Bots are standing in the financial aid line

An instructor opens her roster in week two. Forty names, twelve of whom have never posted, never opened a file, never answered an email. She has done this every term for three years and she knows exactly what she is looking at. Nineteen real students are on the waitlist for a class that is officially full.

Picture the rest of that week honestly, because the accounting matters.

She sends the twelve a message. Nothing comes back. She sends a second one with a deadline, because the college requires documented outreach before anyone can be dropped for non participation. Still nothing. She fills in a form for each of the twelve, which takes a few minutes each, and submits them to a coordinator who is doing the same triage across a few hundred sections. Somewhere in that process the census date passes, which is the point at which enrollment is certified and aid is disbursed.

By the time the twelve are removed, the money is gone. Tuition was paid to the college from the aid package, and the balance, the part intended to cover textbooks and rent and bus fares, was refunded to an account the enrollee nominated. That refund is the point of the entire exercise. Nobody involved ever intended to take a class.

Meanwhile the nineteen students on the waitlist made other plans. Some picked a different course. Some deferred a term. Some, statistically, will not come back at all, because the research on community college attrition is consistent that interrupted momentum is where people are lost.

That is the shape of ghost student fraud. It is usually reported as a money story. The money is the smaller half.

What is a ghost student?

A ghost student is an enrollment that exists to collect financial aid and nothing else. No one attends. Often no one exists.

The mechanics are unglamorous, which is precisely why they work at volume. Open enrollment institutions, community colleges above all, are designed to admit anyone. That is not a flaw, it is their statutory purpose and the reason they matter. Online sections mean nobody has to appear in a room. Financial aid pays tuition to the institution and refunds the remainder to the student, because students have living costs. Put those three facts together and you have an arbitrage: an application costs a few minutes, and a successful one produces a refund measured in thousands.

The operation does not need to be sophisticated. It needs to be numerous. A ring generates applications using stolen or synthetic identities, enrolls each one in online sections, keeps them just active enough to survive the first automated checks, waits for the census date, and collects. If ninety percent get caught, the remaining ten percent still pay for the whole enterprise.

Why the census date is the whole game

Every college has a date on which enrollment is frozen for reporting and funding purposes. Before it, you can add and drop with no financial consequence. After it, the enrollment is real, the institution certifies it, and disbursement follows.

Fraud rings understand this calendar better than most faculty do. All the effort goes into surviving until census. Everything after that is collection. And the college's own detection mechanism, an instructor noticing that a student has never participated, structurally cannot conclude before the deadline it needs to beat, because you cannot prove someone is absent until enough time has passed for their absence to be meaningful.

It is a race between a manual human process and a calendar, and the calendar has never once been late.

How much is this actually costing?

Enough that it has moved from a campus operations irritation to a federal policy question in about two years.

The California Community Colleges Chancellor's Office, which runs the largest system in the country and has been unusually transparent about the problem, tracked roughly ten million dollars in federal financial aid fraud and three million in state aid fraud between March 2024 and March 2025. Reporting during the 2025 and 2026 cycles described several California systems flagging close to a third of applicants as fraudulent or bot generated, which is a number worth pausing on, because it means the majority use case for the admissions pipeline had become filtering rather than admitting.

The pattern is not confined to one state. The Community College of Philadelphia reported more than six hundred fraudulent applications in 2025, about five percent of all applicants. Equifax has written about national exposure in the range of one hundred and eighty million dollars. The US Department of Education said that its identity verification requirement for first time aid applicants, introduced during 2025, prevented on the order of one billion dollars in fraudulent aid by the end of that year.

Treat the individual figures with appropriate care. They come from different methodologies, some from chancellor's office reporting, some from institutional disclosure, some from vendor analysis, and the definition of a flagged application varies. The direction and the order of magnitude are consistent across all of them, and the federal response is the strongest evidence that the problem is real: governments do not add verification steps to student aid for fun, because every step costs them enrollment.

The billion dollar figure cuts both ways

It is worth reading the Department of Education's prevention number carefully, because it is genuinely good news and also the reason this post exists.

A billion dollars prevented means the document check at application worked. It caught a great deal. But a control that operates at application is, by construction, operating before the seat is consumed, before the census date, and before the refund account is nominated. It filters the front door. The money leaves through a different one.

What you would expect to see after a strong application time control is fraud migrating rather than stopping: toward returning applicants who are already past the front door check, toward identities good enough to survive document proofing, and toward the disbursement step. That is the migration institutions should be planning for now.

Why is the seat worth more than the money?

