Resistance Survival Guide #318
Government agencies are rapidly adopting artificial intelligence to review records, identify patterns, evaluate applications, detect suspected fraud, monitor borders, analyze images, sort tips, and assist with decisions that can alter real lives. Many federal agencies must publish inventories describing these systems.
That sounds reassuring until you remember who completes the inventory.
An artificial intelligence inventory is not an independent audit. It is a disclosure prepared by the same institution buying, developing, testing, and operating the technology. The inventory may be accurate. It may be incomplete. It may also describe a consequential surveillance system as an exciting efficiency tool that helps dedicated professionals leverage innovative solutions. Bureaucracy loves a thesaurus when the plain English version sounds alarming.
Today, we are going to audit the audit.
What Is a Government Artificial Intelligence Inventory?
A government artificial intelligence inventory is a list of systems that an agency identifies as using artificial intelligence. Entries may describe the purpose of a system, the agency component responsible for it, its development stage, the data it uses, the vendor involved, and whether officials consider it capable of affecting rights or safety.
Federal guidance requires covered agencies to maintain annual inventories and publish public versions. The current guidance also requires agencies to continue reporting their artificial intelligence uses, compliance plans, and related governance information. However, some uses may be excluded from public inventories because of national security restrictions, legal limitations, classification, or other authorized exemptions.
That distinction matters. A missing system is not automatic proof of misconduct. It is a reason to investigate.
The Government Accountability Office found incomplete and inaccurate information in federal artificial intelligence inventories. Only five of the 20 agencies it reviewed had provided comprehensive information for every reported use case. Some entries lacked required details. Other entries described technology that agencies later determined was not actually artificial intelligence.
In March 2026, the Government Accountability Office reported that the Internal Revenue Service inventory was incomplete. It did not include every way the agency was using artificial intelligence or consistently explain how the technology would benefit the agency.
So yes, the government occasionally loses track of its own artificial intelligence. I am sure the machine making decisions about your tax return is much more organized.
Why These Inventories Matter
An inventory can reveal where an agency is placing automated systems between the public and government power. It may show artificial intelligence being used to screen applications, assess risk, analyze evidence, identify supposed anomalies, monitor communications, recognize faces, or recommend enforcement actions.
Those descriptions can help researchers ask much better questions.
Who supplied the technology? What information enters the system? Who can search it? What decision does it influence? Was it tested on the population it now evaluates? Can a human reject its recommendation? Does the affected person know the system was involved? Is there a meaningful way to challenge an error?
An inventory cannot answer every question. It can give you the first thread to pull.
The Difference Between an Error and an Omission
Finding a contract that does not appear in an artificial intelligence inventory does not immediately prove that the agency concealed a system.
The product may not use artificial intelligence. The contract may support infrastructure rather than a specific use case. The project may have been canceled. It may remain in an early research stage. It may appear under a different name. It may belong to another agency component. It may also fall within an authorized reporting exclusion.
Your job is not to begin with a verdict. Your job is to document the discrepancy, test reasonable explanations, and determine whether the public description matches the available evidence.
Receipts first. Outrage after lunch.
Step by Step Guide
Step One: Choose One Agency and Download Its Inventory
Begin with one federal department or agency. Look for an artificial intelligence page, use case inventory, compliance plan, Chief Artificial Intelligence Officer page, or downloadable spreadsheet.
Download the original file whenever possible. Record the web address, publication date, date accessed, file format, and any update notice displayed on the page. Save a copy instead of relying entirely on a live page that can change without calling you first.
The Department of Homeland Security maintains a public artificial intelligence inventory that can be used as a practical example. Its January 2026 annual refresh includes uses across components such as Customs and Border Protection, the Transportation Security Administration, United States Citizenship and Immigration Services, and the Secret Service.
Create a research folder for the agency. Keep the original inventory untouched. Conduct your analysis on a separate copy so you always have the source document available for comparison.
Step Two: Preserve Earlier Versions
Search the agency website, its document library, the Internet Archive, congressional reports, and public records repositories for earlier inventories.
Compare the versions by year. Note which systems were added, removed, renamed, combined, divided, or moved between agency components. Pay particular attention to a system that disappears without being marked as retired.
A missing entry can have an innocent explanation. It can also mean the agency changed the system name, changed its definition of artificial intelligence, transferred responsibility, or stopped publicly reporting it.
Build a comparison table with fields for the use case name, identification number, responsible component, purpose, status, start date, vendor, risk designation, and the last year in which it appeared.
Do not trust the titles alone. Government technology has an unfortunate habit of receiving a new name whenever the old one begins attracting lawyers.
Step Three: Study the Inventory Language
Read each description as if it were written by someone who desperately wants to avoid alarming a congressional committee.
Circle vague phrases such as decision support, operational awareness, anomaly detection, identity resolution, resource optimization, sentiment analysis, risk assessment, information triage, predictive insight, and enhanced screening.
Translate each phrase into a concrete question.
