Having AI review a contract: what it spots, what it misses
An AI assistant can flag risky clauses in a contract, but it also misses some and makes things up. Measured limits and habits to adopt.
On this page
In 30 seconds
What it is
Researchers have measured how well language models identify legal risks in commercial contracts, clause by clause. Their benchmark, ContractEval, tested 4 proprietary models and 15 open-source models[1]. They conclude that most models perform at a level comparable to junior legal assistants[1].
This article explains to employees what this research shows, what an AI assistant may omit or invent, and which habits to adopt before handing it a document.
Straight answers
Does an AI assistant find every clause in a contract?
The ContractEval research does not allow that claim. The authors used the Contract Understanding Atticus Dataset (CUAD)[1]. They observe that open-source models generate no related clause answers more often, even when relevant clauses are present[1]. They see this as a possible sign of laziness in the model's reasoning or of low confidence in extracting the relevant content[1].
Proprietary models perform better than open-source models on answer accuracy and efficiency, although some open-source models remain competitive on certain points[1].
Does a reasoning mode make the analysis more reliable?
Not necessarily. According to ContractEval, the reasoning mode (thinking) improves answer efficiency but reduces accuracy, probably because it overcomplicates simpler tasks[1].
Can an AI invent legal texts or court decisions?
Yes. OWASP gives the example of ChatGPT, which fabricated fake legal cases, with significant consequences in court[2]. OWASP calls this a hallucination: the model generates content that seems accurate but is made up[2]. A technical report by Alex Dantart analyzes this problem of hallucinations in language models applied to law, with its causes and its manifestations[4].
Why does a wrong answer sound so sure of itself?
OWASP describes a model's misinformation as false or misleading information that appears credible[2]. The NCSC (UK) wrote in 2023 that models have the flaw of sounding convincing whether or not they are right[3]. It added that their answers may be only partly correct, especially on specialized subjects[3].
Can I paste a contract into a public assistant?
The NCSC recommends not including sensitive information in requests to public models, and not submitting requests that would cause problems if they became public[3]. It also describes the practice of copying large amounts of text, or even whole documents, into the question to give context, and does not recommend it with sensitive information on public models[3].
What happens to what I type into the assistant?
According to the NCSC, the request is visible to the organization that provides the model, is stored, and will almost certainly be used at some point to develop the service or the model[3]. At the time of writing, the NCSC stated that a model did not automatically add information from requests to its model for other users[3]. It also notes that requests stored online may be hacked, leaked or, more likely, made public by accident[3].
How it works
According to the NCSC, a language model has been trained on a large amount of text and derives from it a model of probabilities based on the relationships between words[3]. OWASP explains that hallucinations occur because the model fills gaps in its training data with statistical patterns, without understanding the content[2]. It can then produce answers that seem right but have no basis[2].
A common countermeasure is Retrieval-Augmented Generation (RAG). OWASP recommends it: it consists of retrieving relevant, verified information from trusted external databases while the answer is being generated, which helps limit hallucinations[2]. A literature review on RAG presents it as a combination of language models and information retrieval systems, meant to strengthen factual grounding and reduce hallucinations[6]. The same review lists the quality of information retrieval and privacy questions among the persistent difficulties[6]. Alex Dantart's report also examines the effectiveness of RAG and sets out its limits[4].
What you do
- Before pasting a text, check what it contains: the NCSC recommends not including sensitive information in requests to a public model[3].
- Read the tool's terms of use and privacy policy before asking it sensitive questions: the NCSC states that they should be well understood[3].
- For a tool provided by a cloud provider, the NCSC urges organizations to understand how their data is handled: do the provider's researchers or partners have access, and under what conditions can a provider employee see the requests[3]?
- Cross-check answers against trusted external sources. OWASP recommends encouraging this verification and setting up human oversight and fact-checking, especially for critical or sensitive information[2].
- Do not take a no related clause answer as proof that no clause applies: ContractEval recorded such answers while relevant clauses were present[1].
- Keep judgment in your own hands. Alex Dantart's report insists on human oversight, which it considers irreplaceable, and proposes an advisory AI, a tool that amplifies professional judgment instead of replacing it[4].
Pitfalls
- Overreliance. OWASP uses this term for trusting generated content too much without checking its accuracy, which worsens the effect of false information once it enters a decision[2].
- A beginner's level. The ContractEval authors compare most models to junior legal assistants and consider that open-source models need targeted adjustment (fine-tuning) to be accurate and efficient in high-stakes legal situations[1].
- A reasoning mode mistaken for a guarantee of quality: it reduces accuracy in the ContractEval tests[1].
- An answer that sounds right. The NCSC notes that convincing answers may be only partly correct[3].
Sources
- arXiv, “ContractEval: Benchmarking LLMs for Clause-Level Legal Risk Identification in Commercial Contracts”, 5 August 2025. https://arxiv.org/abs/2508.03080 (accessed 14 November 2026)
- OWASP, “LLM09:2025 Misinformation”, 10 April 2024. https://genai.owasp.org/llmrisk/llm092025-misinformation/ (accessed 14 November 2026)
- NCSC (UK), “ChatGPT and large language models: what's the risk?”, 14 March 2023. https://www.ncsc.gov.uk/blog-post/chatgpt-and-large-language-models-whats-the-risk (accessed 14 November 2026)
- arXiv, “Inteligencia Artificial jurídica y el desafío de la veracidad: análisis de alucinaciones, optimización de RAG y principios para una integración responsable”, 11 September 2025. https://arxiv.org/abs/2509.09467 (accessed 14 November 2026)
- NIST, “AI Risk Management Framework”, 12 July 2021. https://www.nist.gov/itl/ai-risk-management-framework (accessed 14 November 2026)
- arXiv, “A Systematic Review of Key Retrieval-Augmented Generation (RAG) Systems: Progress, Gaps, and Future Directions”, 25 July 2025. https://arxiv.org/abs/2507.18910 (accessed 14 November 2026)
Cite this article
Brain, “Having AI review a contract: what it spots, what it misses”, Cyberpedia, 2026. https:/
BibTeX
@misc{brain_cp517_en_2026,
author = {Brain},
title = {Having AI review a contract: what it spots, what it misses},
year = {2026},
howpublished = {Cyberpedia},
url = {https://startbrain.io/cyberpedia/en/articles/ai-contract-review-limits/},
note = {CC BY 4.0}
}