Artificial generative intelligence has made another important leap forward, thanks to the new generator of images based on the GPT-4O model Openaialready available for everyone on Chatgpt, including users who have a free account. And the novelty this time could have an impact on the pockets of consumers and the anti -theft systems of companies. Chatgpt, being now able to create extremely realistic graphics, including readable and consistent texts, is arousing many concerns, after the improper use of Miyazaki’s “Studio Ghibli” style images, for the possibility that it gives to generate false receipts with the AI. This means that, in all respects, it is possible to ask an AI model to create a fictitious receipt for a restaurant, a shop or any commercial activity. And these are not simple abstract examples: online, likely likely generated receipts are already circulating with this technology, in some cases even “dirty” digitally with food stains or folds to seem used. The risk is clear: although the tool has been designed for creative and didactic purposes, it becomes extremely easy to abuse it to implement fraud, for example by inflating reimbursement of expenses or presenting false tests in legal or insurance contexts.
The new skills of the GPT-4O model
The main novelty concerns the ability of the new 4th model to generate perfectly legible text within the imagesa technical obstacle that until recently represented a limit for generative and which allowed to distinguish the images that artificially generated from real images with much more ease than now. Here to produce images of receipts where figures, names and graphic details appear clear and credible is nothing more utopian. Some users on social media, including the X @deedydas account, have shared impressive examples: receipts of real restaurants, digitally modified with plausible names, coherent prices and an almost totally indistinguishable visual aspect from an authentic document. In some cases, to increase its credibility, elements of “disorder” such as spots or folds have been added.
It is important to know that large language models, the so -called Llm (Large Language Models) like GPT-4O they are not yet perfect in the numerical field: often the amounts do not come backoi monetary symbols are not the correct ones. But these small mistakes can be easily corrected with a simple photo editing software or by providing more detailed instructions to the model. This makes the tool very attractive for those who want to falsify documentary tests quickly and economicly.
Openai’s answer
According to what was declared by Openaiall images generated by chatgpt include metadata C2PAa standard developed to indicate whether a content has been created by artificial intelligence or not. Taya Christiansonspokesperson for Openai, asked by Techradarin this regard he explained:
We monitor the generations of images inside and outside our platform, we use internal tools to verify that they have been created by our products and act when we identify the violations of our use policies. “We always learn from the use and feedback of the real world and we will continue to perfect our policies to balance creative freedom with the prevention of improper use. All images include metadata Standard C2PA in the sector that indicate that they were generated by the Ope by Openi.
Too bad that the metadata to which the Openii spokesman refers can be potentially removed in a handful of clicks during the export of the image, making it difficult, in practice, distinguishing a false receipt from a real.
The technological information site Techcrunchin asking Openai for the generation of similar fake content to be allowed, received as a response that “(the goal) of Openai is to offer users as much as more creative freedom possible»And that the false received generated by the AI could be used for”teach people financial literacy to people». To you the judgment on these declarations.
