The apartment in the StreetEasy photo looks brighter and more spacious than the walls allow. Soon, in New York, that digital touch-up will have to be disclosed. On July 16, the administration of Mayor Zohran Mamdani published the Rental Ripoff Report, a 23-point plan requiring landlords to declare when artificial intelligence has been used to enhance listing images. The immediate goal is to protect renters from deceptive representations, but the move uncovers a deeper issue that goes beyond the real estate market: mandatory transparency on AI use forces a rethinking of the entire content production chain.

Labeling an image as “AI-modified” is not just a voluntary declaration; it requires a reliable mechanism that attests to the origin and transformations of the visual data. Current tools for detecting generative AI are far from infallible and, crucially, do not provide incontrovertible proof in case of dispute. The only path to ensure compliance is to embed traceability directly into the pipeline that generates or modifies the images. This is where the debate shifts from regulation to technical architecture: whoever controls the entire stack can integrate cryptographic watermarking, forensic logging, and digital signatures without depending on third-party APIs.

A generic cloud provider offers image manipulation services, but the recording of every inference step remains opaque to the end user. If the New York regulator requested an audit, the real estate company would not have access to the detailed log of the models used on its photos. In contrast, an on-premise infrastructure—which locally manages generative AI models, perhaps with fine-tuning on proprietary data—can expose every step: from model checkpoint to prompt version, down to FP16 or INT8 quantization. This radical transparency becomes a compliance asset, not an added cost.

The Mamdani proposal fits into a broader picture. In Europe, the AI Act imposes transparency obligations for synthetic content; California is discussing similar rules. Companies that handle large volumes of images—from real estate to e-commerce, from publishing to healthcare—face a crossroads: outsource compliance to cloud services hoping that certifications are enough, or internalize the AI pipeline to have full control over the data supply chain. The choice is not only technical but economic: the Total Cost of Ownership of a self-hosted solution must be recalculated by including the reduction of legal risk and the ability to respond in real time to inspection requests.

It is no coincidence that orchestration frameworks like vLLM or TGI are gaining attention precisely for their ability to run in controlled environments. They allow serving image generation models with granular logging and integrating watermarking policies at the inference server level, without ever letting data leave the corporate perimeter. For those evaluating on-premise deployment, there are complex trade-offs—the initial investment in GPUs with sufficient VRAM for inference, the management of scalability—but the regulatory signal indicates that the cost of non-compliance may exceed the infrastructure investment.

Ultimately, the seemingly sector-specific detail of New York apartment photos reveals a structural tension of the generative AI era: trust in digital content can only be built with a level of audit that standard cloud services struggle to offer. On-premise, long considered a relic of the past, returns to center stage not for nostalgia but for governance necessity. And those who today label retouched apartment photos tomorrow might have to certify the entire decision-making process of an AI.