As immigration agents have carried out the Trump administration’s mass deportation campaign, they have used a range of AI-driven technologies – from facial recognition mobile apps to social media monitoring software – to locate individuals and to monitor peaceful protesters. While the administration has revealed few details about these technologies, some much-needed transparency could have come from the Department of Homeland Security’s (DHS) most recent AI use case inventory, published in January of this year (with some revisions in June). By law, federal agencies must publish annual inventories of how they use artificial intelligence technologies (which the administration now calls super intelligence), and OMB guidance requires them to disclose how they manage the risks those technologies pose to people’s rights. As an examination of DHS’s current AI inventory reveals, however, the law has failed to ensure transparency. The information this administration requires its agencies to disclose about their use of AI is too limited to enable accountability, and DHS fails to meet even that minimal standard.
Background
The requirement for agencies to publish AI use-case inventories began with an executive order signed by President Trump in 2020. It was later codified into law in the Advancing American AI Act in 2022, which requires each agency to prepare, maintain, and publicly post an inventory of its AI use cases. President Biden’s Office of Management and Budget (OMB) issued a memo in 2024 that directed federal agencies to implement certain safeguards when using AI in ways that could significantly affect people’s rights or safety and required agencies to disclose how they had implemented those safeguards in their inventories. In April 2025, OMB rescinded and replaced the Biden-era memo on federal use of AI. While the 2025 OMB memo has many similarities to the 2024 version, it cut some crucial protections, such as the requirement that agencies measure and mitigate bias. More recently, DHS issued a directive on the use and acquisition of artificial intelligence that permits the chief AI officer to waive any departmental policy governing AI, including any prohibitions on uses; while the directive adds that waivers may not violate “law, regulation, and government-wide policy,” this authority creates an enormous loophole in the existing guidance.
The inventories that DHS released during the Biden administration had their own transparency problems. Moreover, many of the concerning surveillance tools that Customs and Border Protection (CBP) and U.S. Immigration and Customs Enforcement (ICE) are using today have been in place for years. However, the costs of inadequate transparency are higher now, under an administration that seeks to arrest thousands of migrants a day while undermining efforts to hold federal agents accountable for abuses of power.
How DHS Avoids Required Safeguards
Under the 2025 OMB memo, agencies must report whether a given use of AI is “high-impact” in light of its potential effects on individuals’ rights or safety. The memo requires agencies to implement and disclose certain minimum safeguards for all high-impact AI use cases, including testing the system’s accuracy before deployment, auditing for misuse, and establishing a process to appeal decisions made using AI.
An agency’s use of an AI tool is “high-impact” only if its outputs are “a principal basis” for an official action that affects someone’s rights or safety. As we have written previously, this standard, which the Biden-era memo also relied upon, is far too narrow. Agencies often consider many factors when making decisions, such as a determination by ICE or CBP agents to investigate or detain someone. If an AI’s recommendations, predictions, or other outputs influence such a consequential decision, then agents’ use of that AI should be subject to safeguards, whether or not the AI is a “principal” basis for the decision.
Applying OMB’s narrow standard, DHS avoids implementing safeguards for many uses of AI that directly affect people’s rights. For example, the inventory does not classify the ELITE app as high-impact, although ICE agents use it to find and arrest migrants who may be in the country illegally. The app uses large language models to extract addresses where migrants may live from government records, including data from the Department of Health and Human Services, and displays them on maps, along with “confidence scores” indicating the likelihood that the addresses are accurate, facilitating widescale raids.
DHS also excludes the Investigative Prioritization Aggregator software, which ranks suspects in criminal investigations by Homeland Security Investigations (HSI), a component of ICE, according to their likely importance to an investigation. Detaining or investigating people affects their rights by definition, and the risks are heightened under an administration that is not scrupulously abiding by its legal obligations and has carried out lethal military strikes of people on boats accused of smuggling drugs. Whether or not these AI systems provide a “principal basis” for decisions about whom to arrest or investigate, they directly shape those decisions.
Similarly, the department determined that its use of Dataminr, a tool that scans social media and other public sources to “provide alerts for possible threats related to national security [and] border violence,” among other issues, was not high-impact, because the alerts just supply CBP officers with compiled information. This overlooks the substantial concerns associated with use of social media data, including the barriers to accurately decoding social media communications and the risk of inaccurately labeling someone a threat based on tainted, misinterpreted, or inadequately vetted data.
DHS’s labels have consequences. Excluding technologies from the high-impact list allows DHS to skip safeguards—such as testing and auditing—making it easier for CBP or ICE to deploy technologies that are inaccurate or unreliable, increasing the risk of agents arresting innocent people. Wrongful arrest attempts can be deadly, as tragically underscored when, in two separate incidents during a single week in July, ICE agents killed Lorenzo Salgado Araujo and Joan Sebastian Guerrero, after apparently confusing each of them with other individuals. Interactions with CBP officers can be high stakes as well, from searches of travelers’ electronic devices at the border to fatal encounters. Safeguards are also needed to prevent intentional misuse of AI systems, as detailed below.
DHS Falls Short of the Agency’s Own Rules
The 2025 OMB memo instructs agencies to presume that certain uses of AI are high-impact, and rightly includes activities such as identifying criminal suspects and assessing risks posed by individuals. In follow-up guidance released a few months after the memo, OMB further instructs agencies to explain any deviations from this list.
However, DHS has not always done so. For example, although identifying criminal suspects is presumptively high-impact, DHS does not explain why the “AI-Enhanced ICE Tip Processing” system—which summarizes and ranks the urgency of tips submitted to ICE—is not listed as high-impact. This is not a harmless bureaucratic error. Determining which tips should be investigated determines which threats to public safety are prioritized, affecting the rights of individuals deemed to be suspects and potential crime victims alike.
