Research map / Large language models / AI and society

Watermarking Without Standards Is Not AI Governance

NeurIPS 2026 · Large language models › AI and society · approach: Human-centered trust and agency · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Position paper arguing that AI watermarking without enforceable technical standards, independent audits, and regulatory enforcement becomes symbolic compliance, supported by a small SynthID-Text and AudioSeal robustness check and a three-layer framework.

Problem

Watermarking is cited in AI governance frameworks (EU AI Act, US EO 14110, California SB 942, China's Deep Synthesis Provisions) but current schemes are brittle, hard to audit, proprietary, and lack defined robustness thresholds, so they may provide only symbolic compliance.

Approach

Reframes watermarking as a governance problem and proposes a three-layer framework (technical requirements with shared low-FPR benchmarks, black-box third-party audit infrastructure, and policy enforcement with certification and penalties) to make watermark claims verifiable.

Key results

• SynthID-Text detected 4 of 5 original outputs (threshold 0.52) and 0 of 10 paraphrased or summarized outputs in the small test.
• AudioSeal detected all 5 original clips, but only 1 of 10 speed-modified or noise-added clips remained above threshold 0.5 (only sample 1 with noise).
• Authors report that 9 of 10 original outputs were detected while only 1 of 20 transformed outputs remained detectable across both modalities.

Surprising findings

• Benign edits such as 1.25x playback speed largely eliminated AudioSeal detection across all samples.

Limitations

• Robustness evaluation is minimal: only five samples per modality and a handful of transformations.
• Uses open proxies (Gemma-2-9b-it, AudioGen) instead of the deployed SynthID-Text and AudioSeal systems in production.
• Framework is proposed conceptually without implementation or empirical validation of the audit and enforcement layers.

Experimental setup

Benchmarks: SynthID-Text detection on 5 Gemma-2-9b-it outputs AudioSeal detection on 5 AudioGen clips

Metrics: watermark detection score vs. threshold

Baselines: SynthID-Text AudioSeal

Builds on: SynthID-Text AudioSeal MarkLLM

Closest papers by meaning

• No Free Lunch in LLM Watermarking: Trade-offs in Watermarking Design Choices (NeurIPS 2024 · similarity 0.85)
• Unbiased Watermark for Large Language Models (ICLR 2024 · similarity 0.84)
• WMCopier: Forging Invisible Watermarks on Arbitrary Images (NeurIPS 2025 · similarity 0.83)
• Position: Generative AI Regulation Can Learn from Social Media Regulation (ICML 2025 · similarity 0.83)
• Watermarking Game-Playing Agents in Perfect-Information Extensive-Form Games (NeurIPS 2026 · similarity 0.82)
• Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data? (AAAI 2025 · similarity 0.81)