SagaAI
What it is
AI-generated web text is polluting model-training data while text watermark detection remains fragile.
Overview
Researchers found AI-generated tokens reached 31.1% of FineWeb-filtered web tokens in August 2026. Their tests showed synthetic tokens can eventually hurt models’ ability to predict held-out human text. OpenAI is limiting access to an EU text detector because short passages, edits and translation can remove watermark signals.
Why it matters
The problem matters for model developers, researchers and regulators because contaminated data and weak detection can undermine training quality and authenticity rules.
Key details
• The AI-token share was 10% in June 2024 and 27.5% in June 2026. • OpenAI’s textGrain is planned for ChatGPT and Codex users in the EU. • OpenAI targeted a 1% false-positive rate in internal tests.
How it could play out
Tilt can alert you when its scenarios shift, or send you a recommendation when one crosses a threshold.
Stories in this saga
Portfolio
Portfolio
How it could play out
Saga, how it could play out
SagaAI-generated web text and model-training data pollution
OpenAI treats the low detection rate as a technical compliance failure rather than a legal confrontation. The most likely path is a short remediation cycle with updated watermarking, provenance metadata, and clearer AI Act documentation for EU users.
Regulators decide the failed edit-resistance test exposes a meaningful gap in AI Act compliance. This creates a sector-wide warning that watermarking claims will need independent verification rather than vendor self-attestation.
The weak watermark result accelerates a shift away from hidden text watermarks toward cryptographic provenance, signed metadata, and platform-level disclosure. The market reaction is mixed: lower confidence in pure watermarking, but stronger demand for enterprise verification tooling.
The article becomes part of the broader debate over AI disclosure but does not trigger a major regulatory or market event. Customers treat the 17% result as another sign that watermarking is imperfect, not as a reason to abandon ChatGPT or enterprise AI tools.
Portfolio
Portfolio
The saga on the S&P 500
