Summary

  • Google’s Pentagon agreement authorizes classified military work, but the BBC’s report leaves the verification of deployment limits and meaningful human oversight unresolved.
  • Andreas Kirsch and Sarah Shoker identify time, information and authority to alter outcomes as conditions for meaningful military AI oversight.
  • Sofia Liguori’s resignation places employee speech and Google’s stated ethical values alongside the company’s formal review and anti-retaliation mechanisms.
  • Pete Hegseth, Emil Michael and Jack Shanahan foreground deployment tempo in military AI, while the report supplies no mission-level evidence showing how tempo affects Google’s classified deployments.

The departures of Sofia Liguori and Andreas Kirsch from Google DeepMind put a narrow but consequential question before Google’s military AI arrangements: whether a contract-review process, permissible-use rules and human-oversight conditions provide meaningful assurance when models operate in classified, time-pressured military settings. The BBC report says the Pentagon agreement permits classified military work for “any lawful government purpose” and restricts domestic mass surveillance and autonomous weapons, including target selection, without appropriate human oversight and control. Google describes a consistent review process, acceptable-use policies, AI principles, employee concern channels and anti-retaliation rules, but the report leaves the operation of those mechanisms in classified deployments largely undocumented. Kirsch and Shoker focus on the time, information and authority needed for human review; Hegseth, Michael and Shanahan foreground speed, reach and test-and-evaluation discipline. Liguori’s account adds employee speech and the company’s stated ethical values. The evidence records a disagreement over whether written authorization has become meaningful assurance, not a proven failure or success of safeguards.

The authorized scope and the unverified deployment

The Pentagon agreement defines what Google’s technology may be used for at the level of contractual permission. It allows classified military work and says Google’s models can be used for “any lawful government purpose.” The agreement also says the models “should not be used for domestic mass surveillance or autonomous weapons (including target selection) without appropriate human oversight and control.” It therefore combines broad authorization with specified conditions. One clause identifies purposes the customer may pursue; the other identifies sensitive uses that require human oversight and control.

Contract review can establish approval before work begins, but it cannot by itself show what happens after approval. Verification asks whether actual deployments remain within the restrictions, whether errors can be identified, whether reviewers have enough information and time, and whether anyone has sufficient authority to intervene. The BBC’s account is clearer about the authorization process than about post-deployment verification. It does not disclose how testing, reporting, audit or intervention mechanisms operate in classified work, so the source does not establish that safeguards are absent, inadequate or effective.

The distinction is consequential because a permission can be clear while its practical limits remain unclear. Restrictions on autonomous targeting, for example, depend on more than the presence of a human in the process. They depend on the human receiving usable information, having enough time to examine the output and possessing the authority to change the resulting decision.

Access, institutional interests and accountability

The institutional exchange has two immediate sides. The Pentagon receives access to Google models for lawful government purposes, while Google receives permission to undertake classified military work. The BBC report provides no contract value from which to draw a conclusion about private financial benefit. It establishes an exchange of military capability and contractor permission, but not the financial terms of that exchange.

Military policy assigns particular importance to deployment tempo. Defense Secretary Pete Hegseth made integrating AI into the armed forces a key priority and said in January that the administration would become an “AI-first warfighting force.” Emil Michael, an under secretary at the Department of Defense, said, “Speed defines victory in the AI era.” The Pentagon has maintained that adopting AI is essential to winning wars. Retired Lieutenant General Jack Shanahan, the former director of Project Maven, described the emerging philosophy as: “Go as fast as possible to deploy [AI technology] as widely as possible, potentially at the expense of the rigour and discipline of test and evaluation.” This perspective treats speed, operational relevance and wide deployment as material conditions of military AI rather than questions to defer until after deployment. It does not, by itself, establish that a safeguard works.

Google’s corporate-governance perspective centers on consistent review, permissible use and internal avenues for raising concerns. A spokesperson told the BBC that the company works with public-sector organizations worldwide and follows a consistent process for reviewing and entering contracts, including compliance with acceptable-use policies and AI principles. Google also points to dedicated employee channels for business-ethics and misconduct concerns and workplace policies intended to prevent retaliation. Those mechanisms describe how Google says authorization, reporting and non-retaliation are supposed to function. They do not provide mission-level evidence about whether the mechanisms operated as intended in classified work.

