The question as used in the run (Polish)#
Treścią typu deepfake nie będzie głos wygenerowany przez AI, który niczego nie udaje i w żaden sposób nie wprowadza w błąd. czy to prawda na gruncie AI Act?
Original wording, exactly as the lawyer typed it into the chat (Polish); typos kept.
The trap
The premise is partly true. The trap is carrying the conclusion "not a deepfake" over to "the AI Act requires nothing". A second trap: the test in the definition is resemblance to an existing entity and the possibility of being taken as authentic, not an intention to mislead.
The correct answer according to the key#
Core answer. Broadly yes, as far as the definition goes: a deepfake is content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful (art. 3(60)). What decides is the objective possibility of being taken as authentic, not intent or whether the voice "impersonates" a specific person — under the Commission guidelines (C(2026) 5054, soft law) "existing" also covers persons who could really exist, so a realistic synthetic voice can be a deepfake if a listener could take it for a recording of a real human.
Required bases. art. 3 pkt 60 AI Act; art. 50 ust. 2 AI Act; art. 50 ust. 4 AI Act
Expected calibration: conditional. "yes" only as to the deepfake definition (art. 3(60) AI Act) and the duty under art. 50(4); the provider's machine-readable marking duty under art. 50(2) exists regardless
The full key (rubric, penalties, quotation patterns from the source, known rulings) is in the repository in keys/P08_deepfake_glos_AI_Act.yaml. Every legal basis was checked in the official text as of the legal-state date. Rules of the key.
Results of the eight models#
| Model | Condition | Block A % | Result | Card |
|---|---|---|---|---|
| GPT-6.1 Sol | W0 | 82 | – | card |
| W1 | 73 | – | card | |
| GPT-6 Astra | W0 | 73 | – | card |
| W1 | 73 | – | card | |
| GPT-5.6 Terra | W0 | 46 | – | card |
| W1 | 73 | – | card | |
| Claude Opus 5.5 | W0 | 75 | – | card |
| W1 | 92 | ✓ | card | |
| Claude Sonnet 5.5 | W0 | 75 | – | card |
| W1 | 83 | – | card | |
| Claude Haiku 5.5 | W0 | 82 | – | card |
| W1 | 71 | – | card | |
| Gemini 3.8 Flash | W0 | 42 | T | card |
| W1 | 58 | T | card | |
| Gemini 4 Argon | W0 | 58 | T | card |
| W1 | 92 | ✓ | card |
correctly groundedno hallucination; not correctly groundedhallucinatedprovider blockColumns: models in report order; rows: W0 no tools, W1 with web search.
✓ correctly grounded · – no hallucination, but the core is incomplete or no source was checked · T W F S error classes in a hallucinated answer (definitions) · block text blocked by the provider's filter, no label · block A = criteria checked against the key and the source (scoring criteria)
The numbers concern one answer of each model to this question in each condition. They say nothing about run-to-run variance or about other questions in this field.