The latest wave of AI-powered image trends, from Google’s quirky Nano Banana figurine portraits to the elegant vintage saree photo edits flooding Instagram, has captured the imagination of millions of social media users. These hyper-stylised transformations, which allow ordinary selfies to be converted into cinematic or cartoonish portraits, are spreading at lightning speed. But beneath the charm of glossy plastic skins, retro saree drapes, and whimsical backdrops lies a deeper debate about safety, privacy, and the risks of handing over personal photos to powerful artificial intelligence systems. The question that continues to emerge is whether these fun experiments with Gemini AI tools are harmless self-expression or whether they expose us to long-term vulnerabilities in a digital landscape already struggling with deepfakes, identity theft, and blurred boundaries between the real and the artificial.
The Rise of the Nano Banana and Vintage Saree Craze
Unless you have been entirely disconnected from the online world, chances are you have either experimented with Google’s Nano Banana feature yourself or stumbled upon the viral saree AI portraits circulating across Instagram, TikTok, and countless other platforms. Both phenomena are powered by Google’s Gemini Nano model, part of the company’s suite of generative AI tools designed to make image creation intuitive and engaging for the everyday user.
The Nano Banana trend became the first breakout moment for Gemini’s image capabilities. It transforms an uploaded selfie into a hyper-detailed, three-dimensional figurine-like portrait. The results mimic miniature toys or animated characters, complete with oversized glossy eyes, smooth plastic-like skin, and proportions reminiscent of collectible dolls or action figures. For many, this offered a humorous, nostalgic, or even self-deprecating glimpse of how AI could reimagine personal identity in stylised cartoon form.
Almost in parallel, users began experimenting with another Gemini-powered effect: transforming their images into vintage saree portraits. This trend, which quickly went viral across South Asia and among diaspora communities, generated retro-styled cinematic photographs of mostly women dressed in traditional sarees. The images were often presented with elaborate backdrops resembling old Bollywood posters, sepia-tinted family portraits, or 1980s studio glamour photography. For some, it was a playful homage to cultural heritage; for others, a chance to experience an aesthetic they may never have had in real life.
Both trends highlight the power of AI not merely to alter photos but to rewrite identity and context, enabling people to experiment with entirely new selves. But as with every viral AI-powered phenomenon, these tools raise questions about the safety of our images, the protections in place to prevent exploitation, and whether watermarking and metadata are enough to safeguard authenticity in a world prone to digital manipulation.
The Promise and Limits of Watermarking Technology
Google has attempted to reassure users about the safety of its AI tools by embedding watermarking technology into all images generated or edited using Gemini’s systems. Known as SynthID, this invisible watermark is layered into the pixels of every AI-created photo. It cannot be seen by the human eye but can be detected by specialised tools designed to identify whether an image originated from AI or was significantly manipulated using generative technology. Alongside this watermark, metadata tags are also included to indicate that the file carries an artificial origin.
The theory behind SynthID is straightforward: if AI-generated images can always be tagged and traced, they cannot be easily passed off as authentic photographs. This could, in principle, provide a powerful defense against deepfakes, misinformation, or fraud. Google itself stresses that the watermark is a transparency measure, allowing platforms and researchers to build with confidence, while offering individuals reassurance that there is a technological trail linking back to AI when disputes or misuse arise.
Yet even as the watermark sounds promising, its limitations become quickly apparent. For one, the detection tools required to read SynthID are not currently available to the general public. According to reporting by Tatler Asia, the watermark can only be verified with specific technical systems not accessible to most internet users. This means that while a photo may be watermarked, the vast majority of viewers—including those most vulnerable to being deceived—cannot independently confirm its authenticity.
Experts further caution that watermarking is no silver bullet. Wired recently highlighted skepticism from researchers and industry leaders, including Ben Colman, CEO of Reality Defender, who noted that watermarks can be faked, erased, or simply ignored. Hany Farid, a professor at the University of California, Berkeley, has argued that while watermarking is a step forward, it will never be sufficient on its own. Instead, he suggests combining it with a suite of detection and verification technologies to make the creation of convincing fakes far more difficult.
The limitations of SynthID reflect a broader challenge facing the AI industry. As generative models become more powerful and accessible, any protective mechanism designed to distinguish real from fake must also anticipate attempts to circumvent it. Watermarks are one line of defense, but they risk offering a false sense of security if treated as an absolute guarantee of safety.
Navigating Safety in the Age of Viral AI Images
For everyday users swept up in the thrill of transforming their photos into miniature figurines or timeless saree portraits, the bigger concern is not whether the images carry a hidden watermark but what happens to their photos once uploaded to AI platforms. The vulnerability lies in user intent and platform policies as much as in technological protections.
Whenever we submit images to an AI system, we place trust in how the company handles, stores, and processes our data. Questions about whether uploaded photos are retained, shared, or repurposed for model training remain central to the debate. Even if a company like Google assures that content is handled responsibly, the broader ecosystem of third-party apps and less transparent platforms poses greater risks.
The simplest and most effective safeguard is for users to be selective about what they upload. Harmless selfies may be appropriate for experimentation, but intimate photos, private moments, or images containing sensitive identifiers like children or location cues can pose serious risks if misused. Metadata, too, can quietly reveal more than expected, such as device type or geotagged coordinates. Removing such details before uploading adds a layer of protection.
Another dimension of safety involves controlling visibility. Once an AI-generated image is shared publicly on social media, it enters a space where it can be copied, altered, or used out of context without the original creator’s consent. Privacy settings, restricted sharing, or even simply choosing not to post certain images widely can reduce exposure to potential misuse.
Users are also encouraged to retain original copies of their images and prompts. Having a backup ensures that if an altered or fake version emerges, it can be compared against the authentic source. This is particularly important in a time when manipulated photos are increasingly weaponised for harassment, misinformation, or reputational harm.
Equally significant is the fine print of consent. Many platforms reserve rights over uploaded images in their terms of service, which may include permissions to use data for research or model training. Understanding whether participation in a trend also means granting long-term rights over one’s likeness is essential before hitting upload.
