Contracts from employers and landlords often include unfair or unclear terms that can harm employees and tenants, leading to extra costs or limits on their rights.
To address this, researchers at New York University built a tool called ContractNerd that uses large language models (LLMs) to review contracts. The tool checks clauses and sorts them into categories: missing parts that should be included, unenforceable ones that break laws, legally valid ones, and legal but risky ones rated as high, medium, or low risk. This helps both those writing contracts and those signing them avoid disputes by spotting issues early.
In a press release, the researchers explains that most people lack legal training to fully grasp contracts, so ContractNerd flags biased, illegal, or vague clauses and suggests fixes. “ContractNerd is an AI system that analyzes contracts for clauses that are missing, are extremely biased, are often illegal, or are ambiguous and will suggest improvements to them,” they say.
The tool focuses on leases and employment contracts in New York City and Chicago, pulling information from legal sources like Thomson Reuters Westlaw, a database of court cases; Justia, a site with standard agreement language; Agile Legal, a collection of common clauses; and state rules.
How ContractNerd performs compared to others
To test ContractNerd, the researchers compared it to other AI contract analyzers and found it best at predicting which clauses courts would rule unenforceable. In another test, non-experts reviewed outputs from ContractNerd and the next-best tool, goHeather, on factors like relevance - how well it matches the clause's meaning - accuracy in legal facts, and completeness in covering all key points. The reviewers, unaware of which tool was which, favored ContractNerd overall. A contracts law expert from New York University School of Law also assessed both, noting ContractNerd was more detailed while goHeather was simpler to read.
The researchers have described ContractNerd and the methods and results of this study in a paper (not available at the time of writing) published in MDPI Electronics.