Because the money is recoverable and the seat is not.

When aid is disbursed fraudulently, there is a machinery for clawback. It is slow and painful and the institution eats administrative cost, but the dollars are, in principle, traceable. Nobody has ever recovered a semester.

A community college section has a hard capacity. When ghost enrollments fill it, real applicants see a full class. The consequences are not evenly distributed, which is the part that should bother people most. The students most affected are the ones with the least flexibility: people working shifts who can only take the Tuesday evening section, parents whose childcare fits one specific slot, students who need one particular prerequisite to transfer on schedule. A student with options takes a different class. A student without options loses a term.

Then add the instructional cost. Faculty spend the opening weeks of term doing identity triage instead of teaching, which is both a waste of expensive expertise and demoralising in a specific way: it turns the beginning of a course, the part where you build a class, into an exercise in deciding who is not real.

So the honest ledger for a mid sized system in one term reads something like this. Some quantity of aid disbursed and partly recovered. Several hundred faculty hours of manual triage. An unknown but non trivial number of real students displaced, some of whom do not return. Only the first line item appears in any report.

Why does an identity check at application not stop this?

Four structural reasons, and they compound.

Document proofing is exactly what synthetic identities are built for

A synthetic identity is a constructed persona assembled from real fragments, often a valid identifier belonging to someone with no credit history, combined with plausible supporting detail. The entire purpose of building one is to pass a documentary check. Asking a fraud ring to present a document is asking it to do the one thing it has already optimised for. Generative tools have made the supporting material, essays, correspondence, plausible document images, effectively free since 2024.

The check happens before anything is at stake

This is the deepest problem and it is a timing problem, not a strength problem. Application time verification asks a question at the moment of lowest consequence. Nothing has been consumed. No money has moved. The seat is still theoretical. Then the system assumes the answer holds for everything that follows, across months, through census, through disbursement, through a refund destination that can be changed later.

We call this the enrollment binding gap: a verified human is checked once and then presumed present for a duration nobody re-examines. It appears in this workflow alongside identity discontinuity, the same failure that lets a person interview for a job and a different person do it. Both are catalogued in the Identity Failure Map.

Traffic scoring loses to labour arbitrage

CAPTCHA and bot scoring evaluate a request and return a probability. They are defeated by the simple economics of a human solving farm, where the cost per solved challenge is cents and the payoff per successful enrollment is thousands. Any control whose bypass cost is under a dollar cannot defend a prize measured in thousands. That is not an implementation failing, it is arithmetic.

Every false positive is a real person locked out of an education

Risk scoring returns probabilities, and probabilities generate false positives. In fraud contexts we usually treat a false positive as a minor inconvenience: a declined card, a re-verification, mild annoyance.

Here the false positive is a person who wanted to go to college and was told no by a model. The students most likely to trip a risk score are the ones with the thinnest data trails: recent immigrants, people without established credit, those using a library computer or a shared phone, people who have moved frequently. That is a description of exactly the population open enrollment institutions exist to serve.

Any system that stops ghost students by making it statistically harder for poor students to enroll has not solved the problem. It has renamed it.

Where does the money actually move?

Three moments. Everything else is paperwork.

Moment one, the seat. Enrollment in a section consumes finite capacity. This is where displacement happens, and it happens before any money moves.

Moment two, census confirmation. The institution certifies that this enrollment is real, which triggers disbursement. This is the single highest leverage point in the entire process and at most institutions it is either fully automatic or based on an instructor's judgment made under time pressure.

Moment three, the refund destination. The aid balance goes to an account the student nominates, and that account can typically be added or changed through a self service portal. This is the step where the fraud is finally monetised.

That third one should look familiar to anyone in payments. It is a bank account change on a self service portal, authorised by nothing more than a logged in session. It is structurally identical to payroll diversion, where an attacker with a session changes an employee's direct deposit details and waits one pay cycle. Same payload, same weakness, different building on campus. Investigators have repeatedly found single refund accounts shared across dozens of enrollees, which is the signature of this step being unprotected.

What would actually work?

Move the check from the moment of least consequence to the moments of greatest consequence, and change what is being checked from a document to a live human.

The principle is one human, one enrollment, proved again when it counts. At application the person binds to a key on their device, with an on device face match that produces a one way key and stores no biometric template anywhere. At census confirmation and at refund account setup, they sign the specific thing being asserted.

What the census confirmation looks like

The payload is small and completely uninteresting, which is the point. It says one human confirmed one enrollment at one time.

