AI decoders missed 168 verifiable messages in a controlled audit
Source check: Source checked. Not independent scientific confirmation.
Finding: In an audit of 7,200 language-model outputs, researchers found 168 verifiable hidden messages that the original model decoders failed to recover. The result shows that decoder failure alone can miss successful encoding in this controlled test.
Technology: Gemini, Claude, GPT language-model encoders and decoders
Family Claude: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Learned language models produced and attempted to recover messages. An independent deterministic audit checked recoverability to evaluate what the model decoders missed.
- Test and result
- The researchers compared original model-decoder results with an independent deterministic verification of the output.
- Limitation / still unknown
- The result applies to the tested synthetic messages and verification criteria. It does not estimate prevalence in real communications. Of the 168 cases, 58 were recovered with exploratory procedures added after inspecting the corpus. Verification also allowed normalized letter-sequence matches, rather than requiring byte-exact messages, and does not establish intentional encoding.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.30516 · https://arxiv.org/html/2610.30516v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: arXiv v1 aggregate encoding/decoding contingency results and validity distinctions.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.