AI Reads Your Receipts Well — Until the Scan Gets Blurry
Blurry scans still trip up AI, and that could cost you money.
Every time a company scans your receipt, insurance form, or invoice, software has to turn that picture into typed text — a step called OCR, or optical character recognition. Then an AI reads that text and picks out the important bits: the date, the total, the account number. Two researchers at the University of Texas at Arlington tested how well five popular free AI models handle that second job.
The good news: when the text was clean, these models were excellent. They pulled the right details almost as reliably as expensive, specially built systems. That matters because it means ordinary businesses — a small accounting firm, a doctor's office, a landlord — can automate paperwork without paying enterprise prices.
The bad news: real documents are messy. They're photographed at an angle, photocopied three times, or creased down the middle. The researchers ran the same documents through three common scanning tools and watched accuracy fall off a cliff. Worse, making the AI model bigger — the usual fix — barely helped. Once the scan is bad, no amount of AI cleverness rescues it.
The study also catalogued how the AI fails: it mixes up which label goes with which value, invents details that aren't there at all, and garbles numbers — turning $1,250 into $125. That last one is the scary one. If an AI quietly corrupts a number on an insurance claim or a tax form, a human may never catch it.
The takeaway for anyone deploying this technology: fix the scanner before you upgrade the AI. And for the rest of us, keep an eye on any automated system handling your money — a second pair of human eyes is still cheap insurance.
- Five free AI models (Gemma, Mistral, Qwen2.5, LLaMA 3, DeepSeek) were tested on extracting details like totals and dates from receipts and forms.
- With clean typed text they performed near professional grade, but accuracy dropped sharply on real-world messy scans from PaddleOCR, EasyOCR, and Tesseract.
- A common failure was number corruption — turning $1,250 into $125 — which could quietly cause billing, tax, or insurance errors.
Why It Matters
Automated paperwork is spreading fast; knowing AI misreads blurry scans means you should double-check bills and claims.