{
  "action":      "census_confirm",
  "term":        "2027SP",
  "student_key": "z6MkuT9pQ2vXhL8sD4nR7bW1cY3fA6gK",
  "sections":    ["ENGL-101-4471", "MATH-120-3318"],
  "census_date": "2027-02-08",
  "institution": "did:web:example-college.edu",
  "signed_at":   "2027-02-06T15:41:09Z"
}

And the refund destination, which is the one that stops the monetisation:

{
  "action":       "refund_account_set",
  "student_key":  "z6MkuT9pQ2vXhL8sD4nR7bW1cY3fA6gK",
  "routing":      "021000021",
  "account_last4":"8842",
  "effective":    "2027-02-10",
  "signed_at":    "2027-02-06T15:43:55Z"
}

The verification is local and takes microseconds:

receipt = student.sign(canonical(payload))   # device key, live gesture

# bursar side, before any disbursement
assert verify_ed25519(receipt.sig, canonical(payload), student.pubkey)
assert receipt.freshness < 300          # signed minutes ago, not months
assert receipt.presence == "live"       # liveness challenge passed
assert unique_human(student.pubkey, term="2027SP")

disburse()

Why this inverts the economics

Here is the part that matters more than any cryptographic detail.

For a real student, this costs about twenty seconds, twice a term, on a phone. It requires no document, no upload, no visit to an office, no credit history, no permanent address.

For a ring running four thousand fake enrollments, it requires four thousand distinct live humans to appear on four thousand distinct devices, twice, on the institution's schedule rather than their own. The cost of fraud stops being a per application cost of a few minutes of automation and becomes a per enrollment cost of recruiting and coordinating a real person. That is the difference between a business and a hobby.

Notice that this does not require the college to be better at spotting fakes. It requires the college to ask for something a fake cannot produce at scale. Those are very different engineering problems, and only one of them can be won.

The honest gap in this design

One human enrolling at twelve colleges simultaneously is not addressed by any of the above, because each institution verifies independently and none of them can see the others without building exactly the kind of central registry that should make everyone uncomfortable.

Solving that properly needs a nullifier: a construction that lets a person prove they have not already claimed a benefit in a given period, without revealing who they are or letting institutions correlate them. That is real cryptography, it is well understood in the literature, and it is not shipped here. It belongs on the roadmap and in a standards conversation, not in a sales deck. We would rather say that plainly than imply coverage that does not exist.

But what about students who cannot comply?

This objection is correct, it is the most important one in the entire discussion, and any proposal that waves it away deserves to fail.

Community colleges serve students who are unhoused, students whose immigration status makes them wary of anything resembling government identity infrastructure, students sharing a single phone with a family, students on library computers, students with disabilities that make certain interactions difficult, and students who have every reason from experience to distrust institutions asking them to prove who they are.

A verification requirement that fails those students to stop fraud has not improved the institution. It has just changed which humans get excluded, and it has done so in the direction of the people the institution exists for. Say it as directly as possible: a system that locks out real students in order to stop fake ones has failed, regardless of what happened to the fraud number.

So the design constraints are not negotiable extras. They are the specification.

• No government identity document, ever. If the flow needs a passport or a licence, it has excluded a meaningful share of the intended population at step one.
• Works on a borrowed or shared device. The proof binds to a person, not to a handset. Enrolling on a friend's phone and confirming later on a library machine has to work.
• Nothing biometric is retained. The face match happens on the device and produces a one way key. A public college holding face templates creates state law exposure and, more importantly, is the kind of thing that makes a cautious student decide not to enroll at all.
• A staffed human fallback that is genuinely equivalent. Not a worse path, not a longer queue with a worse outcome. A person in the financial aid office can vouch, and that vouching is itself a signed act by a named staff member, which makes it auditable rather than invisible.
• Failure is never silent. A student who cannot complete a step sees what happened and what to do next. Silent denial is how people disappear from higher education.
• Measure exclusion as rigorously as fraud. If you report prevented fraud without reporting completion rates by demographic, you are reporting half a result and the missing half is the one with a human cost.

How do the available interventions compare?

The two columns that matter are on the right, and most procurement conversations only look at the left.

The bottom row is not better because the cryptography is clever. It is better because it acts at the moment the loss occurs and asks for something that is cheap for one real person and expensive per fake one. Every row above it either acts too early, acts too late, or asks for something a well funded ring finds easier to supply than a struggling student does.

What this does not fix

Five limits, plainly.

