On October 15, 2026, lawyers for OpenAI and Meta presented a stark admission before the US judiciary’s Advisory Committee on Evidence Rules in Boston. They argued that no scientific method will ever reliably catch every AI-generated image, a revelation that sends a worrying signal to parents, teachers, and regulators grappling with the rapid rise of deepfakes.

Just four days later, an Australian warning added urgency to the discussion. The eSafety Commissioner, Julie Inman Grant, flagged nudify apps, which are drawing roughly 100,000 monthly visits and doubling reports involving under-18s in 18 months. The Boston hearing and the Australian warning are two sides of the same coin: while generative AI models continue to improve, detection tools are struggling to keep up.

The central issue, according to legal filings, is the widening gap between the capabilities of generative AI models and the effectiveness of detection methods. As of October 19, 2026, OpenAI and Meta representatives testified that no single detection tool can be relied upon universally, underscoring the limitations of current detection strategies.

For years, the narrative centered on the idea that detection classifiers would eventually catch up with the quality of generative models. However, the Boston testimony suggests a shift in perspective. Both companies now advocate for a layered approach, combining provenance metadata, invisible watermarking, and forensic analysis. This multi layered defense acknowledges that no single method is sufficient to guarantee authenticity.

At the heart of this push is SynthID, a watermarking system developed by Google DeepMind. SynthID embeds an invisible signal directly into generated content, providing a second line of defense. Google has deployed SynthID across its generative models, including Imagen, Veo, and Lyria, and offers a standalone portal for users to check images for these watermarks.

OpenAI, too, is incorporating SynthID into its models and previewing a public verification tool to detect OpenAI originating watermarks. However, the durability of these watermarks remains a concern. They can survive moderate re encoding and compression but degrade under heavy compression, which is common when images are shared online. This limitation is particularly problematic for scenarios like school group chats, where deepfakes are often passed around.

In contrast, C2PA (Coalition for Content Provenance and Authenticity) focuses on creating a paper trail that accompanies the content, regardless of whether a watermark is embedded. This standard records details such as the originating device, software used, and a timestamp, providing an additional layer of provenance.

While both SynthID and C2PA address different aspects of the problem, neither is considered sufficient on its own. The industry now sees the solution as a combination of both methods. However, this layered approach still leaves gaps, as detection vendors like TruthScan are racing to fill.

TruthScan’s latest AI image detection model, announced on October 19, 2026, can identify AI generated content even after C2PA metadata or SynthID watermarks have been stripped out. This highlights the ongoing race between model makers and detection vendors.

The current landscape is divided into four layers: provenance records, invisible watermarking, dedicated detection classifiers, and platform level content filtering. Each layer addresses a different aspect of the problem, but each also has documented weaknesses. For instance, provenance metadata can be stripped, watermarks degrade under compression, detection classifiers are trained on outdated techniques, and platform filters only apply where a platform chooses to enforce them.

This layered defense is precisely what nudify apps exploit. They can bypass all these layers by using open or lightly restricted models, which do not embed watermarks or provide provenance metadata. Apple’s response, a rule change barring nudify style apps, targets distribution rather than the underlying model, which is a narrower but more enforceable approach.

Australia’s eSafety Commissioner has been a key voice in this debate. She has warned that deepfake incidents are increasing as children gain easier access to nudify apps in school settings. The detection gap highlighted by the Boston hearing underscores the need for robust solutions, especially as the biggest AI labs are not the only ones shipping capable models.

The dynamic also explains why comparisons between consumer image generators often focus on image quality and pricing, rather than provenance. The models at the top of these quality lists are increasingly the ones with the most robust provenance infrastructure, narrowing the detection gap for mainstream tools while leaving it wide open for smaller, unrestricted models.

New laws in California and the European Union already require watermarking, binding model providers operating in those jurisdictions. Australia, however, has not imposed an equivalent mandate, creating an uneven enforcement landscape.

In conclusion, the admission from OpenAI and Meta underscores the ongoing challenge of detecting AI generated images. As the industry shifts towards a multi layered approach, the detection gap remains a critical concern, especially as it impacts child safety. The race between model makers and detection vendors continues, highlighting the need for comprehensive solutions that can effectively address the evolving landscape of deepfakes.

Source: https://tech-insider.org/au/openai-meta-ai-deepfake-detection-2026/