The worker and public-accountability perspective centers on whether people inside Google can identify problems, raise concerns and speak without damaging their careers. Liguori and Kirsch are its principal named representatives, alongside other current and former employees quoted by the BBC. Employment contracts prohibit employees from publicly criticizing the company, according to the report, and Liguori described a workplace atmosphere in which staff feared retaliation and career harm.

Three institutional perspectives therefore appear in the record: corporate governance, represented by Google’s documented controls; military policy, represented by Hegseth, Michael and Shanahan; and worker and public accountability, represented by Liguori, Kirsch and other employees. They answer different questions. Corporate governance asks how a deployment is authorized and constrained; military policy asks what operational capabilities the technology provides and how quickly they can be used; worker accountability asks who can challenge the company’s direction. None independently adjudicates whether a particular deployment was lawful, reliable or adequately supervised.

Across those perspectives, one normative proposition remains recognizable: AI deployment should not remove responsibility from people and institutions. The disagreement concerns how contractual restrictions, review practices and internal controls can make that responsibility operational.

Time, information and authority

Kirsch’s account separates two issues that contractual restrictions can otherwise blur. He told the BBC that Google “seems to have no right to control or veto how the Pentagon uses its AI and also seems to have limited visibility into how it’s being deployed.” Limited visibility concerns information: what Google can observe about a customer’s use of its technology. The claimed absence of veto authority concerns control: whether Google can prevent or alter a use that falls outside the agreed limits.

Those are separate questions. A judgment that existing oversight is insufficient does not by itself determine which decisions should remain Google’s to constrain and which should transfer to the Pentagon. The BBC report does not resolve that allocation of authority either.

Sarah Shoker, formerly head of OpenAI’s Geopolitics team, focused on the time available for review. She told the BBC: “When you have hundreds of targets that are selected by AI and a short amount of time, it is almost impossible for a human operator to meaningfully review the AI’s outputs for errors.” Shoker also said AI compresses the time an operator has to make a decision because it works faster than a person, which is especially risky in a high-stress environment.

That observation raises a reflexive question about human oversight. Documenting that a human remained involved does not establish that the human could identify errors or alter a decision. Shoker’s account makes meaningful intervention time-dependent, but it does not prove that oversight has failed in a particular Google mission. It supplies a condition against which actual operations would have to be tested.

Time, information and authority to alter an outcome are therefore the three practical conditions identified by Kirsch’s and Shoker’s accounts. Google’s AI Principles say the company implements AI with “human oversight” and aligns the technology with “international law and human rights.” Former and current Google staff and experts cited by the BBC expressed concern about oversight at critical moments, while tech workers said the military contracts signed that year greatly increased the scope of what the Pentagon could use the technology for.

Model reliability remains a related but incomplete answer to the governance question. Kirsch said current AI models were not robust enough and added that he did not think they had been trained for such tasks. Critics cited by the BBC questioned whether AI is reliable enough for warfare without more testing, stronger guardrails and human oversight. Those concerns do not encompass every dispute: Liguori focused on Google’s direction and stated ethical values, while Kirsch also raised deployment visibility and Google’s authority to constrain use. A more reliable model would not by itself settle who may veto a deployment, what review a mission requires, how employees can report concerns or how responsibility is assigned when AI contributes to an error.

The Pentagon’s own use of AI shows why the distinction matters, but the BBC account does not make Maven Smart Systems a Google-specific example. The United States military has deployed multiple AI-enabled technologies. The most prominent example cited by the BBC was Maven, the Pentagon’s main software platform for organizing intelligence and data. Developed by the US defense company Palantir, Maven uses AI to help personnel select and identify targets to strike. Palantir has integrated technology from several leading AI companies and enabled some processes to be significantly speeded up. The report does not identify a Google model among the technologies integrated into Maven. It therefore does not establish that Google’s controls produced—or failed to constrain—those speedups.

Why the resignations do not settle the question

Sofia Liguori, a London-based research engineer, resigned after three years at Google. She said the company’s increasing work with the US and Israeli militaries led her to lose faith in its stated ethical values. “Realising my company is pushing more and more for AI systems in military operations makes me feel ashamed to work at Google,” she said. Liguori spoke exclusively to the BBC despite a contract restricting her from doing so and spoke in a personal capacity. Her statements document how Google’s military direction affected her decision and how she perceived the company’s values. They are not findings about how a particular model was deployed.