• Eligibility fraud by real people is untouched. A genuine human who misrepresents income or dependency status to receive more aid than they qualify for will sign every proof honestly, because they are present and they are themselves. This control addresses non existent students, not dishonest ones.
• Cross institution rings remain partly open until nullifiers exist, as described above.
• Insider fraud is a different problem. Staff with system access who create or certify enrollments require separation of duties and audit, and the receipt helps by making certification an attributable signed act rather than a database row, but it is not the primary control.
• A signature does not prove learning. It proves a specific human was present and asserted something. A person who confirms enrollment and then never engages has still consumed a seat, and that is an academic policy question, not an identity one.
• Adoption timelines are institutional. Student information systems are procured on multi year cycles. Anyone promising this by next term has not met a registrar.

What to do this week

1. Find out whether your refund account change requires anything beyond a session. Log into your own student portal and try to change the destination. If nothing stops you, that is the monetisation step and it is currently open.
2. Query for shared refund destinations. Group disbursements by account number for the last three terms and look at the tail. Repeated accounts across unrelated enrollees is the clearest single signal available to you today, and it costs one query.
3. Count the faculty hours. Ask a sample of instructors how long the first three weeks of roster triage takes them, multiply by sections, and put a number on it. That number funds the fix and it never appears in the fraud report.
4. Measure displacement. For every section that filled and had a waitlist, count enrollments dropped for non participation after census. That difference is real students who were turned away for seats nobody used.
5. Map your timeline against the calendar. Write down the date each control fires and the census date. Any control that fires after census is documentation, not prevention.
6. Ask your student information system vendor what they support at census confirmation and refund setup, not at application. The answer is usually nothing, and asking the question in a procurement cycle is how that changes.
7. Report exclusion alongside prevention. Whatever you deploy, publish completion rates by student population next to the fraud figure. If you only report one, you will optimise the wrong one.
8. See a presence proof from the student's side. Our bot application demo shows the same primitive applied to a different flood of fake applicants, and the docs show what the verification call looks like in practice.

Frequently asked questions

How do colleges stop ghost student financial aid fraud? By requiring a fresh signature from a live, unique human at the two moments money moves, census confirmation and refund account setup, rather than relying on a document check at application. Application time checks happen before any seat or money is at stake, and they are precisely what synthetic identities are constructed to pass.

What is a ghost student? An enrollment created solely to collect financial aid. A ring applies using stolen or synthetic identities, enrolls in online sections, stays active enough to survive the census date, then collects the aid refund that remains after tuition. Nobody attends, and often nobody exists. The seat is consumed either way.

Why are community college classes full of enrollments that never appear? Open enrollment, online sections and aid refunds combine into an arbitrage where an application costs minutes and a successful one returns thousands. Institutions designed to admit everyone are the natural target, and the same openness that makes them valuable makes them exploitable at volume.

Does the Department of Education verify student identity? Yes. An identity verification requirement for first time aid applicants was introduced during 2025, and the Department reported it prevented on the order of one billion dollars in fraudulent aid by the end of that year. It acts at application, which means it filters the front door but does not cover census confirmation or the refund destination.

Will identity checks lock out low income students? They can, and that is the central design risk. Document based checks exclude students without passports, licences or credit histories, and risk scores misfire on shared devices and public computers. A workable design must require no government document, function on a borrowed phone, retain nothing biometric, and provide a properly staffed human fallback.

Why not just drop non participating students before census? Because absence takes time to establish and the census date does not move. Confirming that someone has genuinely not participated requires enough elapsed time to be meaningful, which is usually longer than the window available, and the process punishes real students who had a difficult first week.

Can one person enroll at many colleges at once? Yes, and no single institution can see it, because each verifies independently. Detecting it without building a central registry that tracks students across institutions requires nullifier cryptography, which lets someone prove they have not already claimed a benefit without revealing who they are. That work is not deployed today and should be argued for openly.

Sources

1. California Community Colleges Chancellor's Office, reporting on financial aid fraud and application screening: cccco.edu
2. US Department of Education, Federal Student Aid, identity verification requirements for aid applicants: studentaid.gov
3. Federal Student Aid partner guidance and programme integrity announcements: fsapartners.ed.gov
4. US Department of Education, Office of Inspector General, audits of student aid programme integrity: oig.ed.gov
5. Information Technology and Innovation Foundation, March 2026 analysis of ghost student fraud and digital identity in higher education: itif.org
6. Equifax, business insights on ghost student enrollment fraud: equifax.com
7. National Institute of Standards and Technology, SP 800-63 Digital Identity Guidelines, identity assurance levels: nist.gov

The stolen money can be clawed back. The seat cannot. Nobody has ever recovered a semester.