The Invisible AI Map
Imagine a first-year student at a community college in upstate New York. Halfway through the semester an adviser gets in touch: the college thinks she might need some support. She never asks why she was picked. Most students wouldn’t. Behind that message sits software that ranks students by how likely they are to drop out. Her college told New York’s state auditors that the software was not AI. The auditors, having read the developer’s own documentation, said it was.
That disagreement is the heart of a state audit of the State University of New York, published by the New York State Comptroller in August 2026. SUNY is the largest public university system in the United States, with 64 campuses, and those campuses already use AI to transcribe doctors’ notes, read license plates in parking lots, flag students at risk of dropping out and blur faces in police video (Office of the New York State Comptroller [OSC], 2026, p. 2).
Our proposition: the most important finding is not that SUNY lacked rules for AI. It is that campuses and auditors could not agree on what AI is, and whoever draws that line decides what gets checked.
Who gets to decide what counts as AI?
The auditors’ headline finding is blunt. SUNY’s central administration had no effective framework for governing AI, no standard definition of AI and no documented policies or procedures for building or using it. Its general IT rules required campuses to keep an inventory of their systems, but nothing in them required anyone to mark which of those systems used AI (OSC, 2026, pp. 2, 11). The audit covered January 2019 to October 2025 (OSC, 2026, p. 23).
The auditors then looked closely at four campuses that use AI: the University at Albany, Stony Brook University, Upstate Medical University and Onondaga Community College. They found governance that varied enormously from campus to campus, and they traced much of the problem to two causes: too little guidance from the center, and poor identification of what was AI in the first place (OSC, 2026, p. 3).
The college that said its early-warning tool wasn’t AI
Onondaga Community College has used its early-alert platform since February 2020. According to the auditors, the software uses machine learning on the college’s own historical student data, such as grades and graduation records, to build a model that predicts which students will stay. Those predictions rank students by how much help they are likely to need (OSC, 2026, p. 13).
The college told the auditors it did not believe the tool used AI. In its formal response it went further, writing in capital letters that its staff do not use the predictive features the report describes (OSC, 2026, p. 31). The Comptroller’s office was unmoved. The developer’s own documentation, it replied, shows the predictions come from machine learning, and as long as those predictions are available to the college, the risks exist and need managing (OSC, 2026, p. 32).
There is a telling detail. The developer had produced a report on how accurate the model was. The college had never received it, and it tracked how many students were contacted, not whether the model was picking the right ones (OSC, 2026, p. 19). In fairness, Onondaga was also the only one of the four campuses with a written rule that a person must review AI-generated output before it is used (OSC, 2026, p. 18). The college has said it will rebuild its procedures from scratch rather than patch them (OSC, 2026, p. 31).
License plates, body cameras and patient notes: the same blind spot
Onondaga was not alone. At Albany, a license plate recognition system launched in June 2025 uses a neural network to read plates and recommend when parking enforcement should act. Campus officials told the auditors they do not classify license plate readers as AI, and that the system is a poor fit for AI governance (OSC, 2026, p. 13).
At Stony Brook, the university police use body-camera software with built-in AI features, from transcription and report writing to detecting people in footage and redacting video. Officials said they used only the redaction feature, though some users have access to the rest, and the tool was left out of the campus AI inventory (OSC, 2026, p. 13).
At Upstate Medical University, clinicians have dictated patient notes into the electronic health record since December 2019 using speech recognition software that learns each user’s way of speaking (OSC, 2026, p. 13). The vendor contract lets the company and its partners use Upstate’s data to train and improve the software, improvements that may also benefit other customers (OSC, 2026, p. 21). Upstate had started building a proper review process and hired an AI Governance Lead, but its AI policy had not been approved during the audit, and its older AI systems had never been through the new review (OSC, 2026, pp. 15, 17).
No list, no tests, no checks on accuracy
Strip away the details and the pattern is simple. None of the four campuses required or carried out specific tests of whether their AI systems’ outputs were accurate or biased (OSC, 2026, p. 3). None had run periodic audits of their AI systems, and only Stony Brook’s medical arm required such reviews (OSC, 2026, p. 19). Three of the four, Albany, Onondaga and Upstate, kept no formal inventory of their AI systems. Stony Brook did keep one, but it left out the very tool the auditors examined (OSC, 2026, p. 18).
The auditors made three recommendations: put policies in place for effective AI governance, provide guidance and oversight for AI systems, and have SUNY’s central administration coordinate with campuses on the rules that apply (OSC, 2026, p. 22).
Who benefits from a blurry line
In our reading, a fuzzy definition of AI suits almost everyone except the people the tools are used on. A campus avoids a review it never knew it owed. A vendor can ship predictive features switched on and let the customer decide whether to call them AI. The student ranked by an early-alert system and the patient whose notes were dictated through learning software carry the risk, usually without knowing any tool was involved, and without a list anyone could check.
SUNY’s answer: the rules exist now
SUNY’s reply deserves a fair hearing. In its formal response of May 2026, it pointed to a new Systemwide Artificial Intelligence Policy Framework, adopted by its Board of Trustees, which sets one definition of AI for the whole system and requires every campus to publish its own AI policy, or update existing ones, by 31 December 2026 (OSC, 2026, p. 25). It says the trustees’ resolution also commits SUNY to issue the procedures, guidance and templates campuses need to put that policy into practice (OSC, 2026, p. 27).