What decision is being supported? Whose identity is being resolved? What behavior counts as an anomaly? What consequence follows a risk score? Which communications are being analyzed? Who becomes the object of enhanced screening?
Record what the description says and what it avoids saying. An entry may explain what the technology produces while remaining silent about the people whose information produces it.
Step Four: Identify the System Behind the Use Case
Search the inventory entry for a product name, project title, contract number, vendor, program office, system acronym, or technical description.
Then search those terms individually and in combinations. Check the agency website, USAspending, SAM.gov, congressional testimony, inspector general reports, privacy assessments, procurement notices, and federal job listings.
A use case may appear in the inventory under an internal project name while the contract identifies only the vendor platform. The vendor may use an entirely different marketing name. Record every known name in your research table.
Search for common technical terms connected to the function, including machine learning, computer vision, natural language processing, biometric matching, predictive analytics, entity resolution, automated classification, and generative artificial intelligence.
The contract may never use the phrase artificial intelligence. Apparently, the machine becomes less artificial when the procurement office calls it advanced analytics.
Step Five: Follow the Contracts and Contract Modifications
Search USAspending.gov for the agency, subagency, vendor, product, and relevant service descriptions. Search SAM.gov for solicitations, award notices, statements of work, and procurement attachments.
Do not stop after finding the original award. Review modifications, option periods, ceiling increases, related task orders, and follow on contracts. A small pilot can become a permanent system through a series of changes that receive far less attention than a new contract.
Record the award amount, potential total value, performance period, contracting office, recipient, subcontractors, and description. Look for language involving data ingestion, model development, automated analysis, identity matching, cloud services, surveillance, or decision support.
The Electronic Frontier Foundation has published a guide to tracing Homeland Security spending. Its methods can be applied to other agencies and technology purchases.
Step Six: Compare the Inventory With Budget Documents
Search the agency budget justification for the system name, vendor, program office, technology category, and purpose. Review congressional budget documents, appropriations reports, performance plans, and information technology spending summaries.
Budgets may describe capabilities that never appear clearly in the public inventory. An agency might request money for automated targeting, biometric analysis, fraud detection, or large scale data processing while its inventory describes the same work as administrative support.
Compare dates carefully. A budget request may describe a proposed system rather than an operational one. Mark the difference between requested funding, authorized funding, obligated money, and actual deployment.
If the agency has money for the system, staff assigned to the system, and a vendor publicly celebrating the system, but the inventory has never heard of it, you may have found a very interesting paperwork problem.
Step Seven: Search Privacy and Records Notices
Search for privacy impact assessments, system of records notices, data retention schedules, records management documents, information collection notices, and matching agreements.
These documents may reveal information absent from the inventory, including data sources, retention periods, information sharing partners, affected populations, access controls, and the consequences of an automated result.
Compare the documents line by line. Does the privacy assessment identify data sources that the inventory omits? Does the system notice describe information sharing with another agency or contractor? Does the inventory say the tool assists employees while another document says it prioritizes, flags, scores, or recommends cases?
Words matter. Assistance sounds gentle. A risk score that pushes someone into an investigation is not merely assisting.
Step Eight: Read the Vendor’s Own Description
Search the vendor website, press releases, conference presentations, case studies, product manuals, corporate filings, job advertisements, and public demonstrations.
Vendors often describe capabilities more enthusiastically than agencies do. The agency may call a system an administrative support tool. The vendor may boast that it predicts threats, identifies hidden relationships, analyzes social media, or processes millions of records in seconds.
Save copies of relevant pages and note the publication date. Marketing language is not proof that an agency activated every available feature. It is evidence of what the product may be capable of doing and a basis for more precise questions.
Ask which modules the agency purchased, which features it enabled, what data it connected, and whether later contract modifications expanded the system.
Step Nine: Examine Government Job Listings
Search USAJobs, archived vacancy announcements, contractor recruitment pages, and professional networking sites for the system name, program office, vendor, or technical skills associated with the project.
A job listing may reveal that an agency is hiring people to train models, label data, evaluate automated decisions, manage a platform, develop facial recognition tools, or integrate information from several databases.
Record the position title, office, duties, required skills, posting date, and employment type. Distinguish federal employees from contractors. A listing can help establish when a capability was being developed, but it does not prove that the system became operational.
Still, it is difficult to claim that a program does not exist while recruiting a senior engineer to run it.
Step Ten: Test the Risk Classification
Current federal guidance requires additional attention for artificial intelligence uses that can affect rights or safety. Examine whether the inventory identifies the system as high impact or subject to additional risk management requirements.
Then examine what the system actually does.
Does it affect access to benefits, employment, housing, health care, immigration status, law enforcement attention, personal liberty, or another important government service? Does it analyze biometric or location information? Does it influence a decision that a person may need to challenge?
If the function appears consequential but the inventory gives it a low risk classification, document the mismatch. Do not simply declare the classification false. Ask which criteria the agency applied, who approved the determination, and whether the agency completed an impact assessment.
