AI wage thief detector exposes $50B in stolen wages for vulnerable workers

A new AI-powered tool is giving millions of workers a fighting chance against wage theft, a billion-dollar problem that quietly strips $50 billion from paychecks in the U.S. every year. Named PaySnap, the application doesn’t just detect violations—it explains them in plain language, cites the exact labor laws broken, and guides users on how to report the theft—all in their native tongue.

PaySnap turns an opaque system into a transparent one. Workers who once struggled to understand confusing pay stubs or feared retaliation can now upload a photo of their pay stub or describe their situation in Hindi, Spanish, or Chinese. In seconds, the AI calculates what they’re owed, identifies the violated law, and provides the Department of Labor’s confidential hotline.

From raw data to real-world impact

The heart of PaySnap is a fine-tuned version of Google’s Gemma 4 E2B model, trained on 365,393 real enforcement cases from the U.S. Department of Labor. Unlike generic AI assistants, this system doesn’t just guess—it applies real-world labor law patterns to detect violations such as unpaid overtime, illegal deductions, and minimum wage breaches.

The team behind PaySnap chose Gemma 4 for three key reasons:

• Edge efficiency — The model runs smoothly on modest hardware, delivering 63 tokens per second on an Apple M3 Pro and fitting within 3.4GB RAM as a Q4_K_M GGUF. Large models simply wouldn’t run on older devices common among workers in high-risk industries.
• Cost-effective training — Using Unsloth LoRA on a free Kaggle T4 GPU, the team fine-tuned the model to a training loss of 0.009. Training a 31-billion-parameter model would have been impossible on limited compute.
• Multilingual reach — Despite its compact size, the model generates coherent responses in 11 languages, a critical feature for reaching non-English speakers who are most vulnerable to exploitation.

How workers reclaim lost wages with AI

PaySnap operates through a simple two-step process. First, a worker uploads a photo of their pay stub or describes their hours, rate, and deductions. The AI then analyzes the input using Gemma 4’s vision and function-calling capabilities.

For example:

• A construction worker in Texas who worked 52 hours at $15/hour with no overtime listed receives an alert: 12 unpaid overtime hours under the Fair Labor Standards Act (FLSA) Section 207(a)(1), totaling $90 in lost wages.
• A restaurant worker in New York earning $16/hour with $35 uniform and $50 breakage deductions discovers both overtime violations and illegal deductions. The AI calculates $149 owed and delivers the explanation in Hindi.

The system doesn’t stop at detection. It always provides the Department of Labor’s toll-free hotline—1-866-487-9243—which is free and confidential—and walks users through the reporting process step by step.

Performance and future directions

In independent evaluations using LLM-as-Judge benchmarks with base Gemma 4 E2B as the adjudicator, the fine-tuned PaySnap model improved accuracy by 11.7%, rising from 8.12/10 to 9.07/10. All five evaluation dimensions saw gains:

• Legal accuracy: +1.73
• Statute quality: +1.33
• Actionability: +0.73
• Dollar accuracy: +0.67
• Worker clarity: +0.27

The open-source releases—including a GGUF model, LoRA weights, training notebook, and the full dataset—allow others to build on this work. The team emphasizes that while technology can empower workers, systemic change requires policy support and employer accountability.

As PaySnap prepares for wider deployment, its creators see potential in expanding language support, integrating with payroll systems, and partnering with labor rights organizations. The tool isn’t just about catching thieves—it’s about restoring dignity to millions of workers who have long been invisible in the shadows of the labor system.