This is the companion technical deep-dive to How Trump's Anti-Woke AI Executive Order Could Actually Make LLMs More Biased. That article examines why the EO banning "woke AI" could paradoxically make LLMs more biased — for government and public users alike. A key part of that argument rests on how AI is actually being deployed to the US government. That is what this piece unpacks.

Trump's Executive Order 14319 requires that LLMs procured by the federal government be free from what it defines as ideological bias. But the order's consequences depend entirely on a question it doesn't answer: when AI companies adjust their models to comply, do those changes stay within government systems...or do they propagate to every user of those models worldwide?

To understand the risk, you need to understand how the infrastructure actually works. This article breaks that down.

Which AI Companies Have Government Contracts

Gemini for Government Image from Google Cloud

As of early 2026, the major frontier model providers with active US government relationships are:

OpenAI has signed a new contract with the US Department of Defense, and in March 2026 signed a deal to sell AI to US government agencies — including for classified workloads — through Amazon Web Services.
xAI (Grok) has entered a $0.42 per agency agreement with the US General Services Administration to accelerate federal AI adoption (September 2025).
Anthropic had won a Pentagon contract worth up to US$200 million in July 2025, but the relationship collapsed in February 2026 after Anthropic refused to allow unrestricted military use of its AI — specifically for domestic surveillance and autonomous weapons. OpenAI subsequently stepped in to fill the gap. According to the most recent statement by Anthropic's CEO in March 2026, they seem to be attempting to legally contest their exclusion from U.S. government contracts after being labelled a supply chain risk.
Meta has been making their LLaMA models available to U.S. Government agencies since November 2024 for the purpose of working on national security applications through partners like Amazon Web Services and Snowflake. The main difference between Llama and the other companies' models is that Llama is open-source, which means it is free to use for both public and government.
Like xAI, Google has entered an agreement with the US General Services Administration for a Gemini for Government offering, with agencies paying $0.47 per agency for Google’s AI tools. They explicitly acknowledge America's AI Action Plan, the one that the EO discussed in this article falls under.

The companies that remain under government contract are expected to be operating under the EO's requirements. All have built dedicated government product packages, which is a clear signal of sunk investment, and pressure to comply in order to recover it.

OpenAI for Government

OpenAI logo against US flag Image from PetaPixel

OpenAI launched OpenAI for Government on June 16, 2025 — a month before the EO was signed. Through the offering, they provide the US government with:

Their most capable models within secure and compliant environments, including through ChatGPT Enterprise and ChatGPT Gov
Custom models for national security, offered on a limited basis
Hands-on support
Insight into upcoming model capabilities so government customers can plan ahead

The models and products are tailored to government needs, but it is unclear whether the customization isolates EO-compliant configurations from the models the rest of the world uses.

ChatGPT Gov: What the Deployment Actually Looks Like

ChatGPT Gov is described as "a new tailored version of ChatGPT designed to provide U.S. government agencies with an additional way to access OpenAI's frontier models."

A snapshot of current partner offerings by Azure OpenAI Service to the US Government From Microsoft Azure's Dev Blog: Azure OpenAI Service now authorized for all U.S. Government data classification levels

The deployment is more technically isolated than the name alone implies. According to Microsoft Azure's developer blog, ChatGPT Gov runs as a containerized image deployed directly into the government agency's own Azure tenant — using services like AKS, Azure PostgreSQL, and Azure OpenAI (AOAI). The application itself lives within the government's infrastructure, not OpenAI's.

Critically, AOAI — Azure OpenAI Service — is Microsoft's hosted version of OpenAI's models. This means the infrastructure chain for government users is distinct from the one public users travel:

