A growing backlash over Google's military artificial intelligence contracts has prompted several employees at its DeepMind research division to resign, while others are preparing to leave, amid concerns that the company's technology could be deployed in warfare without adequate safeguards or meaningful human oversight.
The departures highlight growing tensions inside one of the world's leading AI developers as Google expands its cooperation with the U.S. military, including through an agreement allowing its technology to be used in classified defense operations.
According to a BBC investigation published Wednesday, current and former DeepMind employees have expressed concerns about the company's military partnerships, the reliability of AI systems in combat and the potential professional consequences of speaking out against Google's decisions.
Among those who have left is Sofia Liguori, a London-based research engineer who resigned this week after three years at Google. Speaking to the BBC, she said she was ashamed of the company's increasingly close involvement in military AI development.
Liguori said Google's direction had undermined her confidence in its commitment to its publicly stated ethical principles. She also described a workplace culture in which employees were reluctant to raise objections because they feared retaliation or damage to their careers.
Another departing employee, AI researcher Andreas Kirsch, resigned earlier this month. He told the BBC that Google's internal governance and oversight were inadequate to ensure the responsible deployment of artificial intelligence in classified military environments.
Kirsch also questioned whether the company had sufficient control over what happened to its technology after it was delivered to the Pentagon, suggesting that Google had limited visibility into military deployments and little ability to prevent uses it considered inappropriate.
Both researchers raised concerns about the technical reliability of AI systems in military operations, where inaccurate information or automated recommendations could have life-or-death consequences. Kirsch argued that today's AI models were neither sufficiently robust nor adequately trained for such demanding applications.
The controversy comes as the Trump administration accelerates its adoption of artificial intelligence throughout the armed forces. Defense Secretary Pete Hegseth has made the technology a central priority, promising to transform the U.S. military into an "AI-first warfighting force".
Google is among a growing number of technology companies that have signed significant agreements with the Pentagon this year. Others include OpenAI, Amazon, Microsoft, SpaceX, Oracle and Nvidia, with their technologies being deployed across military applications ranging from logistics and operational planning to battlefield activities.
The agreement Google signed with the Pentagon earlier this year permits its AI models to be used for "any lawful government purpose", including classified military work. It also includes restrictions concerning domestic mass surveillance and autonomous weapons, including target selection, without appropriate human oversight and control.
Google maintains that its contracts comply with the company's AI principles, including commitments to human supervision and respect for international law and human rights. Critics, however, question whether such safeguards can be effectively enforced in classified military operations, particularly when the technology is controlled by government agencies rather than its developers.
One example of the military's expanding reliance on AI is Maven Smart System, a Pentagon platform developed by U.S. defense technology company Palantir. The system processes and organizes intelligence, using artificial intelligence to help military personnel identify potential targets and accelerate operational decisions.
Palantir has integrated technologies from several AI companies into Maven, enabling faster analysis of intelligence and targeting information. However, the expanding capabilities of such platforms have also prompted concerns about whether human operators have sufficient time to verify automated recommendations.
US Defense Secretary Pete Hegseth speaks with troops during a visit to U.S. Central Command's area of operations
(Photo: U.S. Department of Defense)
Retired U.S. Air Force Lt. Gen. Jack Shanahan, who previously led Project Maven, told the BBC that the Pentagon appeared increasingly focused on deploying AI as rapidly and widely as possible, potentially at the expense of rigorous testing and evaluation.
Emil Michael, an undersecretary at the U.S. Defense Department, has defended the emphasis on rapid implementation, arguing that military success in the AI era increasingly depends on speed.
Google's military-related agreements with Israel have generated similar controversy. The company and Amazon were selected in 2021 for Project Nimbus, a cloud computing initiative worth approximately $1.2 billion intended to provide advanced digital infrastructure to the Israeli government and public sector, including defense-related agencies.
The project drew particularly strong criticism from Google employees in 2024, with opponents questioning the potential military applications of the technology and the company's ability to oversee how it was being used.
Google has repeatedly stated that Project Nimbus is not intended for highly sensitive or classified military applications involving weapons or intelligence services. Nevertheless, the agreement has remained a source of internal opposition, particularly as military AI applications have become increasingly sophisticated.
The dispute also reflects a broader concern about the distinction between supplying computing infrastructure to government agencies and directly supporting military decision-making systems. Employees argue that the practical limits of corporate oversight become especially important when advanced AI tools are deployed in classified environments.
Beyond the ethical and technical questions, the BBC investigation exposed concerns about Google's internal culture and restrictions on employees who want to criticize the company publicly.
Liguori said her employment agreement restricted her ability to criticize Google even after leaving. She nevertheless decided to speak publicly because she believed employees and the wider public deserved to understand the company's military involvement.
Describing the atmosphere surrounding internal dissent, she told the BBC: "You feel you are being watched, you can't speak."
Other DeepMind employees interviewed by the BBC anonymously expressed similar concerns. One said that many colleagues who objected to the company's changing ethical direction had already resigned, while others believed there was little prospect of influencing decisions made by such a large organization.
Another suggested that researchers attracted to DeepMind by its original scientific mission could gradually be replaced by employees more comfortable with the company's expanding defense commitments, potentially changing the organization's culture.
Google rejected suggestions that it lacked appropriate safeguards for its government contracts or that employees could not safely raise concerns.
A company spokesperson told the BBC that Google uses consistent procedures to evaluate public-sector agreements worldwide, ensures compliance with its acceptable-use policies and AI principles, and provides dedicated reporting channels for employees to raise ethical concerns or report misconduct.
The spokesperson added that Google maintains workplace policies designed to prevent retaliation against employees who speak up.
The disagreement nevertheless underscores an increasingly difficult question for the AI industry: As its most advanced technologies become embedded in military operations, how much control can the companies developing them retain over their use, and what happens when the researchers who built them no longer trust the safeguards?