AI in Your Phone's Camera — What's Real vs. Marketing

"AI camera" covers two genuinely different things: computational photography (which has quietly used machine learning for years and is mostly real) and generative "AI features" bolted onto the camera app (which range from useful to vaporous).

The clearest real example

Google's Gemini Nano runs entirely on-device on Pixel phones — works offline, no server round-trip — for specific, documented tasks: TalkBack's image descriptions, Pixel Recorder's on-device summaries, and text rewriting. Google's own developer blog is explicit that it runs locally.

Where Apple is more reticent than assumed

Specific Neural Engine "TOPS" numbers circulate constantly online for Apple chips. Checking Apple's own newsroom press release for the iPhone 16 Pro and its A18 Pro chip directly: there is no TOPS figure anywhere in it. Every specific TOPS number attributed to an Apple chip in circulation is a third-party estimate, not an Apple-published spec.

What this means practically

Established computational photography — multi-frame HDR merging, portrait-mode depth segmentation, night-mode stacking — is real, mature technology across Apple, Google, and Samsung, increasingly relabeled "AI" for marketing even though it predates the current hype cycle. Newer generative claims should be verified against the phone maker's own documentation for that specific model, not assumed from a generic "flagship phones have this now" claim.

The honest take

On-device AI camera features are real and increasingly capable, but the specific performance numbers attached to them are frequently third-party estimates dressed up as fact.