What Happened to AI Model Watermarking and "Enshittification"?

The twin concepts of AI model watermarking and "enshittification" represent a critical juncture in the evolution of artificial intelligence, reflecting both the industry's attempts to build trust and the inherent pressures that can lead to platform degradation. AI model watermarking, at its core, involves embedding imperceptible digital markers into AI-generated content—be it text, images, audio, or video—to verify its origin and authenticity. This technology has become paramount in the fight against deepfakes and misinformation, which have proliferated as generative AI capabilities have advanced. By 2026, major players like Google with its SynthID and the broader industry standard C2PA (Coalition for Content Provenance and Authenticity) are widely implementing these solutions, with SynthID having watermarked over 100 billion images and videos and 60,000 years of audio content by May 2026.

The push for watermarking has been significantly accelerated by regulatory mandates. The European Union's AI Act, with its Article 50 transparency rules, became applicable from August 2, 2026, requiring providers of AI systems to ensure their outputs are machine-readable and detectable as AI-generated. Similarly, California's AI Transparency Act (SB 942), also operative from August 2, 2026, mandates invisible watermarking for AI-generated image, video, and audio content from covered providers. These regulations aim to establish a chain of custody for digital media, making it harder for malicious actors to spread deceptive content. However, challenges persist, particularly with text watermarking, which remains more fragile and susceptible to removal through paraphrasing or translation compared to visual or audio watermarks.

In parallel, the concept of "enshittification," coined by journalist Cory Doctorow in 2023, describes the process by which online platforms progressively degrade the quality of their services to extract more value from users and businesses. Doctorow outlines a three-stage cycle: first, platforms are good for users (subsidized by investor funds); second, they exploit users to benefit businesses (e.g., through ads and data mining); and finally, they exploit businesses too, leading to a universally degraded experience due to high switching costs. By 2026, this framework is increasingly being applied to AI services, with analysts noting that AI platforms appear to be transitioning from an initial phase of generous, often free, access to a stage where commercial interests lead to a decline in user experience. This includes the introduction of ads, pay-to-play features, and a perceived reduction in the quality or neutrality of AI outputs, as models become "safer" but also more prone to lecturing or disagreeing without reason.

The consequences of "enshittification" in the AI realm are profound, threatening to erode trust in AI-generated information and diminish the utility of AI tools. As AI content potentially becomes saturated with commercial biases or reduced quality, the very purpose of watermarking—to establish authenticity—becomes even more critical. The current status in mid-2026 shows a race between technological solutions for provenance and the economic forces driving platform degradation. While major tech companies like Meta, Google, and Anthropic are actively implementing watermarking and content labeling policies, often in response to regulatory pressure, the debate continues on how to ensure these measures are robust, interoperable, and genuinely effective against both deliberate manipulation and the subtle erosion of quality inherent in "enshittification." The long-term viability of AI as a trusted resource hinges on successfully navigating these intertwined challenges.