--- title: "Infrastructure Chains: Public vs. Government Access to OpenAI Models (for Educational Purposes)" --- graph TB PU["Public Users"] --> CGPT GA["US Government Agency"] --> CGOV subgraph OPENAI_INFRA["OpenAI Infrastructure"] CGPT["ChatGPT Public-facing web app"] OAPI["OpenAI API"] end subgraph AZURE_TENANT["Government Agency's Azure Tenant"] subgraph CONTAINER["Agency-hosted Container"] CGOV["ChatGPT Gov Containerized application"] end AOAI["Azure OpenAI Service (AOAI) Microsoft-hosted model endpoint"] end CGPT -->|API request| OAPI CGOV -->|API request| AOAI OAPI --> MODEL AOAI --> MODEL MODEL[("GPT-4o Model weights (shared or separate — unconfirmed)")] style PU fill:#fff3c1,stroke:#ffefae,color:#404040 style GA fill:#a2d4d1,stroke:#8cc9c5,color:#262626 style CGPT fill:#fda2a9,stroke:#fc8b94,color:#262626 style CGOV fill:#fda2a9,stroke:#fc8b94,color:#262626 style OAPI fill:#e8e8e8,stroke:#cccccc,color:#404040 style AOAI fill:#a2d4d1,stroke:#8cc9c5,color:#262626 style MODEL fill:#a2805d,stroke:#8b6a46,color:#ffffff style AZURE_TENANT fill:#ffffff,stroke:#8cc9c5,color:#404040 style CONTAINER fill:#a2d4d1,stroke:#8cc9c5,color:#262626

This picture became more complicated in October 2025, when Microsoft and OpenAI announced a restructuring of their partnership. One of its provisions is that OpenAI can now provide API access to US government national security customers regardless of the cloud provider — meaning OpenAI can serve these customers directly, bypassing Azure entirely.

Amazon logo and OpenAI logo Image from Amazon News

This matters because it makes the architecture more complex than a simple shared-model question. The application layer is separate. But the key unresolved variable is not the wrapper — it is the model weights that both paths ultimately call. The next two sections unpack what that means.

The Distinction Between ChatGPT, an AI Wrapper, and GPT, an LLM

While OpenAI's language suggests the LLMs offered to the US government could be customized and hence different from the ones the public uses, it is first important to understand the distinction between what ChatGPT is versus what the GPT models are.

ChatGPT is essentially an AI wrapper — a chatbot application that requests data from OpenAI's frontier models, including the various versions of GPT. It is not an LLM. Rather, it is an application that rides on an LLM.

You can watch our podcast episode discussing AI wrappers like ChatGPT below!

What This Means for the Public

The infrastructure isolation of ChatGPT Gov actually makes the "separate model" scenario — where a distinct, EO-compliant LLM is maintained for government use — more architecturally plausible than it might initially appear. Running entirely within the government's own Azure tenant gives OpenAI and Microsoft the technical means to point government traffic at a differently fine-tuned model, without touching the one public users access.

However, this does not eliminate the risk to public users. The unresolved variable is not the wrapper — it is what happens at the level of the foundational model weights. If OpenAI applies EO compliance changes during training or fine-tuning (for example, by modifying RLHF or removing feature steering at the model level), those changes propagate to every version of the model, regardless of how the application layer is deployed. Azure OpenAI Service typically serves the same underlying model weights as OpenAI's public API — Microsoft does not independently train them.

Microsoft logo and OpenAI logo Image from OpenAI

This is confirmed by the terms of the October 2025 Microsoft-OpenAI partnership restructuring, which explicitly excludes model architecture, model weights, inference code, and fine-tuning code from Microsoft's IP rights. However, the restructuring also introduced a national security carve-out allowing OpenAI to serve US government customers directly — adding a further layer of uncertainty about which model weights government systems are actually accessing.

The diagram below illustrates the scenario where the shared model scenario holds: ChatGPT and ChatGPT Gov are separate chatbot applications (AI wrappers) that both send API requests to the same underlying GPT-4o model. If EO compliance is baked into the model weights, the wrapper separation offers no protection for public users.

