the Department of Defense’s "Maven Project" concerning drone target identification after thousands of employees signed protest letters. This time, Google opted to proceed against employee wishes, completing the agreement and indicating a fundamental strategic shift in its approach to defense business.

A Google Public Sector spokesperson stated that the new agreement is a supplemental revision to the non-classified use contract signed last November. "We are proud to be part of a broad alliance of leading AI labs and tech cloud companies supporting national security," the spokesperson said, "We remain committed to consensus between public and private sectors — AI should not be used for domestic mass surveillance or autonomous weapons lacking proper human oversight."

Looser Terms Than OpenAI’s Agreement

Google’s contract has significant differences in its terms compared to competitors. The agreement OpenAI reached with the Pentagon in February also allowed use for "any lawful purposes," but OpenAI stated in its blog that it retains full autonomy over its "safety systems" and specified prohibitions on using AI for mass surveillance, autonomous lethal weapon systems command, or high-risk automated "social credit" decisions.

Google’s contract requires the company to assist in adjusting AI safety filters at the government’s request. The Google Public Sector spokesperson explained that these filters were originally designed for consumer scenarios, and that Google typically customizes them for corporate clients.

Currently, Google joins Musk's xAI and OpenAI as tech companies signing confidential AI usage agreements with the Pentagon.

Safety Clauses Not Legally Binding

The core clauses limiting AI use in the agreement have drawn skepticism from legal circles. According to sources, the contract includes the statement: "Both parties agree that AI systems are not intended to be, and should not be, used for domestic mass surveillance or autonomous weapons (including target selection), unless there is appropriate human oversight and control."

However, the contract immediately adds: "This agreement does not grant either party control or veto rights over lawful government operational decisions."

Charlie Bullock, Senior Fellow and Lawyer at the independent think tank Legal & AI Research Institute, pointed out that "not intended to be, and should not be" means these clauses "have no legal binding power." He said such wording only expresses that both parties find such uses undesirable, not that they would constitute a breach of contract.

Amos Toh, Senior Advisor for AI and National Security at NYU’s Brennan Center for Justice, noted that "appropriate human oversight and control" does not necessarily mean a human must intervene between target recognition and weapon launch. He said the Pentagon has not ruled out deploying fully autonomous weapon systems. "Requirements for proper human judgment in use of force just means that some form of human participation and decision exists in the overall deployment method of the weapon system," Toh said.

Anthropic’s Precedent Looms Over the Industry

The signing of the Google agreement comes amid ongoing industry controversy over the boundaries of AI military applications. In February, negotiations between Anthropic and the Pentagon broke down publicly, sparked by Anthropic’s refusal to accept the "any lawful purpose" clause and insistence on keeping two clear red lines against mass surveillance and autonomous weapons.

After the breakdown, the Pentagon classified Anthropic as a supply chain risk, and Anthropic is now contesting this in court. At the time, over 900 Google employees and more than 100 OpenAI employees signed a public pledge asking their employers to align with Anthropic’s two red lines.

Now, with Google signing the deal, this internal industry moral consensus is accelerating its unraveling, and doubts about the legal power of safety clauses are deepening concerns about actual constraints on AI in high-risk military scenarios.

Risk Warning and DisclaimerThe market has risks; investment must be cautious. This article does not constitute personal investment advice and does not take into account individual users’ specific investment goals, financial status, or needs. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment is at your own risk.