Instagram’s AI detection is a mess (again)

Jess Weatherbed reports that Instagram's visible AI labels have gone haywire again: users say Meta is auto-applying an "AI Content" tag to images they did not create or edit with generative AI, while actual AI imagery often slips through. Triggers reportedly include Canva's Background Remover and minor blemish fixes — assistive tools, not text-to-image generators. A similar "Made by AI" misfire hit Instagram in 2024 after Meta said it would scan IPTC and C2PA metadata.

Canva told content strategist Jess Bruno that some assistive tools "were being tagged as generative" and claimed a fix; Threads users still report Background Remover triggers, and some pre-fix edits were never tagged. Weatherbed's own tests — Canva, Photoshop, Firefly, Gemini/Nano Banana, Apple Intelligence — found that only images edited or fully generated in Meta's own AI app got labeled. Fully generated Gemini images with C2PA and SynthID stayed unlabeled for nearly two weeks on a new account.

Signal Desk's read: Meta has built a trust theater that currently fails both sides of the test. False positives punish ordinary creators and brands (About Face's social manager said iPhone photos with light Photos-app edits were tagged, with no AI used). False negatives mean the label cannot be trusted as a synthetic-content detector. When the only reliable trigger in a reporter's battery is Meta's own generative stack, the system looks less like industry-standard provenance and more like self-preferencing detection.

The 2024 precedent matters. Meta already promised to weigh "the amount of AI used" after retouching metadata swept up minor edits. Repeating the same confusion in 2026 suggests either the metadata ecosystem is still incoherent — Canva's assistive tools getting generative tags — or Instagram's scanners are opaque and poorly calibrated. Weatherbed is right to ask what signals Meta actually reads; secrecy that protects against gaming also protects against accountability when the product gaslights users.

The poison-image case is especially ugly: a poisoned photo tagged, its unpoisoned twin not. That is the opposite of a coherent integrity story. Expect creators to strip metadata, re-export, or avoid assistive editors — exactly the arms race labeling was supposed to reduce.

Opinion: until Meta publishes clearer thresholds and third parties can reproduce results, "AI Content" on Instagram is a liability signal, not a consumer protection.

Meta announced in February 2024 that it would scan images for IPTC and C2PA metadata to label AI-made or AI-manipulated content, then said it would tweak labeling after minor generative retouches were swept up. Last week Meta also announced crackdowns on undisclosed AI accounts — which makes inconsistent image labels harder to defend.