One question runs through the entire week: what counts as proof? Apple wants to sign photos at capture time to show they weren't generated, ChatGPT now turns a back-of-envelope sketch into a realistic render, hikers had to be rescued after planning their route with Gemini, a French decree will move 300,000 homes out of the worst energy rating, and courts just tightened the rules on agency client databases. Five signals about what stays verifiable when everything becomes generatable.
A whole week around a single question: what counts as proof?
A photo, a render, an assistant's answer, an energy label, a client database. Five very different objects, and the same question behind each of them this week: what can you still rely on, and how do you prove it? In a trade where everything hinges on trust between a listing and a viewing, that's far from theoretical.
Apple wants to prove a photo wasn't generated
Apple is preparing a camera mode that would sign images at the moment of capture, attesting that they came from a sensor rather than a generator. The idea isn't to detect AI after the fact - probably a losing battle - but to certify provenance at the source, which is a far more tractable problem.
If this becomes standard, the question asked of a listing changes in nature. It won't be "is this photo retouched?" but "can you prove where it came from?". Agencies that start from a real property have nothing to fear from that shift, quite the opposite: traceability becomes a free selling point. Those publishing entirely fabricated interiors will discover that the absence of a signature is information too.
Source: The Verge, "Apple's new iPhone camera mode promises to prove your photo isn't AI", 2026.
From scribble to render: ChatGPT moves to sketching
OpenAI updated its image generation with a sketch feature: you draw roughly by hand, and the model produces a detailed visual from the lines. It's a meaningful shift, because it moves past the text prompt to give users spatial control - exactly what professionals were missing when they wanted a specific room rather than a nice random one.
For real estate, it means producing a credible interior from a scribbled floor plan is now within everyone's reach, in seconds. The consequence isn't that you should avoid it, it's that a generated visual no longer proves anything on its own. What keeps its value is the verifiable link between the image and the property: a render derived from a real capture, with the right dimensions and the right circulation, isn't in the same category as a living room invented from a sketch - even when both look good.
Sources: The Verge, "ChatGPT Sketch turns your bad drawings into detailed AI images" and Numerama, "OpenAI updates ChatGPT Images to version 2.5", 2026.
Hikers rescued after planning their route with Gemini
The story went around all week: hikers had to be rescued after following an itinerary prepared with Gemini, which described a route that didn't match the actual terrain. The model didn't lie in the usual sense - it produced a plausible, well-written answer about a place it has never seen.
That's the most useful reminder of the week for our trade. A general-purpose assistant asked about a neighborhood, a local regulation, or a property's potential will likewise produce a fluent, confident answer, sometimes correct, sometimes not, never flagging the difference. In the same vein, OpenAI's chief scientist admitted this week to losing visibility into his own models' reasoning. The practical rule hasn't moved: an AI answer is a hypothesis to verify, never a source.
Sources: TechCrunch, "Hikers rescued after using Google Gemini for planning" and Numerama, "OpenAI's chief scientist says he's losing control over his AI's reasoning", 2026.
300,000 French homes about to leave the worst energy rating
A change in calculation method planned for 2027 would move roughly 300,000 homes out of the worst energy performance bracket. The buildings themselves won't move an inch: same walls, same insulation, same heating bill. Only the label changes - and with it, the right to rent it out, the negotiating price, and sometimes the whole balance of a deal.
Set that against another piece published this week on the growing difficulty of producing an objective valuation in the energy-certificate era. Both say the same thing: a significant share of a property's value now rests on an administrative indicator whose calculation method keeps evolving. That's a direct concern for anything automating valuation, ours included: a model trained on yesterday's labels will produce wrong estimates in 2027, with exactly today's confidence. On this topic, knowing the regulatory calendar beats any algorithm.
Sources: Immobilier 2.0, "2027 energy rating: status change for 300,000 poorly rated homes" and Journal de l'Agence, "Property valuation: value opinions, agent liability and the objectivity challenge in the energy-certificate era", 2026.
The client database: an asset, and therefore a liability
Two stories this week answer each other. On one side, French courts hardened their position on agents joining a new agency with their previous employer's commercial files: the mere presence of those files on a computer is now enough to establish unfair competition risk. On the other, a cyberattack on the Zéro Logement Vacant service exposed the data of 48 million people.
Side by side, they show the two faces of the same subject. An agency's database is recognized as a valuable, protected asset that can't be moved around freely - that's good news. But an asset of that value is also a target and a liability, and many agencies can't say precisely where their data lives, who accesses it, or what walks out the door when someone leaves. Before plugging an AI tool into that database, it's worth having answered those three questions.
Sources: Journal de l'Agence, "An agent arrives with his former agency's commercial files: beware the unfair competition risk" and Immobilier 2.0, "Cyberattack: 48 million people exposed in a massive Zéro Logement Vacant breach", 2026.
The thread this week: what can be proven gains value
Photos signed at the source, renders produced in three seconds, an energy label that depends on a decree, a client database that carries legal exposure: everything verifiable gains weight, everything unverifiable loses it. That's good news for anyone working from the real property, and a warning for anyone getting comfortable publishing images they couldn't account for.
So, on your latest listing: could you say where every published visual came from?