--- title: "Scenario A — Shared Model Weights (Risk Scenario, for Educational Purposes)" --- graph LR PU["Public Users"] -->|prompt| CGPT GA["US Government Agencies"] -->|prompt| CGOV subgraph WRAPPERS["Wrapper Layer — Chatbot Applications"] CGPT["ChatGPT OpenAI infrastructure Standard bias mitigations active"] CGOV["ChatGPT Gov Containerized in agency Azure tenant EO 14319 compliant"] end OAPI["OpenAI API"] AOAI["Azure OpenAI Service (AOAI) Microsoft-hosted model endpoint"] GPT4o[("GPT-4o Shared Foundation LLM Same model weights serve both paths EO compliance at this level affects all users")] CGPT -->|API request| OAPI CGOV -->|API request| AOAI OAPI --> GPT4o AOAI --> GPT4o GPT4o -->|response| OAPI GPT4o -->|response| AOAI OAPI -->|response| CGPT AOAI -->|response| CGOV style PU fill:#fff3c1,stroke:#ffefae,color:#404040 style GA fill:#a2d4d1,stroke:#8cc9c5,color:#262626 style CGPT fill:#fda2a9,stroke:#fc8b94,color:#262626 style CGOV fill:#a2d4d1,stroke:#8cc9c5,color:#262626 style OAPI fill:#e8e8e8,stroke:#cccccc,color:#404040 style AOAI fill:#a2d4d1,stroke:#8cc9c5,color:#262626 style GPT4o fill:#a2805d,stroke:#8b6a46,color:#ffffff style WRAPPERS fill:#ffffff,stroke:#cccccc,color:#404040

The Azure deployment detail makes the second scenario — a separately maintained, EO-compliant LLM — more credible. OpenAI also offers "Custom models for national security, offered on a limited basis", which signals both the willingness and the technical capacity to fork their model development for government use. Combined with the containerized, tenant-isolated deployment of ChatGPT Gov, a possible architecture for that separation might look like this:

--- title: "Scenario B — Separate Model Weights (Safer Scenario, for Educational Purposes)" --- graph TB PU["Public Users"] -->|prompt| CGPT GA["US Government Agencies"] -->|prompt| CGOV subgraph WRAPPERS["Wrapper Layer — Chatbot Applications"] CGPT["ChatGPT OpenAI infrastructure Standard bias mitigations active"] CGOV["ChatGPT Gov Containerized in agency Azure tenant EO 14319 compliant"] end OAPI["OpenAI API"] AOAI["Azure OpenAI Service (AOAI) Microsoft-hosted endpoint routes to government model"] subgraph TRAINING["Model Development & Versioning"] BASE[("Shared Pre-trained Base Common foundation weights before fine-tuning fork")] BASE -->|fine-tuned with standard safety & bias mitigation pipeline| GPT4o BASE -->|fine-tuned under EO 14319 compliance separate update cycle| GPT4oGOV end GPT4o[("GPT-4o Public Foundation LLM Bias mitigations intact Public users unaffected")] GPT4oGOV[("GPT-4o-Gov Government Foundation LLM EO 14319 compliant Separately versioned")] CGPT -->|API request| OAPI CGOV -->|API request| AOAI OAPI --> GPT4o AOAI --> GPT4oGOV GPT4o -->|response| OAPI GPT4oGOV -->|response| AOAI OAPI -->|response| CGPT AOAI -->|response| CGOV style PU fill:#fff3c1,stroke:#ffefae,color:#404040 style GA fill:#a2d4d1,stroke:#8cc9c5,color:#262626 style CGPT fill:#fda2a9,stroke:#fc8b94,color:#262626 style CGOV fill:#a2d4d1,stroke:#8cc9c5,color:#262626 style OAPI fill:#e8e8e8,stroke:#cccccc,color:#404040 style AOAI fill:#a2d4d1,stroke:#8cc9c5,color:#262626 style BASE fill:#a2805d,stroke:#8b6a46,color:#ffffff style GPT4o fill:#fda2a9,stroke:#fc8b94,color:#262626 style GPT4oGOV fill:#a2d4d1,stroke:#8cc9c5,color:#262626 style WRAPPERS fill:#ffffff,stroke:#cccccc,color:#404040 style TRAINING fill:#ffffff,stroke:#cccccc,color:#404040

Even so... These custom models OpenAI refers to might not mean customizations of its LLM range (i.e. the GPT-4o range), but rather new AI models trained from scratch that may not be LLMs in the first place.

Wait... Aren't AI Models LLMs?

Yes, this is an important distinction — while LLMs are AI models, AI models go well beyond LLMs alone.

