IC-1177LLM-generated misinformation is harder for LLM detectors to detect than human-written misinformation with the same semantics

Canyu Chen, Kai Shu

SourceCan LLM-Generated Misinformation Be Detected?

Eight LLM detector configurations (ChatGPT-3.5, GPT-4, Llama2-7B-chat, Llama2-13B-chat, each with and without zero-shot CoT) were evaluated on PolitiFact, GossipCop, and COAID. Detection accuracy on LLM-generated misinformation is consistently lower than on human-written misinformation. For example, GPT-4 (no CoT) on PolitiFact drops from 48.6% (human-written) to 22.0% (open-ended generation); Llama2-7B-chat (CoT) drops from 47.4% to 27.8% on rewriting generation (19.6% drop). Results hold across all six open-source generators (Llama2-7B/13B/70B, Vicuna-7B/13B/33B) and all three datasets, with most p-values below 0.05.

Evidence
correlational
Key metric
GPT-4 no-CoT PolitiFact: human-written 48.6%, paraphrase 41.7%, rewriting 34.8%, open-ended 22.0%; Llama2-7B-chat CoT PolitiFact: human-written 47.4%, rewriting 27.8% (drop 19.6%); GPT-4 CoT COAID: human-written 81.0%, open-ended 52.7%
Caveat
Zero-shot setting only; average over three runs for API-based detectors due to output variance; GossipCop and COAID results show less consistent drops than PolitiFact for some detector-generator pairs.
Model
ChatGPT, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, Llama 2 / Llama 2 base, Vicuna
Concepts
Failure mode
Datasets
GossipCop [eval], COAID [eval]
Methods
Zero-shot Chain-of-Thought / Wei et al. 2022 (Chain of Thought) / Zero-shot chain-of-thought prompting [primary], Paired t-test [validation]
Related findings
IC-1175, IC-1176
Extraction
automatic-extraction