That has been proven not to work and a variation of the "make no mistakes" meme. What about the lawyer doing the actual work, he/she is being paid for?
?! It does work but you need separate (sub)agents (and potentially loop it until the reviewer agent finds nothing). You also need to set it up a bit so all citations mentioned are easy to identify. Ideally this is also done programmatically, ie parse text find citations and ensure they say what you think they say (llm part).
A second AI agent could have verified the sources and citations before submission. Easy extra layer of protection against hallucinations.
That has been proven not to work and a variation of the "make no mistakes" meme. What about the lawyer doing the actual work, he/she is being paid for?
?! It does work but you need separate (sub)agents (and potentially loop it until the reviewer agent finds nothing). You also need to set it up a bit so all citations mentioned are easy to identify. Ideally this is also done programmatically, ie parse text find citations and ensure they say what you think they say (llm part).
Sure, and there are existing legal research tools which would help here (LexisNexis, amongst others, will happily sell you a suitable service).
Constraining an agent by forcing it to pass some form of external validation is one of the best ways we have of minimising hallucinations.
But is a lawyer who is this lazy and this dishonest ever likely to bother doing that?
It does not work because the errors compound and also are most times correlated.
"Nine Judges, Two Effective Votes: Correlated Errors Undermine LLM Evaluation Panels" - https://arxiv.org/abs/2605.29800
"From Spark to Fire: Modeling and Mitigating Error Cascades in LLM-Based Multi-Agent Collaboration" - https://arxiv.org/abs/2605.29800
"Large Language Models Cannot Self-Correct Reasoning Yet" - https://arxiv.org/abs/2310.01798
"Correlated Errors in Large Language Models" - https://proceedings.mlr.press/v267/kim25e.html
"Partially Correlated Verifier Cascades in LLM Harnesses: Concave Log-Odds, Polynomial Reliability, and Blind-Spot Ceilings" - https://arxiv.org/abs/2607.13918
"Cross-Context Verification: Hierarchical Detection of Benchmark Contamination through Session-Isolated Analysis" - https://arxiv.org/abs/2603.21454
"AI Hallucination Cases Database" - https://www.damiencharlotin.com/hallucinations/