Metaesthetic AI

Metaesthetic AI: the aesthetics of machine-made media

Metaesthetic AI is a free study of the look of machine-made images: the tells that make them recognizable, the training and tuning steps that produce them, and how to use or break the look on purpose. Every figure links to its source, from the Northwestern guide that sorts AI tells into five categories to the 2022 PNAS study in which people judged AI faces with 48.2% accuracy, and the vendor docs behind the print size planner.

The machine has a look. These pages name it, trace where it comes from, and show how to choose whether to keep it.

The signatureThe five categories of AI image tells, and how well people spot AI faces. OriginsTraining data, aesthetic filters, preference tuning and guidance, with figures. Working with itDesign rules, and the canvas sizes GPT Image and Nano Banana output. Breaking itGuidance, conditioning and texture: the levers that move output off the default. Labels and provenanceWho adds C2PA Content Credentials and watermarks, and who reads them. Print size plannerTool: print size in, model output sizes, crop and upscaling out.

Who marks or labels AI images?

The look is one signal; provenance is another. From each company's own pages, checked 13 November 2026.

Company Signal What happens
Google SynthID invisible watermark In every image generated with the Nano Banana models in the Gemini API; Gemini, Search and Chrome can check for it
OpenAI C2PA Content Credentials and SynthID Its provenance tool checks images for both; it does not detect other companies' AI content
Adobe C2PA Content Credentials Attached to Firefly generations, Adobe said in March 2024
Meta Visible marker, invisible watermark, IPTC metadata Reads C2PA and IPTC "AI generated" signals and shows an "AI info" label
TikTok C2PA Content Credentials Reads them to auto-label AI content uploaded from other platforms (May 2024)
YouTube AI disclosure label Applied automatically to content with C2PA metadata, content made with its own AI tools, and content it detects

Details and sources are on labels and provenance. A missing label is not proof that an image is real: editing, converting or sharing a file can strip its metadata.

Will it print?

Nano Banana 2.1 reaches 4096 x 4096 pixels at 1:1; GPT Image tops out at 8,294,400 pixels in total. The print size planner turns any print size into the pixels it needs and shows which output covers it.

Open the print size planner