Venture capitalist DeedyDas posted on X a photo of a fake receipt from a (real) San Francisco steakhouse that he said was made with 4o.
Others have been able to replicate similar results, including ones with food or drink stains to make them look more authentic:
The most realistic example actually comes from France, where a LinkedIn user posted an AI-generated crumpled receipt from a local restaurant chain:
TechCrunch tested 4o and successfully generated a fake receipt from an Applebee's in San Francisco:
Several telltale signs point to it being a forgery. First, the total uses a comma instead of a period. Second, the math doesn't add up accurately. Large models still have a hard time getting basic math right, so that's not surprising.
But it's not hard for fraudsters to quickly fix a few numbers using photo editing software or more precise tips.
Clearly, making it so easy to forge receipts opens up huge opportunities for fraud. It is not difficult to imagine that this technology could be used by criminals to "reimburse" completely false expenses.
OpenAI spokesperson Taya Christianson said all of its images include metadata indicating they were produced by ChatGPT. Christianson added that OpenAI "takes action" when users violate its usage policies and that it is "always learning from real-world usage and feedback."
When asked why ChatGPT allows people to generate fake receipts and whether this is in line with OpenAI’s usage policy (no fraud). Christiansen responded that OpenAI's "goal is to give users as much creative freedom as possible" and that the fake AI receipts could be used in non-fraudulent situations, such as "teaching people financial literacy" and creating original artwork and product advertising.