Artificial intelligence is becoming a real headache for the accounting departments of many companies, and not precisely because of the automation of tasks that it requires. can leave you without work. Now, the big threat comes from an unexpected source: the ease with which fake expense receipts can be generated thanks to the skills of generate images hyperrealistic with AI.
AI at the service of fraud. Until recently, expense fraud was concentrated in already known areas: small traps in subsistence, transportation or purchase receipts that employees presented for reimbursement. However, as highlighted by a published article in it Financial Timesthe number of refund requests based on falsified receipts has increased dramatically in recent months.
As detailed in the analysis From AppZen, since OpenAI launched ChatGPT and Google launched its Gemini with image generative AI, expense receipt fraud has increased by 14%. The problem is not that they have increased since then, but that they have improved their quality so much that each time it is more difficult to detect them. “These receipts have gotten so good that we tell our customers, ‘don’t trust your eyes,'” Chris Juneau, senior vice president and director of product marketing at SAP Concur, told Financial Times.
Indistinguishable receipts. As they practically showed on the podcast The Accounting Podcastgenerating receipts for fraudulent expenses that can pass as authentic, is within anyone’s reach. It’s not even necessary be a paying user from ChatGPT or Gemini.
With just a handful of commands entered with a natural language In ChatGPT, the ticket obtained would fit perfectly as one of the thousands of receipts for food or transportation expenses that companies process daily.
AI to detect AI. It may seem that the problem of counterfeit tickets is a minor problem because they involve relatively small amounts. However, the average economic loss from false receipt fraud in the US is $133,000 on average, according to a survey prepared by Medius in 2024 to senior financial executives.
The expansion of this type of financial fraud has reached the ears of OpenAI, one of the main generators of these receipts, has taken letters in the subject including metadata in your images to identify that it is AI-generated content. This has allowed fintech companies like Ramp to develop AI-based tools that analyze the metadata of this type of receipts to stop fraud. The fintech claimed that with this software it had detected false receipts worth one million dollars in just 90 days.
The law is made, the trap is made. Faced with this movement to detect fraudulent receipts through metadata, “cheating” employees have also chosen to refine their techniques using more ingenuity: take a photo of the receipt generated by AI and attach it as a receipt, just as you would do as a real receipt.
That way, the image no longer includes the AI agent’s metadata, returning to square one. “It will be a continuous arms race. We just have to stay ahead,” he said. in statements to Bloomberg Karim Atiyeh, co-founder and CTO of Ramp.
Spending patterns. In response to this new strategy, AI uses what it does best: analyze patterns. These expense receipt “verifier” agents take into account the context in which these receipts could have been generated if they were authentic (travel schedules, date coincidences, repetition of names, etc.).
Based on this data, the system determines whether the receipt is false or not, without needing to verify the metadata. “Technology can analyze everything in such detail, so much so that humans, after a while, miss things; they are human,” he assured the Financial Times Calvin Lee, Senior Director of Product Management at Ramp
Image | Gemini
GIPHY App Key not set. Please check settings