Viral 1980s AI Photo Trend: How Safe Is It To Upload Your Face?

Social-media feeds are filling up with AI-generated portraits that transport users back to the 1980s, complete with voluminous hairstyles, colourful outfits, neon lighting and grainy film effects.

The trend involves uploading a personal photograph to AI chatbot and asking it to recreate it as a retro portrait. The realistic results have helped the format spread, but they have also revived privacy questions previously raised during viral Ghibli-style and caricature trends.

The central issue is not the visual style. It is the information users hand over when they upload an identifiable photograph to an online service.

A Selfie Can Reveal More Than Face

A clear photograph of someone’s face is personal information because it can identify that individual. The original image may also contain embedded EXIF metadata recording when it was taken, the device used and, if location services were enabled, GPS coordinates.

Wired Magazine that photographs can carry location data that is invisible in the image itself. Although some online platforms remove this information during processing, users should not assume that every service handles metadata in the same way.

The visible background can be equally revealing. A photograph may inadvertently show a home address, vehicle registration plate, workplace, identity document, school uniform or another person who has not consented to having their image processed.

Uploading a photograph also means transferring a copy from a device to the platform’s systems. How that content is subsequently handled depends on the service, account type and privacy settings.

What OpenAI Says About Uploaded Images

OpenAI’s privacy policy says the company collects user content submitted to its services, including prompts, files, images, audio and video. This does not mean that every photograph uploaded to ChatGPT is automatically used to train a model.

For personal ChatGPT accounts, users can decide whether new conversations contribute to model improvement. OpenAI says the option can be disabled by going to Settings, selecting Data Controls and turning off “Improve the model for everyone.” Conversations remain in the user’s history after the setting is disabled, but are not used to train ChatGPT. OpenAI's Data Control FAQ says the preference applies across the user’s account.

Temporary Chat offers another option. According to OpenAI's guidance, temporary conversations are not used to improve its models and ordinarily do not appear in chat history unless the user saves them. The company may retain a copy for up to 30 days for safety purposes.

Different rules apply to organisational products. OpenAI says data from ChatGPT Business, Enterprise, Edu and its API platform is not used for model training by default.

Earlier Trends Prompted Similar Warnings

Cybersecurity specialists raised comparable concerns when users uploaded photographs and personal details for AI-generated caricatures.

“Every time these trends pop up,” ESET global cybersecurity adviser Jake Moore told Forbes, people are often “quicker to jump on the bandwagon” than to examine the implications.

Matt Conlon, chief executive and co-founder of security company Cytidel, told Forbes that users were “actively feeding increasingly detailed personal information into generative models” while trying to improve inaccurate results.

The concern is that a user may initially submit only a photograph but later disclose their age, occupation, location, family details or other identifying information while refining the image. Each additional instruction creates a more detailed collection of information within the conversation.

During the earlier Ghibli-style trend, Luiza Jarovsky, co-founder of the AI, Tech & Privacy Academy, also highlighted the distinction between photographs collected from public websites and images submitted voluntarily by users.

Writing on X, she said users were teaching themselves to upload personal pictures whenever they wanted “a fun avatar of themselves”. Her comments centred on the access AI services could gain to private photographs that might never otherwise have been publicly available.

The same distinction applies to the 1980s trend: users are knowingly submitting the source photographs, but may not have examined the platform’s settings before doing so.

How Users Can Limit What They Disclose

People participating in the trend can reduce their exposure by choosing a photograph that contains little information beyond what is needed to create the portrait.

Cropping the image can remove house numbers, documents, vehicle plates, computer screens and recognisable locations. Users can also check whether the original file contains location metadata and remove it before uploading.

Highly sensitive photographs—including intimate images, medical records and identity documents—should not be used as source material for an entertainment trend. Images containing children or other people should also be avoided unless those individuals, or their guardians where appropriate, have agreed to the processing.

Users concerned about model training can disable “Improve the model for everyone” before beginning the conversation or use Temporary Chat after reviewing its retention conditions. Someone who does not want a photograph processed in the cloud can instead use an image-editing application that works locally on their device.

AI Images Also Carry Energy Cost

Producing an AI image requires considerably more computation than applying a conventional filter. However, estimates of the electricity consumed by each generation vary widely according to the model, hardware, image resolution and measurement method.

A 2023 study involving researchers from Hugging Face and Carnegie Mellon University found that generative systems generally consumed substantially more energy than task-specific models. The researchers did not test ChatGPT’s proprietary image system, meaning their results cannot establish the electricity consumed by one portrait in this particular trend.

A separate 2025 study of 17 image-generation models found a 46-fold difference in energy consumption between the systems tested. That variation makes it misleading to present a single figure as the universal energy cost of an AI-generated image.