Step Eleven: Request the Missing Records
Once you identify a specific discrepancy, submit a narrow Freedom of Information Act request. Ask for records tied to a defined system, office, vendor, and time period.
Request the agency’s internal artificial intelligence inventory entry, inventory submission records, risk classification, impact assessment, validation reports, testing results, model cards, data sheets, procurement documents, approval records, governance reviews, and communications about whether the system should appear in the public inventory.
You can also request records explaining why a use case was removed, renamed, consolidated, excluded, or designated as not publicly releasable.
Avoid requesting every artificial intelligence record created since the invention of electricity. An enormous request is easier to delay, narrow, or reject. Precision is your tiny bureaucratic crowbar.
Step Twelve: Build an Evidence Matrix
Create a table with one row for each factual claim and separate columns for the inventory, contract, budget, privacy document, vendor statement, job listing, oversight report, and agency response.
Classify every conclusion as confirmed, strongly supported, possible, contradicted, or unresolved.
A contract confirms that the government purchased something. It does not necessarily prove deployment. A job listing confirms recruitment. It does not prove operational use. A vendor announcement may establish a relationship, but vendors also enjoy making themselves sound indispensable.
The strongest finding appears when several independent records point to the same conclusion.
Step Thirteen: Give the Agency a Chance To Explain
Send the agency a short list of precise questions before publishing your findings.
Ask whether the system is operational, why it does or does not appear in the inventory, which office controls it, whether a contractor operates any portion of it, what data it uses, whether it affects rights or safety, and which reporting exception applies if it was omitted.
Give a reasonable response deadline. Preserve the questions and the response. If the agency does not answer, state that accurately.
Silence is worth reporting. It is not permission to invent the answer.
Step Fourteen: Publish the Receipts and Protect the Public
When you publish, link directly to the inventory, contracts, budget pages, privacy documents, oversight findings, and agency responses.
Explain the difference between confirmed facts and reasonable questions. Remove personal information that could expose ordinary people. Do not publish license plate numbers, private addresses, immigration details, medical information, or data that could help someone target a vulnerable person.
The goal is to expose government power, not to create a complimentary surveillance database for internet creeps.
Warning Signs Worth Investigating
A use case deserves closer examination when its description changes substantially without explanation, its start date predates its first public appearance, or it disappears while related contracts continue receiving money.
Other warning signs include missing vendor information, generic descriptions, unclear data sources, inconsistent risk labels, repeated name changes, absent testing records, and claims of human review that never explain what the human can actually change.
Look closely when a system is described as a pilot for several years, when an agency calls it inactive while recruiting employees to support it, or when a vendor describes operational capabilities that the agency presents as research.
None of these details alone proves deception. Together, they can establish a pattern that deserves formal oversight.
What Not To Claim
Do not claim that an absent use case is secret merely because you did not find it.
Do not describe a system as operational when the evidence only establishes procurement or development. Do not assume the agency activated every feature advertised by the vendor. Do not equate artificial intelligence with fully automated decision making when a system may perform a narrower analytical function.
Most importantly, do not treat the inventory as either sacred truth or useless propaganda. Treat it as one statement from an institution whose work must be tested against the rest of the public record.
Why This Investigation Matters Now
Federal artificial intelligence use is expanding rapidly. The federal budget for fiscal year 2027 states that more than 3,500 artificial intelligence use cases have been reported across the government. That scale makes accurate inventories essential.
The Government Accountability Office has repeatedly identified inventory weaknesses, including missing information and inconsistent reporting. It reported in 2026 that some agencies still had unresolved inventory problems, while others had corrected earlier deficiencies.
At the Department of Homeland Security, the public inventory includes hundreds of uses across immigration, border security, transportation security, emergency management, cybersecurity, and departmental operations. Tech Policy Press reported on the growth and surveillance implications of the DHS inventory.
An inventory is supposed to make government technology visible. It should not function as decorative transparency, where the agency publishes enough information to claim openness but not enough for the public to understand what the system actually does.
Closing
Artificial intelligence does not enter government through a single dramatic announcement. It arrives through pilot projects, contract modifications, cloud platforms, analytics subscriptions, vendor partnerships, and systems with names apparently selected by a committee trapped inside a beige conference room.
The inventory is where the agency tells us what it believes the public should know. The contracts tell us what it bought. The budget tells us what it values. The privacy documents tell us whose information is involved. The vendor tells us what the product can do. The job listings tell us what the agency is building.
Place those records beside one another.
Then ask whether the official story survives contact with its own paperwork.
Sources
- Office of Management and Budget Memorandum M 25 21
- Government Accountability Office Report on Federal Artificial Intelligence Inventories
- Government Accountability Office Report on Internal Revenue Service Artificial Intelligence
- Government Accountability Office Artificial Intelligence Accountability Framework
- Department of Homeland Security Artificial Intelligence Use Case Inventory
- Electronic Frontier Foundation Guide to the Homeland Security Spending Trail
- Tech Policy Press Report on the DHS Artificial Intelligence Inventory
