Fake receipts generated with ChatGPT.
Credit: x.com/deedydas and Raphael Chenol
Image generation is one of the most eye-catching elements of the new generative AI craze, and it continues to have serious real-world ramifications. Following the latest update to ChatGPT's image generation with its 4o model, users have found that it is particularly good at generating realistic-looking receipts, complete with food and drink or retail orders. The potential ramifications for customer service are wild.
Alongside text generation, most of the latest generation large language model (LLM) artificial intelligence tools, like OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and Microsoft's Copilot, can all create images from a simple text prompt. Although those have been used to great effect to generate all sorts of images—lately, a lot of Studio Ghibli-inspired artwork—they've also been used in nefarious ways. The latest attempt at that is to create realistic-looking receipts, which could conceivably be used to "prove" you ate at a restaurant or purchased something from a store.
As with most things AI-generated, they're not perfect, with some discrepancies in the cleanness of the text or the fact the lighting doesn't match the crinkles in the paper. But it's good enough that an employee isn't going to think twice before approving this as an official receipt.
Tweet (opens in a new tab)
TechCrunch dug up a few other examples that are as good, if not better, and suggests that the war on whether these are realistic or not has already been lost. We went from a world that could trust images of receipts to one that can't overnight.
While you can find problems in each of the examples, they would be easy to fix with further AI prompting, or better yet, manual Photoshop manipulation. While that might be slightly more time-consuming, it's easy to see how getting a refund for $100 or so would be worth 10 minutes in Photoshop and the ChatGPT app. They wouldn't even need to work all the time. As with most scams, if it works once in a while, that's more than enough to make it profitable.
Unfortunately, the longer-term fallout of this will just be that image verification becomes obsolete. They can join the cloned voices of family and friends, deepfaked videos, and personalized spam emails as byproducts of the growth of AI. How fun.
In the near term, however, OpenAI has said it will crack down on the use of its AI for this. But the genie is already out of the bottle and Pandora's box is well and truly open. The receipts are in, and they're phony. Maybe.
Tagged In
ChatGPT Artificial Intelligence Scams
More from Computing