What Really Happens When You Turn Your Selfie Into a 1980s AI Pic?

We, as human civilisation, are having a small identity crisis online. 

It started out as fun, chasing endorphins over AI slop. Then we began feeding our selfies to AI. The latest trend is people giving ChatGPT a photo, asking for a 1980s makeover, and within seconds we have a version of ourselves from a decade we may not even remember, or were alive in, creating a nostalgia that never existed. 

At first, the images seem like harmless fun. But behind that one-click transformation is the considerable computing power it takes to generate the image and what AI companies get out of our sudden obsession with looking retro.

Then there’s what happens after the image is generated. OpenAI’s privacy policy says it collects user content, including prompts and other uploaded content. It also collects usage data, including features used, actions taken, and feedback on responses. 

That makes such trends almost like a data mine for AI labs. However, OpenAI says Temporary Chat conversations do not appear in history or get used to improve its models. It also says it generally removes deleted personal data from its systems within 30 days, subject to exceptions.

For someone following a trend, however, these choices may be easy to miss.

Your Selfie is Not Just a Selfie

Anwesha Sen, Programme Manager for Technology and Policy at the Takshashila Institution, says the issue is not only about what the policy says. It is whether users understand the choices before uploading an image.

“The actual privacy outcome for any given uploaded photo is essentially unknowable from the outside,” Sen tells AIM.

Privacy outcomes depend on whether users are aware of such settings. She also questions whether a setting buried in a product menu creates meaningful consent for casual users.

“Consent is closer to nominal than informed,” she says.

The concern becomes sharper with personal photographs. A selfie can reveal a person’s face and other details. A family photograph can also contain people who never chose to interact with the AI service.

OpenAI’s recently launched ChatGPT Images 2.5 uses safety checks on image inputs and generated outputs. Its system includes image classifiers, prompt checks, and a safety reasoning model.

OpenAI also says it trained its image models using several sources. These include public internet information, third-party sources, and information users, human trainers, and researchers provide or generate. That does not establish whether a particular viral selfie becomes training data. However, users need to distinguish between consenting to service improvement and model training.

Gemini’s policy makes the trade-off even clearer. 

If Keep Activity is on, Google says what users share with Gemini, including photos, can be used to improve its services and train generative AI models. Some of that material may also be seen by human reviewers, and reviewed chats may be retained for up to three years after disabling the user account. 

Users can switch Keep Activity off or use Temporary Chats, although Google says it still retains those chats for 72 hours for service and safety purposes.

The Hidden Bill

The other side of the transaction is compute. Generating an image requires inference, which means running a trained model to produce an output. One image may not look significant on its own. A viral trend can change the scale.

Images also demand considerably more computing power than ordinary text generation. Adarsh Ranjan, Research Analyst at Carnegie India, says available estimates suggest generating an image can require around 50–100x more energy than a simple text generation.

“When millions of users simultaneously generate images as part of a viral trend, such as the Ghibli-style trend or the more recent trend of generating 1980s-style photographs, the resulting surge in image-generation requests can substantially increase inference workloads,” Ranjan explains.

A December 2025 assessment by Germany’s Gesellschaft für Informatik, a network of computer scientists, says inference uses less compute per request than training. However, generative AI services run inference at enormous scale, and billions of daily queries can make inference a sustainability concern.

The same report links AI operations to higher electricity use, greenhouse gas emissions, and water use. Water comes from data centre cooling and indirectly from electricity generation. The impact varies with model size, hardware, workload, data centre design, and location.

However, no reliable figure exists to gauge the electricity or water used for generating one 1980s-style image. Ranjan says technology companies may publish the overall electricity and water consumption of their data centres, but rarely provide enough detail to isolate what AI inference consumes.

“I do not think frontier AI companies are sufficiently transparent about the infrastructure and resource costs associated with their products,” he says.

But the broader scale is easier to see. Non-profit Epoch AI estimated global AI compute had reached about 20 million NVIDIA H100 equivalents by the end of 2025. It estimated OpenAI accounted for around 10% to 15% of the world’s operational AI compute.

For India, that growth carries an environmental cost. “India already faces significant water stress,” Ranjan says. 

As data centre capacity expands, he argues that stronger safeguards will be needed around cooling, water and power efficiency, waste management, and where facilities are built.

AI Makeovers

The 1980s trend is hardly the first time people have fed their images to an AI model. In late 2022, Lensa AI’s Magic Avatars became one of the first major AI-photo crazes, turning users’ selfies into stylised portraits. 

The trend was big enough to push AI photo apps to more than 4.3 million daily downloads at their peak, according to Apptopia data. 

Then came the nostalgia phase. In 2023, Epik’s AI Yearbook trend turned people’s photographs into versions of themselves styled like 1990s American school portraits. The trend spread particularly quickly on TikTok, helped by K-pop celebrities and creators. 

Around the same period, AI-powered Barbie avatars also took off, with PhotoRoom’s official Barbie selfie generator reportedly used more than 13 million times after its launch. The appeal was becoming familiar: upload a photograph, choose a recognisable visual world, and let the model put you inside it.

By 2025, the scale and sophistication had changed. 

OpenAI’s image generation brought the 2025 Studio Ghibli-style trend, as people transformed selfies, family photographs and famous scenes into Ghibli-inspired images. 

The trend quickly sparked a debate over copyright, consent, and AI’s relationship with human art. Hayao Miyazaki’s 2016 criticism of AI-generated animation resurfaced, including his description of the technology as “an insult to life itself”. Artists and commentators also questioned whether OpenAI’s models had been trained on copyrighted works without permission, adding the Ghibli trend to a wider legal fight over AI training data.

Soon after, ChatGPT users turned themselves into boxed action figures, while Google’s Nano Banana pushed the format further with collectable figurines, retro portraits, younger-self images, and other highly specific edits. 

Google said more than 500 million images had been generated with Nano Banana since launch, with the feature helping bring more than 10 million new users to Gemini. 

A Business Model on Virality?

A deeper question is whether these trends are simply organic internet behaviour or also serve a business purpose. 

While evidence suggests that image generation based on trends can drive usage, it does not prove that companies intentionally engineer trends to collect data. But AI giants have the opportunity to turn the product feature into a user-acquisition mechanism.

The environmental report raises a related idea. It describes a “rebound effect”, where efficiency gains can increase demand rather than reduce total resource use. It says market incentives can reinvest efficiency gains to expand the market and stimulate new demand.

The report also points to opacity as a barrier to independent assessment. Companies may withhold information about model architecture, training data, and environmental impacts because of intellectual property concerns.

That leaves the 1980s selfie as more than an internet fad. It represents a microcosm of a much larger AI economy.

The user sees a retro photograph. The company sees another interaction with its product. Behind that interaction sit data policies, safety systems, GPUs, electricity, cooling infrastructure, and a race to grab the growing market.