NVDA 162.85 ▼2.01%GOOGL 295.77 ▼0.98%MSFT 523.88 ▲0.17%AMD 534.06 ▼8.75%INTC 59.66 ▼2.52%TSMC 398.73 ▼1.67%AMZN 266.66 ▼0.29%META 641.69 ▲2.90%AAPL 356.27 ▲0.39%PLTR 167.99 ▼2.09%
Markets at last close

Policy

Image labels cannot establish truth, standards experts warn

·1 min read

AI-generated image labels can improve transparency, but they cannot establish whether an image is true, according to standards specialists and photographers participating in discussions hosted by Giga in Geneva. As synthetic images become increasingly realistic, visual intuition is also proving unreliable. A Microsoft research study of over 12,500 participants worldwide found an overall detection accuracy of 62 per cent.

Since 2 August 2026, the EU’s AI act transparency rules require generative AI providers to add machine-readable marks and users deploying these systems to disclose deep fakes. China requires visible labels and embedded metadata, while India requires visible watermarks. No binding global rules exist, and experts advising the World Standards Cooperation warn that jurisdictions differ over disclosure requirements, responsibilities and enforcement.

Implementation remains uneven. An AI Forensics investigation found that 23 per cent of Instagram’s AI content was properly labelled. Markers can also be removed, limiting their reliability. Newsrooms and photography competitions instead combine professional ethics, metadata checks, reference images and fact-checking.

Experts identify cryptographic signatures applied at the camera as a way to strengthen provenance records. Such tools can help establish where an image came from, but cannot independently verify its meaning or accuracy. Trust still requires context, corroboration, journalism, forensic analysis and human judgement.

Originally reported by genevasolutions.newsRead the source →
Related coverage