Hallucinations

Here's the thing about LLM "hallucinations" that every auditor needs to understand: these models aren't actually knowledge databases like your firm's technical guidance library. They're sophisticated pattern-matching systems that predict what words should come next based on statistical relationships they learned during training. Think of it like this: the LLM doesn't "know" that GAAP requires specific disclosures the way you do. Instead, it has learned that when people write about financial statements, certain words and phrases tend to appear together in predictable patterns.

When an LLM hallucinates, it's generating information that sounds authoritative and professional but is actually incorrect, misleading, or completely made up. Yet it presents this information with the same confidence it would show when giving you accurate guidance. This is like getting an audit opinion that reads perfectly but is based on completely fabricated evidence. For auditors, this phenomenon is particularly dangerous because we're trained to rely on authoritative sources and documented support for our conclusions.

Basic strategies to minimize hallucinations

• Allow the LLM to say "I don't know": Just like you'd tell a staff auditor it's better to escalate a question than guess, explicitly tell the AI it's okay to admit uncertainty. This simple permission can dramatically reduce the chances of getting confident-sounding but wrong information.
• Require a list of sources: Treat the LLM like a junior staff member submitting workpapers. Ask it to provide specific quotes and cite sources for each claim it makes. Remember though, you still need to verify that those sources actually exist and say what the AI claims they say, just like you'd review any work before signing off on it.