iOS 27's new generative photo features come with a specific list of things they will not do, and those refusals are doing more work than the superpowers line.
Apple's iPhone camera chief Jon McCormack framed the company's generative photo strategy this week as a way to give users "superpowers." The actual product, however, arrives with a list of things it will not do, and those refusals are the more interesting part of the announcement.
In a WIRED interview published this week around WWDC, McCormack said Apple is "not doing AI for the sake of AI." That line was meant to distinguish Apple from competitors, but it also captures the structure of what iOS 27's Photos app will actually ship. Two new generative features are coming later this year: Extend, which expands a photo's framing by up to 25% in a single shot by generating background pixels, and Spatial Reframe, which changes the perspective of an image. The existing Clean Up tool, which erases unwanted objects, is being upgraded using Apple's improved AI models.
Each feature is hedged with a specific constraint, and the constraints are consistent enough to read as a posture rather than a series of product calls.
Extend cannot be chained. A user gets one expansion of up to 25%, and that is the end of the chain. Clean Up cannot remove the primary subject of a frame, only surrounding clutter. The stated rule across these tools is that the main subject's face is not altered; only background pixels are generated or modified. McCormack used exactly that limit to position Apple against Google and Samsung, whose generative photo features allow users to erase and move people or insert objects into scenes.
The contrast is deliberate. Apple is publicly naming itself the more measured of the three, on the record, in a feature interview timed to its own developer conference. The bet is that users will treat measured edits as more trustworthy, and that the resulting photos will hold up better as records of what was actually in front of the camera.
The trade-off is real. Synthetic pixels are still being created and distributed as photographs. The background-only rule is partly a marketing line, since generating a plausible sky or wall is still generating something that was never seen. The 25% cap is an arbitrary product choice dressed as a principle. And the upcoming integration of SynthID, an invisible watermarking standard from Google DeepMind, is a partial signal that most viewers will never see or know to look for, which shifts the burden of provenance onto the audience rather than the platform.
What Apple is doing, then, is not just shipping features. It is publishing a set of fences around what a phone-generated photograph is allowed to be, and asking users to trust the fence. The features are bounded; the question of whether bounded fakery is the right answer to unbounded generative tools is the one this bet actually answers.
McCormack's superpowers line works as a hook. The constraints, named one after another, work as the argument.