In March, ChatGPT introduced an enhanced image generator as part of its 4o model that significantly improves the generation of text within images. Individuals have already started employing this feature to create counterfeit restaurant receipts, potentially expanding the already extensive toolkit of AI deepfakes utilized by fraudsters.
Deedy Das, a prominent social media user and venture capitalist, shared on X a photo depicting a falsified receipt for a legitimate San Francisco steakhouse, claiming it was generated using the 4o model.
Others have managed to produce similar outcomes, including instances where the receipt appears more authentic with the addition of food or drink stains. One such example by Michael Gofman demonstrates efforts to make the text appear more realistic by integrating it with the paper’s texture.
A particularly convincing example discovered by TechCrunch originated from France, where a LinkedIn user posted an image of a crumpled AI-generated receipt for a local restaurant chain.
TechCrunch conducted its own test of the 4o model and successfully fabricated a fake receipt for an Applebee’s in San Francisco. However, this attempt exhibited several signs of forgery, such as using a comma instead of a period in the total and inconsistent mathematical calculations. While language models struggle with basic arithmetic, these errors are not unexpected.
Nevertheless, it would not be difficult for a fraudster to rectify these numerical discrepancies using photo editing tools or more precise prompts. The ease with which fake receipts can be generated presents significant opportunities for fraudulent activities, enabling bad actors to claim reimbursements for non-existent expenses.
Taya Christianson, a spokesperson for OpenAI, informed TechCrunch that all images produced by ChatGPT contain metadata indicating their origin. Christianson emphasized that OpenAI enforces action against users who breach its usage policies and continuously learns from real-world application and feedback.
When questioned about the allowance of fake receipt generation by ChatGPT and its alignment with OpenAI’s usage policies prohibiting fraud, Christianson responded that OpenAI aims to provide as much creative freedom as possible. She also noted that AI-generated fake receipts could serve non-fraudulent purposes, such as education in financial literacy, besides being used for original art and product advertisements.
