Adobe’s latest move in Project Indigo adds a layer that goes far beyond simple background removal: the camera app now critiques your photos using artificial intelligence. It’s not an aesthetic filter to apply during editing, but a judgement that arrives before your finger leaves the shutter button. As Adobe describes it, the AI analyzes the shot and provides immediate feedback. Behind that seeming simplicity, however, lies an architectural decision that will resonate in the ongoing debate over data sovereignty and on-device deployment.

A bit of technical context is in order. Removing a background in real time demands precise semantic segmentation, a task that until recently was the domain of cloud servers with massive GPUs. Shifting that operation to a smartphone—and adding a visual critique module that relies on vision models trained on aesthetics and composition—means squeezing complex inference pipelines into a handful of watts. Adobe is betting that the neural engines embedded in modern mobile SoCs are now capable of running these workloads without calling on remote data centers. If that bet pays off, the gain isn’t just in lower latency: it would mark a decisive step toward local processing, with user data never leaving the device.

This choice has a direct impact on privacy and on how users perceive control over their images. In an era when every photo uploaded to the cloud raises questions about model training, profiling, and GDPR compliance, the ability to keep the entire workflow on-device becomes a differentiating asset. Adobe, traditionally anchored in a subscription and cloud services ecosystem, seems to be taking the opposite road with Indigo: minimize data transfer and reinforce a distributed computing model that multiplies the value of the hardware we already carry in our pockets. It’s not an isolated revolution; it follows the path already set by other photo-editing apps that have progressively moved inference to the edge. But the novelty here is substantial: aesthetic critique requires a level of semantic understanding that strikes at the heart of the relationship between human creativity and automated judgment.

The second-order implications are significant. If an AI becomes capable of evaluating a shot against recognizable criteria, the photographer—professional or amateur—gains an instant assistant that can correct composition errors or suggest alternative angles in real time. This shortens the learning cycle and reduces reliance on prior expertise, but at the same time raises the risk of stylistic homogenization: if the feedback is generated by a model trained on a corpus of “successful” images, it may discourage experiments that deviate from the norm. Who wins and who loses? Creators seeking technically sound images get a personal tutor without sacrificing data privacy; cloud service providers, on the other hand, see a slice of the workload that previously ran on centralized infrastructure migrate away. For Adobe, the move can become a competitive advantage, differentiating it from rivals that double down on the cloud.

From a structural perspective, Project Indigo signals that the battle for on-device AI has now reached mainstream creative tools. It’s no longer about voice assistants or trivial photo filters, but functions that require sophisticated content interpretation. This pushes mobile hardware manufacturers to invest in specialized silicon for inference, while framework developers must optimize ever-larger models for resource-constrained environments. For enterprise teams evaluating on-premise deployment of AI workloads, the parallel is clear: what a phone can do with one or two watts could translate into local servers that offer minimal latency and data sovereignty without relying on hyperscalers. The scale differs, but the logic is the same: move computation where privacy and responsiveness are needed most.

One open question lingers: to what extent will an AI’s aesthetic judgment become a perceived standard, and what will that mean for creative diversity? Automatic critique marks a subtle transition: we’re not just delegating the execution of a task, but beginning to entrust an algorithm with a qualitative assessment of our work. The line between assistance and conditioning becomes blurred, and the history of creative tools teaches us that every new technology reshapes not only what we can do, but also what we consider desirable.