Some campuses moved quickly. Upstate told the auditors it agreed with the findings and that its AI policies had since been formally approved. They require human review before any AI system is deployed, a monitoring plan for every AI system, old and new, and an institutional AI inventory (OSC, 2026, pp. 35–36). Onondaga said it would apply the systemwide policy, including its definition of AI (OSC, 2026, p. 29).
We read the policy itself, SUNY Document 6904, which took effect on 30 April 2026. Its definition of an AI system is deliberately wide, written to cover academic, research and administrative uses alike (State University of New York [SUNY], 2026, p. 3). It asks for oversight in proportion to risk, and says systems that can control or meaningfully shape outcomes for people’s safety, rights or wellbeing deserve closer scrutiny, including bias reviews and risk assessments (pp. 2–4). A definition that wide makes it hard for any campus to wave away a tool with predictive features built in. That answers the Onondaga problem on paper.
There is one wrinkle. The policy’s own list of the institutions it covers names state-operated campuses, health science centers, university hospitals, and statutory and contract colleges, but not community colleges (SUNY, 2026, p. 1). That fits what the auditors were told: SUNY’s authority over community colleges is limited to general supervision and approving academic programs (OSC, 2026, p. 2). Onondaga has said it will apply the systemwide policy anyway (OSC, 2026, p. 29). Whether every community college does the same is up to them.
Our verdict: SUNY now knows what AI is, but not where it is
SUNY has answered the auditors’ first complaint. It now has one definition of AI for the whole system, and a broad one (SUNY, 2026, p. 3).
The second problem is still open. Before anyone can check an AI system, someone has to know it is there. The audit found that three of the four campuses it examined kept no list of their AI systems, and the fourth, Stony Brook, kept one that left out the very tool the auditors chose to look at (OSC, 2026, p. 18). The new policy does not ask campuses to keep such a list. It sets out what each campus’s own AI policy must cover, from who is responsible to how AI is bought, trained for and reviewed, and it says SUNY will run regular surveys of how AI is used across the system (SUNY, 2026, pp. 3–4). A survey records what people remember to report. A list is a record that someone is responsible for keeping complete. Upstate has promised to keep one (OSC, 2026, p. 36). Nothing in the systemwide policy requires the other campuses to do the same.
We have seen this gap before. At the US Department of Agriculture, the inspector general found that the department’s list of AI relied on staff reporting their own tools and could not be confirmed as complete, while 73 of its 82 AI systems in use had never received the required security sign-off, as we reported in 73 of 82. In Queensland, the state auditor found two layers of human review on the state’s AI road cameras but no central record of where AI was used across the transport department, as we reported in 208 Million Checks, 114,000 Fines. Careful work at the edge, no list at the center.
What the audit does not tell us
The audit is about governance, not harm. It describes no student or patient who was hurt by any of these systems, and it does not say the tools made wrong decisions; it says nobody was checking whether they did. It sampled four of SUNY’s 64 campuses, so it cannot tell us how the rest compare. And the real test, what campuses publish by 31 December, is still ahead.
Questions you can ask
If you are a SUNY student or parent, ask whether your campus uses an early-alert or retention tool, what data it uses, and whether a person reviews every flag before anyone acts on it.
Has your campus told you about an AI tool? Tell us. Whether you are a student, a parent or a member of staff, if you have been told about an AI system on your campus, or found one nobody mentioned, write to us at info@fairbydesign.org. We read every message, and what you tell us helps decide what we investigate next. To hear what we find, get the free Mercury Brief.
If you are a patient at a SUNY hospital, or at any hospital, it is worth knowing whether software helped write your record. Upstate’s doctors have dictated notes through speech recognition since 2019 (OSC, 2026, p. 13).
Questions to ask about AI in your care
- Was speech recognition or any other AI tool used to write or summarize my notes?
- Who checked what the tool produced before it went into my record?
- Can I have a copy of my notes, so I can check them myself?
- Does the company that makes the tool get to use my information to improve its product?
For a case where an AI tool sat inside decisions about patients’ care, see our page for Episode 1: Is your AI really just a suggestion?
If you sit on a campus committee, ask to see the list of AI systems in use, who owns each one, and which are treated as higher risk. If there is no list, put it on the agenda before the December deadline. The first section of our free readiness assessment, Inventory and scope, walks you through writing a one-page record for each AI system: what it is for, who owns it and what it depends on.
The new question will be answered in public within weeks. When SUNY’s campuses publish their policies at the end of December, will any of them include an actual list of the AI they run? We will check the policies after the 31 December deadline and report back in January. Get the free Mercury Brief to hear what we find.
Is any of your software AI without you knowing it? If you run AI anywhere, from a campus office to a small business, the free readiness assessment starts where this audit starts: knowing what you run and who owns it. Try it on the awkward cases, such as vendor tools with AI features you say you don’t use. Take the free readiness assessment.
References
Office of the New York State Comptroller. (2026, August). State University of New York: Artificial intelligence governance (Report 2024-S-33). Division of State Government Accountability. https://www.osc.ny.gov/files/state-agencies/audits/pdf/sga-2026-24s33.pdf
State University of New York. (2026, April 30). SUNY systemwide artificial intelligence policy (Document No. 6904). https://www.suny.edu/SUNYPP/pdf.cfm?doc_id=933


Leave a Reply
You must be logged in to post a comment.