As ChatGPT becomes increasingly integrated into our daily lives, many users don’t hesitate to upload a photo to ask what a rash on their hand might be, to identify a flower they spotted on an afternoon walk, or even to polish their profile picture for LinkedIn. Others do it for fun, asking the chatbot to turn a selfie into a cartoon or to show how they might look when they grow old.

It may seem harmless, but experts warn that every time we upload a photo to ChatGPT, we share far more than we think.

What happens to these images once the interaction with the chatbot ends remains unclear. Different AI companies follow different policies, and there is always the risk of technical errors or the unpredictable evolution of this fast-moving technology.

Most users assume a level of privacy that may not actually be guaranteed.

“Do not upload any photos you would never want anyone else to see,” warns Jacob Hoffman-Andrews of the digital rights organization Electronic Frontier Foundation, speaking to The Wall Street Journal.

According to Hoffman-Andrews, users should treat chatbots like cloud storage platforms such as iCloud or Google Photos—but with added risks. Chief among them is security, since chatbots could become targets for hackers who may gain access to user accounts.

Beyond hackers, users’ privacy is also at risk from the AI companies themselves. These companies can access user data and images. To evaluate how well their models work, employees may review sampled conversations, meaning that an OpenAI staff member, for instance, could potentially see a user’s interaction with ChatGPT—even if that user has deleted it.

Even a simple photo can become dangerous in the wrong hands. Photos often contain metadata, such as the location and time they were taken, and can reveal sensitive details like bank account numbers, medical diagnoses, or home addresses.

What many users don’t realize is that, in some cases, their data is used to train AI models—even without explicit consent.

While a vacation snapshot or a picture of a recipe may get lost among the flood of data processed by chatbots, some photos carry a higher risk of being retained in AI “memory” and could resurface in future chatbot outputs in recognizable forms, WSJ reports.

According to Hoffman-Andrews, two categories of images are particularly “dangerous”:

• Photos that appear thousands of times online, such as popular memes.

• Photos with highly distinctive features that stand out from the statistical norm.

For the average user, an exact reproduction of their personal photos by AI systems is unlikely. However, uploading a high-resolution selfie increases the risk that a chatbot could generate images retaining identifiable traits.

Even if current AI company policies appear strict, no one can predict how the landscape will evolve as technology advances rapidly. In the long run, a photo uploaded today could be reused in ways very different from its original purpose.

“Whatever you upload is going to have a life far longer than the moment you use the system,” says Sarah Myers West of the AI Now Institute, which studies the societal impacts of artificial intelligence.