Why Legal AI Needs Better Architecture, Not Better Guesswork
When AI fabricates legal authority — citations to cases that were never decided, quotations from opinions never written, constitutional provisions that don’t exist — we’re not looking at a “bad prompt.” We’re looking at an architecture problem. And the problem isn’t going away.
In early 2025, Kyle Kjoller, a welder in Nevada County, California, was held without bail. Prosecutors opposed his release in a brief his lawyers say carried generative-AI fingerprints: misread law, fabricated quotations, misstated constitutional provisions. A federal judge in Oregon sanctioned two lawyers $110,000 in December 2025 after they submitted 23 fabricated citations and eight invented quotations. And in April 2026, Sullivan & Cromwell, the elite Wall Street firm, apologized for filing a brief with AI hallucinations.
This isn’t fringe risk. It’s a category problem that most legal AI tools are exposed to.
The Wrapper Wave
In its 2024 Legal Technology Survey Report, the ABA found 30% of attorneys in private practice say their firms are using AI tools. Many are working on top of large language models wrapped with legal-friendly UX and guardrails. Harvey, for example, is used by more than half of the top U.S. law firms and is explicitly an LLM-powered generative platform.
These tools can be useful. But most share the same underlying risk because they’re still generative-first. Even when they add retrieval or knowledge grounding, the model is typically allowed to “free-generate” around what it retrieved, which means hallucinations remain possible. That’s not a vendor problem. It’s an architectural one. Generative AI predicts. Closed loop AI surfaces what’s in the record.
Generative vs. Closed-loop AI
Generative models like ChatGPT predict what should come next based on patterns. Asked to draft a brief, they can produce reasoning that looks right while quietly drifting into fiction.
Closed-loop AI works differently. The model is constrained to the verified source material you provide—transcripts, exhibits, discovery files, case records—and outputs stay tethered to that record. It can summarize, extract, surface contradictions, and cite the source. It cannot introduce authority that isn’t supported by a record.
Closed-loop AI isn’t perfect. But it materially reduces hallucination risk because the system isn’t allowed to wander.
The Case for Closed-Loop Investigative Intelligence
In deposition or evidence review, generative systems can produce a clean, confident story of what a witness “probably meant”—even when the record doesn’t support it. Closed-loop systems behave differently. When support is thin, they surface uncertainty. When testimony conflicts, they flag it. When evidence exists, they point you back to it.
Greening Law Group used Rev’s Multi-File Insights to surface contradictions across body-cam and interview footage in real time, building cross-examination from the record itself, not from a chatbot’s best guess. The firm credits that closed-loop workflow on Rev’s Investigative Intelligent platform with helping secure dramatically reduced sentences for their client.
That’s what better architecture changes in your day: faster review, defensible output, and a system that fails visibly when it doesn’t have support — not silently, in the middle of a filing.
Closed-loop architecture is what makes investigative intelligence possible—software that doesn’t just retrieve evidence but surfaces what matters across it.
Three questions to ask before adopting any legal AI tool
1. Verifiability—Can every output be traced to a source in your record?
2. Auditability—Is there a clear trail from the result back to the underlying file?
3. Defensibility—Could you show your work to a judge?
If a tool can’t answer those, you’re not adopting AI. You’re adopting risk. Learn more about how Rev can help your practice. ABA Members enjoy special pricing.
Write a letter to the editor, share a story tip or update, or report an error.This content is advertising.
