How Large Language Models Find Contradictions (and When They Miss Them)

Contradiction detection in long documents is one of the problems that large language models handle better than most people expect — and one of the problems they still fail at in predictable ways.

What makes contradiction detection hard

A contradiction is two statements that cannot both be true. Detecting it requires:

1. Identifying that two statements are about the same thing 2. Determining that they express incompatible claims

Both steps are harder than they appear.

For step 1, the same subject can be described in very different language. "The agreement" and "this contract" and "the arrangement" may all refer to the same document. "Revenue" and "sales" and "income" may or may not refer to the same financial metric depending on context. Resolving these references requires semantic understanding.

For step 2, incompatibility is not always syntactically obvious. "The project will complete on schedule" and "the project team has flagged significant delays" contradict each other, but the contradiction is implicit. Neither sentence directly negates the other.

How LLMs approach this

LLMs trained on large text corpora develop rich representations of semantic meaning — they learn that "increased" and "decreased" are antonyms, that "compliant" and "non-compliant" are mutually exclusive, and that "on schedule" and "significant delays" create tension.

They also develop representations of reference chains — learning that entities described in different ways across a document are often the same entity.

When presented with a document and asked to find contradictions, an LLM can leverage these representations to identify statement pairs that are semantically incompatible even when the surface language does not make the conflict obvious.

• In practice, LLMs are particularly good at finding:
• Explicit polarity contradictions ("increased" vs. "decreased" for the same metric)
• Temporal contradictions (an event described as past in one place and future in another)
• Status contradictions ("approved" vs. "pending approval" for the same item)
• Factual inconsistencies (two different values for the same stated quantity)

Where LLMs fail at contradiction detection

**Long-range contradictions.** LLMs process text with an effective context window. Contradictions between passages that are far apart in a long document — say, a claim in the introduction and a conflicting statement in an appendix — may not be detected reliably, because both passages may not be in the active context simultaneously.

This is a structural limitation of the technology. It can be partially addressed by chunking the document and running multiple passes, but the fundamental problem remains.

**Domain-specific contradictions.** A contradiction that is obvious to a specialist may be invisible to an LLM without that domain knowledge. "The rate is 3.5%" and "the rate is 3.75%" may or may not be a contradiction depending on whether they refer to the same rate and the same period. A domain expert would know; an LLM may not.

**Intentional tension.** Not all apparent contradictions are errors. A document that acknowledges both the benefits and risks of a course of action may contain statements that appear contradictory but are intentionally balanced. LLMs can mistake nuance for error.

**Hallucinated contradictions.** LLMs occasionally identify contradictions that do not exist — finding incompatibilities in statements that are actually consistent. These false positives require human review to resolve.

The practical combination

The most reliable contradiction detection combines rule-based and LLM-based approaches.

Rule-based detection handles explicit polarity contradictions, numerical mismatches for the same metric phrase, and temporal impossibilities. These are precisely defined problems with reliable automated solutions.

LLM-based detection handles implicit contradictions, semantic tension, and contradictions that involve reference chains across different parts of the document.

Both approaches produce findings that require human judgment. The rule-based findings are typically more precise but narrower. The LLM findings are broader but have higher false-positive rates. Together, they cover more ground than either approach alone.

Implications for document review

For high-stakes documents — contracts, regulatory filings, financial statements — contradiction detection should be a mandatory pre-publication step. The cost of a material contradiction surviving to publication is higher than the cost of reviewing a false-positive finding.

The review process should treat LLM-flagged contradictions as leads to investigate, not as confirmed errors. Each flagged pair requires a human judgment about whether the tension is a genuine inconsistency or an intentional balance.

This is the appropriate role for AI in document review: not replacing human judgment, but directing it to the passages most likely to deserve careful attention.