These deepfake receipts are so sophisticated workers could cheat their expenses
Since the dawn of expense claims, workers have been adding a few extra drinks for themselves on the company card (legal disclaimer: of course never us).
But the possibilities for dodgy claims are getting much more sophisticated, with businesses warned to take note.
Image generation by AI is now so advanced that fake but convincing receipts can be created in seconds.
OpenAI released its latest image generation model last week, an update which caused so many people to create Studio Ghibli-style drawings that its servers ‘melted’.
But a less wholesome use case also emerged, as users showed how easy it is to create fake receipts.
Many companies ask employees to upload photos of invoices they need reimbursing for, before their expense claims are approved.
Metro quickly made our own fake receipt using ChatGPT, for an Italian restaurant called Mikey’s on West Street in London.
We had minestrone, lasagne, tiramisu and a bottle of house red according to the receipt, costing £63.80 including the service charge.
However, not only did we not really visit, but the restaurant doesn’t even exist.
Sure, you could work this out by googling it, or you might have suspicions from the commas and full stops being confused, but many times this would just be waved through.
Jon Clay, VP in threat intelligence at Trend Micro, told Metro that realistic AI images will take fraud and scams to the next level.
He said: ‘When you submit submit stuff for HR, if you’re an employee that’s a little shady, for example, you could potentially get some some money from fraudulent receipts.’
This isn’t the only headache for businesses, however, with business e-mail compromise attacks a lucrative target for scammers.
This is when a criminal gains access to a company email account with the intention to trick a worker with the ability to transfer funds, often a senior executive or account manager.
Jon explained: ‘A scammer could potentially impersonate an employee, submit a whole bunch of receipts, get payment for those receipts, and make money that way. They could definitely defraud businesses.’
Artificial intelligence expert Raphael Chenol shared photos on LinkedIn of a real receipt and a fake one he created with ChatGPT.
He said: ‘I just asked it to change the date and price… and my expense report entry application doesn’t see anything. It only takes a few seconds where before you needed Photoshop skills.’
Clarifying that he ‘wouldn’t risk being fired for fraud by my employer’ , he said he was sharing the images to denounce possible abuses of the technology.
Mr Chenol suggested that to avoid ‘montains of false invoices’, they should include a QR Code-style cryptogram as standard, or that they should be submitted with proof of purchase such as via Blockchain.
Meanwhile, to reduce the risk from scammers Mr Clay called on businesses to have a two step verification process by which someone is asked on a different platform if it is really them asking for money to be transferred.
He said that AI photos are becoming so realistic that it will be very difficult for people to spot them just by looking at them, and although metadata will say they were edited by AI, this is easy for scammers to strip out.
One way to reduce the risk would be to insist that AI generated images contain a visible watermark, potentially due to laws that insist on it.
‘As this technology gets out there and if there if we see a lot more abuses of the technology, we might see lawmakers start to recognise that,’ he said.
‘But then again, if I’m a bad actor, I’m not going to follow the law.’
The makers of Chat GPT, Open AI, explicitly ban the use of the service to harm others, including using any output to defraud or scam.
Spokesperson Taya Christianson told TechCrunch that their ‘goal is to give users as much creative freedom as possible’ and that fake receipts aren’t necessarily fraudulent as they could be used for ‘teaching people about financial literacy’ as well as for art.
Get in touch with our news team by emailing us at webnews@metro.co.uk.
For more stories like this, check our news page.