--- title: A Simplified Family Tree of AI Model Types --- graph TD AI["AI Models"] LLM["LLMs Language Models"] NonLLM["Non-LLM AI Models Specialised for non-text modalities"] AI --> LLM AI --> NonLLM Frontier["Frontier LLMs General-purpose, state-of-the-art e.g. GPT-4o, Claude, Gemini, Grok, LLaMA"] Coding["Task-Specialised LLMs LLMs fine-tuned on specific task data e.g. coding-focused models like Codex"] LLM --> Frontier LLM --> Coding Video["Video Generation Models e.g. OpenAI SORA"] ASR["Speech Recognition Models (ASR) Converts audio to text e.g. OpenAI Whisper"] Image["Image Generation Models e.g. DALL·E, Midjourney"] NonLLM --> Video NonLLM --> ASR NonLLM --> Image Multimodal["Multimodal Capability Handles multiple input types simultaneously e.g. text + image in one prompt An increasingly common capability in modern frontier LLMs"] Frontier -.->|"evolving to support"| Multimodal style AI fill:#8b5a49,stroke:#764735,color:#ffffff style LLM fill:#fda2a9,stroke:#fc8b94,color:#262626 style NonLLM fill:#a2d4d1,stroke:#8cc9c5,color:#262626 style Frontier fill:#fdb9be,stroke:#fda2a9,color:#262626 style Coding fill:#fdb9be,stroke:#fda2a9,color:#262626 style Video fill:#b8dedd,stroke:#a2d4d1,color:#262626 style ASR fill:#b8dedd,stroke:#a2d4d1,color:#262626 style Image fill:#b8dedd,stroke:#a2d4d1,color:#262626 style Multimodal fill:#fff3c1,stroke:#ffefae,color:#404040

In the LLM family, frontier models are general-purpose models — GPT-4o, Claude, Gemini, Grok, LLaMA. Coding models are still LLMs: language models fine-tuned on large corpora of code, optimized for software engineering tasks rather than general conversation.

Then there are specialised models for other modalities: video generation models like OpenAI's SORA (being discontinued April 26, 2026), and automatic speech recognition (ASR) models like OpenAI's Whisper, which converts spoken audio to text. And there are multimodal models — a single model that handles multiple input types simultaneously. Modern frontier LLMs like GPT-4o and Gemini are increasingly multimodal: you can feed them a photo and ask a question about it in the same prompt.

There are a total of 83 models OpenAI has made publicly known on their official website, including those that have been deprecated.

This means that OpenAI's OpenAI for Government offering — which includes "Custom models for national security, offered on a limited basis" — may or may not include customization of their LLMs specifically. It remains unconfirmed how OpenAI intends to comply with the EO, and whether their compliance would affect users beyond the US Government.

OpenAI's Expansion Through AWS

OpenAI's government footprint expanded significantly beyond Azure in 2026. In November 2025, OpenAI and AWS announced a $38 billion multi-year strategic partnership, making OpenAI's models available through Amazon Bedrock — AWS's managed AI service. This was directly tied to the national security carve-out in the October 2025 Microsoft-OpenAI restructuring, which permitted OpenAI to serve US government national security customers regardless of cloud provider.

Robotic Hand over the background of OpenAI and AWS logos Image from The Straits Times obtained through Reuters

That provision was exercised almost immediately. By March 2026, OpenAI had signed a deal to sell AI to US government agencies through AWS, covering both unclassified and classified workloads. OpenAI's models are now available in AWS GovCloud and AWS Classified Regions — environments rated for Secret and Top Secret workloads — giving the Pentagon and intelligence agencies access to OpenAI models through an entirely separate infrastructure chain from Azure.

Government agencies can now access OpenAI models through at least two distinct cloud paths — Azure OpenAI Service and AWS GovCloud — and potentially directly from OpenAI for national security workloads. OpenAI retains control over which models are made available through AWS, and requires advance notice before AWS can enable access for especially sensitive intelligence customers. But the public has no visibility into which model versions, fine-tunes, or EO compliance configurations are deployed in each environment.

