Google Gemini saree trend sparks privacy scare after AI adds unseen body detail

A saree trend goes viral — and a privacy scare follows

A viral Instagram craze that turns everyday selfies into glossy 1990s Bollywood-style saree portraits has taken an uneasy turn. An Instagram user, who posts as Jhalakbhawani, says the AI-generated image she got from Google Gemini added a mole on her left hand — a real body detail she never showed in the photo she uploaded. Her post set off a wave of comments split between alarm and skepticism.

The “Nano Banana” saree edits — a nickname users have given to the Gemini Nano-powered look — pack a very specific visual punch: chiffon sarees, golden-hour lighting, jasmine-studded hair, and cinematic backdrops. Skin looks lacquered and doll-like, with 3D figurine vibes. It feels nostalgic, familiar, and hyper-shareable, which is why timelines across India are full of retro Bollywood makeovers.

The scale is staggering. By mid-September 2025, users had created or edited more than 500 million images in the Gemini app, with many more spilling across other platforms. In India, Gemini shot to the top of both Apple7s App Store and Google Play, adding an estimated 23 million users between August 26 and September 9. DeepMind chief Demis Hassabis publicly praised the surge, calling it a beginning rather than the peak.

Inside that flood, Jhalakbhawani7s story stood out: the AI-generated saree portrait showed a mole on her left hand. It exists on her body, she says, but wasn7t visible in the original selfie. She called it creepy. Commenters went in two directions — some saw it as a red flag for privacy, others chalked it up to coincidence or attention seeking.

Why would a model add a detail like that? Generative systems don7t just copy pixels; they synthesize new ones based on patterns learned from huge image datasets. When they 22upgrade22 a selfie into a stylized portrait, they can insert realistic texture — freckles, moles, pores, light hair — because those details make skin look convincing. Depending on the prompt, the seed, and the model7s training, a mole might get 22imagined22 in a spot that happens to match someone7s real body.

That doesn7t mean the system actually knew anything private. Diffusion models generate from noise toward an image that fits the prompt and the learned distribution of 22how skin tends to look.22 If the reference hand was partly cropped or flat-lit, the model might still add contrasty skin features to sell the cinematic look. Sometimes texture packs or style presets contribute similar effects over and over, which can feel eerily specific.

Still, the unease isn7t baseless. Could an app infer details from other photos on a device? On modern iOS and Android, apps have to ask for media access, and users can limit permission to selected images. Some creative apps upload photos to the cloud for processing unless an on-device mode is available. If cloud processing is used, images can be briefly stored for rendering or quality checks before deletion — the specifics depend on the app7s policy.

Gemini Nano is designed to run on-device for speed and privacy, but not every effect or filter is guaranteed to stay local. Sophisticated styles, bundled models, or queue-based upscaling often rely on servers. The 22Nano Banana22 label speaks to the model family, not a promise that every transformation avoids the cloud. Without a technical breakdown from the developer, users have no easy way to know which path their selfie took.

That7s where data practices matter. The key questions are simple: Did the app upload the image? Was it stored, even briefly? Is it used for improving models by default, or is there a clear opt-in? Are images or prompts retained in logs? Transparent answers to those questions would cool a lot of nerves — especially as the edits get more lifelike.

India7s Digital Personal Data Protection Act (2023) leans heavily on consent and notice. A selfie is personal data, and a stylized portrait that infers or emphasizes body characteristics is still personal data. If any processing happens off-device, the app must be upfront about what7s collected and why. Even on-device features can raise trust issues if the app requests broad media permissions it doesn7t need.

There7s also recent history. In late 2022, AI avatar tools surged, then drew backlash when some outputs sexualized users or invented intimate body details. The Lensa boom was a lesson: stunning results, messy edges. The saree trend is gentler, but the core tension is the same — the tech is great at cinematic fantasy, and it sometimes crosses a line.

Culturally, this moment makes sense. The 90s Bollywood aesthetic is a comfort look — dreamy lighting, billowing fabric, saturated romance. AI makes that production-level fantasy as easy as a tap. You get the flattered version of yourself, the poster you never got to pose for. The dopamine hit is real, and it7s why these filters spread like fire across WhatsApp groups and family chats.

How AI might add 22new22 body details, and what you can do

Think of generative AI as a very confident painter. It doesn7t copy an image; it paints a new one guided by your face, pose, and prompt. In that repainting, it uses its memory of what skin, fabric, and light usually look like. Moles, freckles, and tiny hairs often make an image read as 22real22 — so the painter adds them. If one lands where you actually have a mole, it7s uncanny. It can also be pure coincidence.

There are also technical reasons for these uncanny hits:

• Style presets: Some styles add freckles or moles as part of a 22cinematic skin22 recipe. The position can vary run to run.
• Pose completion: If a hand is partly obscured, the model 22completes22 it, adding texture for depth and realism.
• Seed randomness: Small changes in the random seed can change tiny features while keeping the overall look.
• Training priors: Models learn that realistic hands often have small marks, veins, or spots; they insert them to avoid the 22plastic22 look.

Could a system peek beyond your chosen photo? If you gave broad photo-library access, that7s a theoretical risk, but mobile OS controls now limit that by default. If you allowed only one or a few photos, the app shouldn27t see the rest. Cloud processing, if used, would involve the images you submit — and policy decides how long those remain on servers and whether they27re used to improve the model. As of now, there27s no public technical explanation tied to this case that shows a data leak or gallery scan.

For users who want the look without the stress, a few settings and habits help:

• Use 22Selected Photos22 (iOS) or the system photo picker (Android) so the app sees only what you choose.
• Check if there27s an on-device or 22process locally22 toggle; prefer that for face edits.
• Opt out of 22Use my data to improve services22 if that27s offered in settings.
• Avoid uploading images that reveal scars, tattoos, or kids27 faces if you don27t need to.
• Read the app27s data retention policy; look for clear deletion timelines.
• Revoke camera or photo permissions after you27re done, then grant again when needed.
• Clear in-app caches; many editors store recent files locally for convenience.
• If a result feels too revealing, don27t post it; the safest data is what never goes online.

Developers can meet users halfway by labeling when an effect runs on-device versus in the cloud, adding a simple 22why we ask for this permission22 screen, and making opt-outs obvious. People will still chase the fantasy shots — but they27ll do it with less guesswork about where their faces go and what the model might infer.

The saree portraits aren27t going anywhere. But a single mole in the wrong place has reminded millions that our most playful tech is also our most intimate — and that trust is now part of the filter.