What the Inventory Doesn’t Say
For high-impact uses of AI, the OMB memo requires agencies to implement certain measures to manage the technologies’ risks. ICE classifies six and CBP classifies 19 actively deployed use cases as high-impact, but the inventory reveals little about what safeguards are in place. OMB’s follow-up instructions on the AI inventories require agencies to answer nine questions about their compliance with the minimum safeguards for high-impact use cases. All but one of those questions are multiple choice, each with a list of three to six possible responses that convey very little.
For example, CBP and ICE have deployed a high-impact facial recognition app called Mobile Fortify, which agents use to identify individuals they stop to question, potentially matching them to records in federal databases. This is indeed high-stakes technology. A false match could result in wrongful arrest or deportation. However, the agencies are not required to tell the public any meaningful details about their efforts to mitigate these risks. In answer to a question about whether they monitor the use of Mobile Fortify for harms to privacy, civil rights, or civil liberties, DHS simply selected a response that says: “Yes, sufficient monitoring protocols have been established.”
This generic response is essentially useless, particularly in light of DHS’s removal of the only relevant policy from its website, along with reporting suggesting that ICE officials have told agents they can rely on Mobile Fortify results as “definitive” evidence of someone’s migration status and can ignore contrary evidence – even a birth certificate. The administration’s decision to stop requiring agencies to address demographic disparities in AI models compounds this problem, since facial recognition systems are generally less accurate on women and minorities.
In addition, many of the questions that the Biden-era OMB instructions required agencies to answer were dropped from the Trump-era instructions, including:
- whether the AI use case has been tested in operational environments to understand its performance and potential impact on affected individuals or communities;
- whether a senior privacy official has assessed the privacy risks;
- whether the AI is used to significantly influence or inform decisions or actions that could have a negative impact on specific individuals or groups; and
- what steps the agency has taken to detect and mitigate significant disparities in performance across demographic groups.
Excluding these kinds of granular questions deprives the public of important insights into how their government is using AI.
Another gap in transparency stems from the fact that agencies are only legally required to release inventories annually, which allows them to deploy technologies in potentially harmful ways for months before the public learns much if anything about them. For instance, Mobile Fortify became operational at the start of May 2025, according to the inventory, but its existence would have remained a secret until January 2026 if 404 Media had not broken the story. As it was, the public did not learn key details, such as the company that developed the app, until the inventory was published; based on that information, WIRED reported that the app was not designed to confirm identities, as ICE agents are reportedly doing. If ICE agents deployed new AI surveillance tools during Operation Metro Surge in Minnesota in January 2026, the public might not learn about it until 2027.
Similarly, recent reporting indicates that U.S. Citizenship and Immigration Services (USCIS) is seeking to hire a contractor who will use artificial intelligence tools to help sift through mountains of governmental and commercial data as part of an effort to ramp up criminal investigations of fraud in immigration benefit applications. If the contract is awarded this year, an analysis of the impact of the tools and any risk mitigation efforts would be released in January 2027 at the earliest; if it is not awarded until next year, any disclosures are not likely to appear in the inventory until 2028.
ICE’s Contradictory Use of Sentiment Analysis
The inventory also reveals that ICE is violating its own policies on internet surveillance by employing automated sentiment analysis. An entry titled “Open-Source Intelligence for Investigations” reveals that HSI uses software systems that search online content for undefined “risks” and suggest “investigative leads.” These systems utilize sentiment analysis: AI methods for classifying the emotions or opinions expressed in text. The entry itself cites an official ICE policy that states the agency is prohibited from applying sentiment analysis technologies to online content during law enforcement investigations.
The policy as written is right to ban sentiment analysis as a tool of criminal investigations. There is little reason to think that sentiment analysis systems could reliably identify real threats to public safety or legitimate criminal investigative leads, since AI models often misinterpret the tone or context of informal comments people make online. Even if such models were perfectly accurate, equipping ICE investigators with tools to search social media for ideas they consider threatening is ripe for abuse, particularly in an administration that conflates criticism with threats and regards viewpoints it dislikes as signs of terrorism. Moreover, Trump administration officials openly announced their intentions to use AI to monitor the internet for perceived support for Hamas among foreign students, which has often meant simply expressions of support for Palestinian rights, and ICE has sought to monitor social media for criticism of the agency. If this is the kind of AI usage the administration is willing to tell us about voluntarily, stronger transparency requirements would likely reveal considerably more.
What Congress Should Do
In the long term, Congress should pass several legislative reforms to rein in dangerous uses of AI. The Advancing American AI Act leaves it to each administration to decide what details agencies must disclose, which the current administration has interpreted narrowly to weaken the requirement. The Act should be amended to require federal agencies to disclose more meaningful detail in their inventories about how they are using AI. That must include reporting on the safeguards that agencies have put in place for any AI system that impacts people’s rights or safety. To limit ambiguity, the law should explicitly list certain applications of AI for which agencies must report safeguards, including AI systems used to inform decisions about whom to investigate, arrest, or deport.
Agencies should also be required to update their inventories when a new use of AI is rolled out that has a significant impact – or would be presumed to have such an impact – on rights or safety. In the meantime, the House and Senate committees with jurisdiction over DHS ought to exercise their full oversight powers to bring this administration’s use of AI surveillance technologies out of the shadows, including compelling DHS officials to answer questions under oath.
The post DHS’s AI Inventory Reveals the Limits of Federal AI Transparency Law appeared first on Just Security.