Under a worst-case formal model, attributing LLM outputs to a specific model is provably impossible except in the narrow case of finitely many deterministic models.
Proc Natl Acad Sci USA 111(4):1298– 1303
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Can adversarial attacks by large language models be attributed?
Under a worst-case formal model, attributing LLM outputs to a specific model is provably impossible except in the narrow case of finitely many deterministic models.