What happened

The updated model’s image generator can produce receipts with realistic text, wrinkles, stains, and background details, making them nearly indistinguishable from genuine ones. 

Examples include fabricated bills from real restaurants like San Francisco’s Epic Steakhouse, complete with itemised purchases and totals. While minor errors (eg. comma-decimal confusion, arithmetic mistakes) persist, these can be easily corrected with basic editing tools.

The risks span fraudulent expense reimbursements, bogus tax deductions, fabricated crowdfunding claims, and forged financial records. Systems relying on receipts for verification - such as workplace reimbursement platforms, gig economy apps, and online marketplaces - are considered particularly vulnerable.

Why it happened

The technical leap stems from ChatGPT-4o’s improved ability to render text within images, a feature intended for creative applications like educational content or advertising. 

However, users quickly exploited this capability for malicious purposes, sharing prompt templates and refining outputs to enhance authenticity.