‘We detected unusual activity’: the scam that uses AI to exploit your holiday photos

You are on a short break in Porto and post some pictures of your family on Instagram or Facebook. With a small section of the Douro river in the background, you think it could have been taken anywhere. A few days later, you get a text saying that your card was compromised.

You click on the link to confirm your bank details. Since you have not put any details of your trip on any of your social networks – bar the indistinct pictures – you don’t suspect there is anything suspicious about the messages. But the text was a fraud designed to extract your financial details.

The criminals behind it figured out where you had been on holidays – lending credibility to their text – by using AI to analyse the image for the most sparse signs of where it was taken. New research has shown that by picking up on details in the picture – such as the background, the architecture, the signage or the light – the AI agent can pinpoint where it was taken. For criminals, this information gives attempts to defraud by text message or email an added legitimacy.

McAfee, the producer of anti-virus software, used two freely available AI models to test more than 21,000 travel images. One of the models identified 91% of images accurately while the other got 87%. Staff were then asked to replicate the experiment with their own pictures and became uncomfortable with how easy their travels were pinpointed.

The company says that it shows that computers don’t need photos to be tagged or attached metadata to identify where they are taken. Pictures are more likely to be identified by AI if they have recognisable landmarks, skylines, signage and street markings. Food stalls and storefronts can also pinpoint quickly where the picture was taken.

If the picture is taken on a beach or a hotel room, the accuracy of the AI is lowered. But McAfee reports that it is likely that the system can identify which country they were taken in – which is all the scammers need. When a staff member tested ChatGPT to identify a picture of a river with some trees in the foreground, it correctly pinpointed it as Hastings-on-Hudson, an area in New York state.