Exclusive: Elon Musk's DOGE Used Meta's Llama AI to Analyze Federal Worker Emails

Elon Musk’s ambitious, and often controversial, initiative known as the Department of Government Efficiency, or DOGE, has reportedly deployed artificial intelligence developed by Meta to sift through and analyze communications from within the federal workforce. This revelation, based on materials viewed by WIRED, indicates a significant step in the new administration's push to integrate advanced AI tools into the machinery of government, particularly within the Office of Personnel Management (OPM), the federal government's central human resources agency.

Specifically, DOGE affiliates operating within OPM tested and utilized Meta’s Llama 2 model. The primary task assigned to this AI was to review and classify responses from federal workers to a now-infamous email, widely dubbed the “Fork in the Road” email, which was disseminated across government agencies in late January.

The Genesis of the "Fork in the Road" Email

The "Fork in the Road" email was a stark directive, offering federal employees a clear choice: accept significant changes being implemented by the Trump administration or opt for deferred resignation. These changes included a mandated return-to-office policy, potential downsizing initiatives, and a requirement for employees to demonstrate "loyalty." The email's structure and tone bore a striking resemblance to a similar ultimatum issued by Elon Musk to employees at Twitter (now X) shortly after his acquisition of the social media platform in 2022. To accept the resignation offer, employees simply needed to reply with the word "resign."

This approach, while perhaps efficient from a purely logistical standpoint, sent ripples of uncertainty and anxiety through the federal workforce. It represented a dramatic departure from traditional government communication and personnel management practices, signaling a new era of direct, top-down directives aimed at rapidly reshaping the civil service.

Llama 2: The AI Classifier

According to the materials reviewed, Meta's Llama 2 model was specifically deployed to process the influx of responses generated by the "Fork in the Road" email. Its task was to sort through these emails and determine how many federal workers chose to accept the offer of deferred resignation by replying with the designated keyword. The use of AI for this purpose suggests an attempt to automate the process of quantifying the impact of the directive and identifying those who chose to leave.

Crucially, the materials indicate that the Llama 2 model appears to have been run locally. This detail is significant as it suggests the processing of potentially sensitive government emails occurred within OPM's own infrastructure, rather than being sent to external servers or cloud services managed by Meta or a third party. Running the model locally could mitigate some data security and privacy concerns, although questions about the handling and storage of the data, as well as the model's accuracy and potential biases in classification, would still remain.

Neither Meta nor OPM provided comments when contacted by WIRED regarding the use of Llama 2 for this specific task. This lack of official confirmation leaves many details about the scope and specifics of the AI's deployment, including the exact number of emails processed and the outcomes of the analysis, undisclosed.

The Open-Source Advantage (and Challenge)

Meta CEO Mark Zuckerberg has been seen alongside other prominent tech leaders, including Elon Musk and Jeff Bezos, at events signaling proximity to the new administration. However, the extent to which Meta's technology is being directly utilized within government operations has largely remained out of the public eye. The case of Llama 2 at OPM offers a glimpse into this. Because Llama models are open-source, they are freely available for anyone to download, modify, and deploy. This open nature means that government agencies or affiliated groups like DOGE can utilize the technology without requiring explicit consent, formal agreements, or direct technical support from Meta. This accessibility allows for rapid deployment but also potentially bypasses traditional procurement processes and oversight mechanisms that might apply to proprietary software or services.

The ease with which an open-source model can be integrated into government workflows presents both opportunities for innovation and challenges for accountability and transparency. While it can accelerate the adoption of powerful AI capabilities, it also makes it harder to track exactly how and where these tools are being used, who is using them, and what safeguards are in place.

DOGE's Entrenchment at OPM

Soon after the new administration took office in January, DOGE operatives reportedly embedded themselves within OPM. As an independent agency, OPM plays a vital role as the federal government's central human resources department, managing everything from hiring and benefits to retirement and workforce policy. The presence of DOGE affiliates within this critical agency signaled the administration's intent to implement rapid, potentially disruptive changes to the federal workforce.

One of the initial major goals for DOGE within OPM was the creation of a new, government-wide email service. This project, according to current and former OPM employees, involved individuals like Riccardo Biasini, a former Tesla engineer, who was reportedly involved in building the infrastructure for the system that would eventually send out the "Fork in the Road" email. This suggests a coordinated effort to establish new communication channels and tools that could facilitate the administration's workforce management objectives.

Beyond the "Fork": The "Five Points" Directive

Weeks after the initial "Fork in the Road" email, OPM issued another directive to all government workers in late February. This subsequent request asked employees to submit five bullet points each week outlining their accomplishments. This instruction, like the first, generated significant confusion and, in some agencies, led to what has been described as chaos. Federal workers grappled with how to comply with the directive while remaining mindful of security clearances, the need to protect sensitive information, and the sheer volume of reporting required.

Adding to the uncertainty, some reports indicated that responses to these weekly "five points" emails might not have even been opened, with some workers who enabled read receipts finding that their submissions remained unread. This raised questions about the true purpose of the directive and how the collected information was intended to be used.

NBC News reported in February that these weekly emails were expected to be fed into an AI system for analysis. While the materials viewed by WIRED specifically detail the use of Meta's Llama models for the "Fork in the Road" emails, they do not explicitly confirm their use for the "five points" submissions. However, two federal workers suggested to WIRED that reusing the code and system developed for the "Fork" emails for the "five points" analysis would be a logical and straightforward step if DOGE intended to process that data using AI.

As one federal worker put it, "We don’t know for sure [if Llama was used for the five points emails]... Though if they were smart they’d reuse their code.” This highlights the potential for rapid, undocumented expansion of AI use once a system is in place.

