Сравнение hallucination и confabulation. Разберите, как оба термина описывают plausible but unsupported or fabricated AI output, и почему hallucination часто используется как broader operational term.
Краткий вердикт: Используйте hallucination для общего AI safety term про false or unsupported generated outputs; используйте confabulation, когда акцент на confident fabrication presented as plausible.
Hallucination describes AI-generated output that appears plausible or confident but is false, unsupported, misleading, or fabricated.
Контекст: Наиболее уместно при describing factual reliability failures in generative AI outputs.
Confabulation describes generation of false, unsupported, or fabricated information by an AI system while presenting it as plausible.
Контекст: Наиболее уместно, когда акцент на confident fabrication or unsupported plausible generation.
| Аспект | Hallucination | Confabulation |
|---|---|---|
| Определение | Hallucination — plausible or confident AI-generated output that is false, unsupported, misleading or fabricated. | Confabulation — generation of false, unsupported or fabricated information while presenting it as plausible. |
| Практическое различие | Often broad operational label for generative AI factual failures. | Often emphasizes fabricated or plausible-story quality of inaccurate output. |
| Типичный сценарий | Evaluation, grounding, verification and human review discussions. | Confident but inaccurate model outputs or fabricated explanations. |
| Распространённая ошибка | Assuming confident answer reliable because it sounds coherent. | Using confabulation as if it always described separate technical mechanism. |
| Значение для governance | Requires grounding, constraints, verification, evaluation and human review. | Requires similar controls where plausible fabrication can mislead users or reviewers. |
На практике governance response важнее label. Teams should record examples, source checks, mitigations and residual risk instead of debating terminology alone.
Используйте Hallucination для AI-generated outputs, которые false, unsupported, misleading или fabricated. Это common term for reliability failures: incorrect facts, invented sources, false citations or ungrounded reasoning.
Используйте Confabulation, когда акцент на generation of plausible but false or unsupported information. Полезно, когда risk is confident fabrication rather than simple classification or retrieval error.
Для NIST AI RMF и ISO 42001 controls hallucination/confabulation должны связываться с evaluation, grounding, verification and human review evidence. В high-risk contexts plausible false outputs may create deception, safety or accountability risks.
Они близки. Hallucination — more common AI term for plausible but false or unsupported generated output; confabulation often emphasizes confident fabrication.
Through grounding, constraints, verification, evaluation and human review. Mix depends on risk level and use case.
False citations make unsupported output appear evidence-based. Это common hallucination pattern and should be tested and logged.
No recently viewed comparisons yet.