TL;DR
The Pandemic Response Accountability Committee (PRAC) trained a fraud-detection model on roughly five million Small Business Administration COVID loan applications and says it could have flagged potentially tens of billions of dollars in questionable payments before disbursement. PRAC is now deploying the tool across oversight work funded by the recent reconciliation law and has had its mandate extended through 2034.
What happened
PRAC developed a proof-of-concept Fraud Prevention Engine using about five million SBA Economic Injury Disaster Loan applications collected after pandemic programs were already operating. The system combines unsupervised anomaly detection, supervised machine learning trained on known pandemic fraud patterns, and rules-based checks such as invalid Social Security or employer identification numbers. PRAC told the House Oversight and Government Reform Committee that the engine can process roughly 20,000 applications per second and that, had it been in place in March 2020, it likely would have flagged potentially tens of billions of dollars in payments for further review before funds were sent. To date, PRAC reports it has helped recover just over $500 million. Following last year’s budget reconciliation, PRAC’s remit was extended through September 30, 2034, with $88 million to support expanded oversight; the committee says it is working with agency Inspectors General to deploy the engine across programs funded under that law.
Why it matters
- Early automated screening could reduce the need for costly post‑payment recovery by identifying suspicious claims before funds are disbursed.
- A validated fraud‑detection model trained on pandemic data can be repurposed for oversight of other large federal spending programs.
- PRAC’s expanded mandate and funding create a path to scale analytics across agencies that manage reconciliation law programs.
- The existence of a large, searchable dataset of applications increases the government’s ability to spot connections and systemic fraud patterns.
Key facts
- PRAC trained the Fraud Prevention Engine on roughly five million SBA COVID EIDL applications.
- The engine uses a mix of unsupervised anomaly detection, supervised machine learning, and rules-based checks.
- PRAC estimates the model could have flagged potentially tens of billions of dollars in suspect payments if used in March 2020.
- The system is reported to process about 20,000 applications per second.
- PRAC says it has helped recover just over $500 million so far from pandemic relief programs.
- Last summer’s reconciliation bill extended PRAC’s mandate through September 30, 2034, and provided $88 million for expanded oversight.
- PRAC has begun working with agency Inspectors General to apply the Fraud Prevention Engine to programs funded under the reconciliation law.
- PRAC developed the engine as a proof of concept to test feasibility and resolve technical hurdles.
What to watch next
- Whether a permanent institutional home is found to preserve the engine and its database beyond PRAC’s extended mandate — not confirmed in the source.
- The scope and pace of deployment across federal agencies and programs, where PRAC says it has begun work with Inspectors General.
- How much additional recovery or prevention the engine will produce once applied to reconciliation-law programs — not confirmed in the source.
Quick glossary
- Pandemic Response Accountability Committee (PRAC): A federal committee set up to oversee pandemic-related spending and recovery efforts; it conducts audits and investigations to detect waste, fraud and abuse.
- Economic Injury Disaster Loan (EIDL): A Small Business Administration loan program used during the COVID-19 pandemic to provide financial assistance to businesses suffering economic injury.
- Unsupervised machine learning: Algorithms that identify patterns or anomalies in data without labeled examples of correct outputs.
- Supervised machine learning: Algorithms trained on labeled examples to recognize patterns associated with known outcomes, such as confirmed fraud cases.
- Rules-based checks: Deterministic validation steps that flag outright inconsistencies or invalid identifiers, such as incorrect Social Security or employer ID numbers.
Reader FAQ
Was the Fraud Prevention Engine in use during the initial COVID relief payouts?
No. PRAC trained the model on application data that became available after the programs were already operating.
How much potentially fraudulent payment could the model have flagged?
PRAC told Congress it could have flagged potentially tens of billions of dollars for further review.
How much money has PRAC recovered so far from pandemic relief programs?
PRAC reports just over $500 million recovered to date.
Will the engine be preserved permanently within the federal government?
Not confirmed in the source.
PUBLIC SECTOR There was so much fraud on COVID loans, the feds trained an anti-fraud AI on the applications Had it been around in 2020, it could have flagged tens…
Sources
- There was so much fraud on COVID loans, the feds trained an anti-fraud AI on the applications
- Pandemic watchdog builds AI fraud prevention 'engine' …
- Feds pitch anti-fraud AI trained on COVID loan data
- AI helped the feds catch $1 billion of fraud in one year. And …
Related posts
- Datacenter expansion forecasts may be overstated due to power shortfalls
- US pauses visa processing for applicants from 75 countries over public charge
- Facing an Age Gate Online? How to Decide and Protect Your Data