A Growing Portfolio of AI Tools

The use of Llama 2 for email analysis is not an isolated incident but rather part of a broader pattern of DOGE rolling out and testing various AI-based tools across government agencies over the past few months. These initiatives appear aimed at implementing the administration's agenda, often with a focus on efficiency, productivity, and potentially, workforce restructuring.

In March, WIRED reported on the US Army's use of a tool called CamoGPT. This AI was reportedly designed to identify and remove language related to Diversity, Equity, and Inclusion (DEI) from training materials, reflecting a specific ideological goal of the administration.

The General Services Administration (GSA), another key independent agency, also launched its own AI tool earlier this year, dubbed "GSAi." This chatbot was presented as a means to boost overall agency productivity, offering automated assistance and information access to GSA employees.

Furthermore, OPM itself has reportedly accessed software known as AutoRIF. This tool is designed to assist with the process of Reduction in Force (RIF), a mechanism used for mass firings of federal workers. The availability and potential use of such a tool, especially in conjunction with AI-driven analysis of workforce data, raises significant concerns about job security and the potential for automated personnel decisions.

The Potential Entry of Grok

While Meta's Llama 2 was reportedly used for the initial email analysis, Elon Musk's own AI model, Grok, developed by his company xAI, did not appear to be the primary tool used in the early weeks of the administration's government-wide email system project. At that time, Grok was a proprietary model with limited API access, making it less readily available for widespread government deployment compared to an open-source alternative like Llama 2.

However, the landscape is rapidly changing. Earlier this week, Microsoft announced that it would begin hosting xAI's Grok 3 models as options within its Azure AI Foundry. This development significantly increases the accessibility of Grok models, particularly within Microsoft environments, which are widely used across the federal government, including at OPM. This move potentially paves the way for Grok to become a viable AI option for DOGE's future initiatives, should they choose to integrate it.

Adding to Grok's potential reach within government, Palantir, a company with extensive contracts across federal agencies, struck a deal in February to include Grok as an AI option within its software platforms. Given Palantir's deep integration into various government data systems, the inclusion of Grok in their offerings could further facilitate its adoption for government tasks.

The potential shift towards or inclusion of Grok alongside models like Llama raises further questions about the criteria for selecting AI tools for government use, the evaluation of their capabilities and biases, and the influence of specific tech leaders and companies on public sector technology adoption.

Implications and Concerns

The reported use of Meta's Llama 2 by DOGE within OPM, coupled with the rollout of other AI tools like CamoGPT, GSAi, and AutoRIF, highlights several critical implications and concerns regarding the rapid integration of AI into the federal government:

• Transparency and Oversight: The use of AI models, especially open-source ones deployed locally, can make it difficult to track exactly how decisions are being made or data is being analyzed. This lack of transparency hinders effective oversight by Congress, watchdog groups, and the public.
• Data Privacy and Security: While local deployment of Llama 2 might reduce some risks, the sheer volume of sensitive information contained in federal worker emails raises significant data privacy and security concerns. How is this data being stored, protected, and used beyond the initial classification task?
• Algorithmic Bias: AI models, including large language models like Llama 2, can inherit biases from the data they were trained on. If used for tasks like classifying employee responses or potentially evaluating performance or loyalty, these biases could lead to unfair or discriminatory outcomes for federal workers.
• Workforce Impact: The use of AI for tasks like analyzing responses to resignation offers or assisting with mass firings (via tools like AutoRIF) directly impacts the federal workforce. It raises questions about job security, the nature of work, and the potential for automated personnel decisions without adequate human review or appeal processes.
• Influence of Private Tech: The deep involvement of figures like Elon Musk and the use of tools from major tech companies like Meta, Microsoft, and Palantir raise questions about the influence of private sector interests and technologies on public sector functions and policies.
• Policy and Governance Gap: The rapid deployment of AI tools appears to be outpacing the development of clear, comprehensive policies and governance frameworks for their ethical and responsible use within the federal government.

The Future of AI in Public Service

The events at OPM and other agencies under the purview of DOGE signal a determined effort to leverage artificial intelligence to reshape the federal government. Proponents might argue that AI can bring much-needed efficiency and data-driven decision-making to bureaucratic processes. However, the manner in which these tools are being introduced, often with limited transparency and seemingly tied to specific political objectives, raises serious alarms.

The use of AI to analyze employee sentiment or compliance, particularly in response to directives perceived as loyalty tests or precursors to downsizing, ventures into sensitive territory. It blurs the lines between administrative efficiency and surveillance or ideological screening.

As AI becomes more powerful and accessible, its integration into public service is inevitable. However, the case of DOGE's activities underscores the critical need for robust ethical guidelines, clear governance structures, independent oversight, and a focus on ensuring that AI is used to serve the public good, not specific political agendas or the interests of private tech companies. The experience with the "Fork in the Road" email and the subsequent AI analysis serves as a cautionary tale about the potential pitfalls of rapid, opaque AI deployment in the core functions of government.

The coming months will likely reveal more about the extent and impact of DOGE's AI initiatives. The potential for models like Grok to enter this ecosystem further complicates the picture, adding another layer of corporate influence and technical complexity. Ensuring accountability, protecting federal workers' rights and privacy, and maintaining the integrity of government operations in the age of AI will be paramount challenges.

The narrative unfolding within OPM and other agencies is not just about technological adoption; it is about the fundamental nature of the civil service, the relationship between the government and its employees, and the democratic principles of transparency and accountability in the digital age. The quiet deployment of AI to analyze sensitive communications is a stark reminder that the future of government efficiency is deeply intertwined with complex questions of ethics, power, and control.