Andreas Kirsch, an AI researcher, had resigned earlier in the period covered by the report. He said Google’s governance and oversight were “not sufficient or meaningful enough” to deploy AI safely and responsibly in military and classified settings, or to develop artificial superintelligence safely in the future. He also raised the separate deployment-visibility and veto-authority questions. His position combines an assessment of model readiness with concerns about who controls the technology and whether Google can observe its use.

One current Google DeepMind employee offered a possible explanation for why two departures need not indicate widespread attrition. The employee, who requested anonymity because of fear of retaliation, said a pipeline of early-career workers could replace motivated employees who joined under the company’s original mission, presenting this as a possible “changing of the guard.” Another anonymous employee said a lot of people had left who felt strongly about the company’s changing ethics position, with some people continuing to challenge the direction and others expecting a large company to remain unchanged. This account cautions against treating two resignations as proof of widespread departures. It does not explain why Liguori and Kirsch associated their decisions with military deployment, governance and restrictions on public criticism, and it does not establish the scale of any turnover.

One possible reading focuses on employee speech. The BBC reported that employment contracts prohibit public criticism and that staff are frightened to speak about AI safety and uses in warfare. Liguori said in principle she was prohibited from criticizing Google for the duration of her life, and she said, “people are scared of retaliation and putting their career at risk. You feel you are being watched, you can’t speak.” Google’s concern channels and anti-retaliation policy coexist with that reported experience. The source does not establish that retaliation occurred against Liguori or another employee. If people best placed to identify problems regard raising them as professionally hazardous, however, formal reporting mechanisms are difficult to test.

Another possible reading treats the resignations as evidence of a change in institutional direction. Under that account, Google’s mission may have shifted while formal contracts remained compliant, allowing employees attached to its earlier purpose to leave without demonstrating a broad institutional rejection of the new direction. The account identifies a possible mechanism, not its scale.

Three implementation mechanisms sit behind the competing accounts. Formal policies define acceptable use; contract terms establish permissions and restrictions; and employees’ ability to question deployment supplies an internal check on both. The available evidence does not show that any of those mechanisms has failed or identify a particular deployment that violated the Pentagon agreement. It does show unresolved concerns about model readiness, deployment visibility, meaningful review and the conditions under which employees can speak.

The wider pattern and the separate Israeli record

The BBC reported similar military agreements involving OpenAI, Amazon, Microsoft, SpaceX, Oracle, Nvidia and the start-up Reflection. Those agreements allow AI technology to be deployed in military settings such as transport, planning and battlefield operations. The list establishes that Google’s arrangements belong to a wider contracting pattern. It does not establish that the agreements have identical scope, restrictions or oversight arrangements.

Google’s agreements with Israel require separate treatment from the Pentagon agreement. In 2024, staff criticized a joint contract with Amazon worth $1.2bn (£960m) called Project Nimbus, which provides Israel’s government and military with AI technology and cloud-computing infrastructure. At the time, Google said Project Nimbus “is not directed at highly sensitive, classified, or military workloads relevant to weapons or intelligence services.”

Liguori linked her resignation to Google’s increasing work with both the US and Israeli militaries. Her statement does not establish that the Pentagon and Israeli agreements have equivalent terms or uses. Nor does it turn the Nimbus arrangement into evidence of a specific Google deployment under the Pentagon contract.

The documentary Naza examines the role of AI-powered targeting systems in bombarding the Gaza Strip and makes claims about resulting high civilian casualties. An Israeli government minister said the claims presented in the film were false and distorted. The source does not establish that the systems examined in Naza are the systems provided under Project Nimbus, or that either is connected to the Pentagon agreement or the resignations. The Pentagon contract, Project Nimbus and the systems examined in Naza therefore remain separate evidentiary strands, even though Liguori’s stated reason encompasses both US and Israeli military work.

A governance problem without a single solution

The structure matches the wicked-problem framework articulated by Horst Rittel and William J. Webber in their 1973 article Dilemmas in a General Theory of Planning. They characterized such problems as lacking a definite solution, explicit criteria and an agreed causal explanation. The competing assessments over Google’s military AI arrangements differ over what counts as success, what evidence would demonstrate it and which trade-offs should be accepted.

Several constraints interact. Additional contract review can establish approval, but it cannot by itself reveal later use. Additional human review encounters the time compression Shoker described. Employee reporting depends on people believing they can raise concerns without harming their careers. Classified work limits what public reporting can establish about monitoring. A safeguard in one part of the system can therefore become less effective when it meets another constraint.

