GoKawiilGoKawiil Original How to Tell Whether an Image Was Generated by AI
The visual tells that worked two years ago are disappearing. What still works is checking where a picture came from rather than how it looks.
C2pa
9 GoKawiil briefs on this topic
Apple details cryptographic chain behind iPhone 18 Pro's Reference Image mode
Apple published a Security Research blog post explaining how its new Reference Image camera feature, launching on the iPhone 18 Pro, verifies that photos are authentic from the moment they're captured. The system creates a secure digital negative with signed metadata and timestamps inside Apple's Private Cloud Compute environment, and uses a sensor-level trust chain with revocation lists to flag compromised hardware or fraudulent images.
Apple proposes hardware-based system to verify authentic iPhone photos
Apple has outlined a new approach called Reference Image aimed at proving whether a photo genuinely came from a real camera sensor, rather than relying on after-the-fact metadata like the C2PA standard. The method uses iPhone's secure hardware and Apple's Private Cloud Compute infrastructure to build a chain of trust starting at capture, rather than tracking edits after the fact.
Apple adds hardware-based 'Reference Image' authentication to iPhone 18 Pro camera
Apple's upcoming iPhone 18 Pro and Pro Max will include a Reference Image mode that uses the camera sensor to cryptographically sign captured photo data, producing an unalterable record of what the camera originally captured. Developers will get access to a Reference Image API on iOS, iPadOS, and macOS so third-party apps can eventually let users verify these signed images, though the feature won't ship at launch in the EU.
Apple adds hardware-based 'Reference Image' mode to iPhone 18 Pro to verify authentic photos
Apple has introduced a new camera mode called Reference Image, launching with the iPhone 18 Pro and Pro Max, that cryptographically signs sensor data at capture to prove a photo hasn't been altered. Using Private Cloud Compute, the phone generates an unchangeable reference version of the image that users can compare against edited copies in the Photos app. Apple is also opening a Reference Image API for developers across iOS, iPadOS, and macOS, though the feature won't be available at launch in the EU.
Anthropic launches tool to detect Claude-generated images, audio and video
Anthropic has released a web page at claude.com/check-content that lets users upload files to see if Claude was involved in creating them, using embedded content credentials under the C2PA standard. The checker currently supports only multimedia formats like JPEGs, GIFs, MP3s and MOVs, not text documents. A separate text-detection API exists but remains in private preview for select organizations.
Google Messages adds Content Credentials to reveal AI editing history
Google Messages now supports Content Credentials, a feature based on the C2PA standard that shows compatible photos, videos, and audio's origin and editing history, including whether AI tools were used. Users can press and hold supported media, tap Info, then select 'How this was made' to view this data. The feature is currently limited to Android devices.
Security researcher shows C2PA image-authentication scheme breaks down on Android
Researcher David Buchanan demonstrated that C2PA, the standard meant to cryptographically verify photos come straight from a camera, can be defeated on Android devices. He shows that root exploits, including a currently in-the-wild one-click attack (CVE-2026-43499) against fully patched Google Pixel phones, let attackers bypass Android's Key Attestation and Play Integrity protections that C2PA camera apps rely on, allowing arbitrary files to be signed as if they were genuine camera captures.
Microsoft Paint and Photos embed hidden server-issued ID in AI images, even offline ones
Reverse engineering of Windows Paint and Photos apps found that both send user prompts to a Microsoft server for moderation even when image generation runs locally via onboard AI models. That server returns a GUID alongside the approved prompt, and this GUID gets invisibly embedded into the pixels of the generated image, separate from the visible watermark toggle users can control. This applies even on Copilot+ PCs where generation happens entirely on-device.