ChatGPT’s New Picture Generator Makes Faking Receipts Easy

In a post on X, CEO Sam Altman announced that all users (yes even for free users) may now access OpenAI’s new picture generator, which is based on its GPT-4o model. Up until now, only ChatGPT subscribers who paid could access the service.

Altman stated last week that customers on the free tier may only create three photographs per day, however it is unclear how many images they can produce.

Altman said that the demand for OpenAI’s picture generating tool was so great that the company’s GPUs were “melting” as a result of its rapid success. The program also soon became well-known for being used to transform images into the aesthetic of the Japanese animation business Studio Ghibli, which raised questions about copyright and training data utilized by the company due to the stylistic similarities.

It was also used by some to create phony restaurant bills and other receipts. According to an OpenAI representative who talked to TechCrunch, ChatGPT created all of these photos, and the firm “takes actions” if the photos don’t follow its policies.

At a $300 billion value, OpenAI said today that it has secured $40 billion in investment, backed by SoftBank. Additionally, according to the corporation, ChatGPT now has 700 million monthly active users and 500 million weekly active users.

As part of its 4o model, ChatGPT debuted a new image generator this month that is far more adept at producing text inside images. 

People are already using it to create phony restaurant receipts, which might add another tool to the fraudsters’ already vast arsenal of AI deepfakes.

Effective VC and social media poster Deedy Das shared a picture of a phony receipt for a (actual) steakhouse in San Francisco on X, claiming that it was made using 4o.

Similar outcomes were reproduced by others, one of which included food or drink stains to add even more authenticity:

The example that TechCrunch considered to be the most realistic came from France, when a LinkedIn member shared a wrinkled AI-generated receipt for a nearby restaurant chain:

In addition to testing 4o, TechCrunch was able to create a phony receipt for a San Francisco Applebee’s:

However, there were a few dead signs that our endeavour was a fake. One is that a comma is used in place of a period in the total. The math doesn’t add up for another. It’s hardly shocking that LLMs still have trouble with simple math.

However, a fraudster might easily correct some of the figures with photo editing software or maybe more specific instructions.

It’s obvious that fraud prospects are much increased when it’s so simple to create phony receipts. It would be easy to foresee evil actors using this type of technology to receive “reimbursed” for completely fictitious charges.

All of its photos include information proving they were created by ChatGPT, according to OpenAI representative Taya Christianson, who talked to TechCrunch. Christianson went on to say that OpenAI is “always learning” from comments and real-world use, and that it “takes action” when users break its usage guidelines.

TechCrunch then questioned if ChatGPT’s usage standards, which forbid fraud, were consistent with its initial policy of permitting users to create phony receipts.

OpenAI’s “goal is to give users as much creative freedom as possible,” Christianson said, adding that phony AI receipts may be used for non-fraudulent purposes such as “teaching people about financial literacy” in addition to producing original artwork and advertisements.