The persistent trade-offs do not resolve through a single technical improvement. A more reliable model might reduce errors, but it would not by itself determine who may veto a use, what review a mission requires, how employees can report concerns or how responsibility is assigned when AI contributes to an error. Each mechanism can shift a residual risk rather than remove it. The competing readings define acceptable military-AI success differently, no single technical measure resolves their disagreement, and the trade-offs recur as models, missions and operating conditions change.

The control-centered assessment associated with Kirsch, Liguori and Shoker asks whether governance has enough information, review time and authority to constrain military uses. Liguori’s concern also asks whether Google can examine its own direction while employees face contractual and career constraints. The tempo-centered assessment associated with Hegseth, Michael and Shanahan treats speed, operational relevance and wide deployment as material conditions of military AI. The BBC report supports both assessments but does not establish that either is dominant or minority across the military-AI field. It also supplies no comparative evidence showing how governance arrangements function across Google and other contractors.

The assessments could cease to be contradictory in a particular mission if evidence showed that validation was adequate, reviewers had enough time, information and authority to affect outcomes, use remained within contractual restrictions and the deployment remained operationally useful. Assurance and tempo would then be tested against the same mission rather than treated as substitutes.

The failure condition is equally concrete. If operational speed makes meaningful intervention impossible, tempo advances at the cost of assurance unless review procedures change. A formally compliant process that no reviewer can meaningfully exercise does not resolve the tension between the two assessments.

Governance details that would make assurance testable

Any intervention expected to address the tension would need to specify who may suspend a deployment, what information an operator receives, how much decision time is protected, what post-deployment access or verification Google has and how responsibility is assigned when AI contributes to an error. Those details determine whether contractual restrictions function as governance or remain declarations.

Possible institutional measures include contract amendments providing audit access and deployment-review milestones; staffing and minimum-time requirements for human review; validation and incident-reporting conditions; verification of how employee-concern and anti-retaliation channels operate in classified work; and staged deployment gates.

Each measure carries a cost. Review milestones impose additional approval checkpoints. Minimum-time requirements require additional personnel and longer decision windows. Validation and incident reporting carry testing and reporting burdens. Protected concern channels require resources and procedures. Staged deployment gates reduce immediate availability. These are possible responses rather than recommendations, and the BBC report does not establish that Google has adopted any of them. Under the tempo-centered assessment, their costs are operational. Under the control-centered assessment, they are mechanisms for making oversight meaningful.

Evidence needed to resolve the question

The report establishes that two employees have left over a perceived mismatch among military deployment, corporate governance and Google’s stated values. Google documents contractual restrictions, review procedures, employee-concern channels and an anti-retaliation policy. Departing and current staff have raised unresolved concerns about model readiness, deployment visibility, meaningful review and employee speech.

The report does not establish widespread departures, a violation of Google’s AI Principles, prohibited use, an unsafe Google deployment, a contractual breach, or a connection to Project Nimbus or the systems examined in Naza. It also does not establish that Google’s safeguards are absent or adequate. The decisive missing evidence is operational rather than contractual: audit access, deployment logs, validation results, incident records, direct evidence about how human review operates in a Google military mission, and evidence about how employee-concern channels function in classified work.

The BBC report remains the controlling supplied source for the Pentagon contract language and the attributed statements by Shoker, Hegseth, Michael and Google. The available source material does not independently corroborate the exact contract text, Shoker’s remarks, Hegseth’s exact statement or Google’s original Project Nimbus statement. A supplied excerpt corroborates the related company list but not equivalent contractual terms. Comparisons with OpenAI, Amazon, Microsoft, SpaceX, Oracle, Nvidia and Reflection therefore cannot be represented as established contractual equivalence.

Until mission-level evidence is available, the resignations show a disagreement within Google about whether written authorization has become meaningful assurance. They do not establish that the controls have failed or that they have worked.

Analytical techniques used in this piece

This analysis applies the methods below. Each links to a short, plain-English explainer you can read and reuse.

Dialectical Analysis
Holds thesis against antithesis and works toward a higher synthesis.
Wicked Problems
Treats a problem as wicked — no stopping rule, no clean test of success, every attempt consequential.
Worldview Cartography
Maps the clashing worldviews underlying a dispute.