--- title: "OpenAI Government Access Pathways — and the Model Weights Question (for Educational Purposes)" --- flowchart TD subgraph agencies[" "] direction LR AG1["🏛️ Standard Federal Agency\n(unclassified workloads)"] AG2["🔒 Pentagon /\nIntelligence Agency\n(Secret / Top Secret)"] end subgraph paths["Access Pathways"] direction LR P1["☁️ Azure OpenAI Service\n(Microsoft-hosted,\nagency Azure tenant)"] P2["☁️ AWS GovCloud /\nAWS Classified Regions\n(Secret & Top Secret rated)"] P3["🔗 Direct from OpenAI\n(national security\ncarve-out, Oct 2025)"] end AG1 -->|"Path 1"| P1 AG2 -->|"Path 2"| P2 AG2 -->|"Path 3\n(limited basis)"| P3 subgraph unknown["❓ Model Weights — Unconfirmed"] direction LR MW1["Model Version A?\nEO-compliant fine-tune?"] MW2["Model Version B?\nSeparate weights?"] MW3["Model Version C?\nCustom national\nsecurity model?"] end P1 -->|"Which version?\nWhich fine-tune?"| MW1 P2 -->|"Which version?\nWhich fine-tune?"| MW2 P3 -->|"Which version?\nWhich fine-tune?"| MW3 PUBLIC["🌍 Public Users\n(ChatGPT / API)"] PUBMODEL["GPT-4o\n(public model weights)"] PUBLIC --> PUBMODEL PUBMODEL -. "Same weights as\ngovernment paths?\nUnknown." .- MW1 PUBMODEL -. "Same weights as\ngovernment paths?\nUnknown." .- MW2 PUBMODEL -. "Same weights as\ngovernment paths?\nUnknown." .- MW3 style AG1 fill:#fff3c1,stroke:#ffefae,color:#404040 style AG2 fill:#a2d4d1,stroke:#8cc9c5,color:#262626 style P1 fill:#fda2a9,stroke:#fc8b94,color:#262626 style P2 fill:#fda2a9,stroke:#fc8b94,color:#262626 style P3 fill:#fda2a9,stroke:#fc8b94,color:#262626 style MW1 fill:#a2805d,stroke:#8b6a46,color:#ffffff style MW2 fill:#a2805d,stroke:#8b6a46,color:#ffffff style MW3 fill:#a2805d,stroke:#8b6a46,color:#ffffff style PUBLIC fill:#fff3c1,stroke:#ffefae,color:#404040 style PUBMODEL fill:#fdb9be,stroke:#fda2a9,color:#262626

The timing is also notable: OpenAI's AWS government deal came directly after Anthropic's Pentagon contract collapsed in February 2026, when Anthropic refused to allow unrestricted military use of its AI — specifically for domestic surveillance and autonomous weapons. OpenAI, now the last major frontier model provider with Pentagon access, is distributing that access through two of the world's largest cloud platforms.

What Remains Unknown

While this article does a deepdive on OpenAI and does not expand into other companies' model offerings to government, across all major models government offerings, the same gap exists: the public has no visibility into which specific model versions are deployed in government environments, what EO compliance configurations have been applied, or whether those configurations affect the foundational model weights that public users share.

The companies operating under government contracts are the same companies whose models are used by students, doctors, job seekers, and small business owners around the world. Until this is transparent, the question of whether government AI policy is shaping the AI tools the rest of us use every day cannot be answered.

For the broader argument about why the EO's approach to bias is technically flawed — and why banning bias correction could make LLMs more biased for everyone — read the companion piece: How Trump's Anti-Woke AI Executive Order Could Actually Make LLMs More Biased.

ragTech is a podcast by Natasha Ann Lum, Saloni Kaur, and Victoria Lo where real people talk about real life in tech. Our mission is to simplify technology and make it accessible to everyone. We believe that tech shouldn't be intimidating, it should be fun, engaging, and easy to understand!

✨ragTech Spotify: https://open.spotify.com/show/1KfM9JTWsDQ5QoMYEh489d

✨ragTech YouTube: https://www.youtube.com/@ragTechDev

✨Instagram: https://instagram.com/ragtechdev

✨Other Links: https://linktr.ee/ragtechdev