Who built VeriLex, what it claims and doesn't claim, known limitations, and how errors are handled.
VeriLex was built by a UC Berkeley undergraduate student pursuing a pre-law track, with a summer internship at Kirkland & Ellis. It is a free, non-commercial tool with no investors, no revenue model, and no data collection.
The motivation was straightforward: low-income individuals, immigrants, and tenants routinely receive legal documents they cannot afford to have professionally reviewed. Eviction notices, immigration forms, court summons — documents with serious consequences — sit unread or misunderstood because legal help is expensive and legal aid is underfunded. VeriLex was built to close that gap, not to replace attorneys.
Questions or concerns about VeriLex can be directed to terachen@berkeley.edu.
What it does:
What it does not do:
We believe users and referring organizations deserve an honest accounting of where VeriLex can fail. The following limitations are real and should inform how the tool is used.
AI accuracy: VeriLex uses Claude AI (Anthropic) to analyze documents. AI can misread text, misidentify applicable statutes, miscalculate urgency, and occasionally generate plausible-sounding but incorrect information. Every result includes a reminder to verify deadlines against the original document. Statute citations link directly to official sources so users can check them independently.
OCR and image quality: When a document is uploaded as a scanned PDF or photo, AI image recognition is used to read the text. Blurry, low-contrast, or handwritten documents are prone to misreads — particularly of dates and numbers. Results from scanned documents include a visible warning.
Non-English output: VeriLex's AI output in non-English languages has not been reviewed by bilingual attorneys or certified legal translators. While the underlying AI model is capable in 15 languages, legal translation is a specialized discipline and errors are possible. Users viewing results in non-English can flag translation issues directly from the results page.
No attorney supervision: No licensed attorney currently reviews VeriLex's outputs before users see them. There is no retraction mechanism for an analysis already acted upon. This is described further in the oversight section below.
Local resources: Legal aid organizations listed in results are drawn from the AI's training knowledge and may be outdated. Phone numbers are omitted when uncertain. Users should independently verify that organizations are currently operating before relying on contact information.
Jurisdiction: VeriLex handles documents from all U.S. jurisdictions but was built with California context in mind. Federal immigration documents, out-of-state matters, and multi-jurisdictional situations may receive less precise analysis.
If you believe VeriLex provided incorrect legal information — a wrong statute, a misread deadline, a mistranslated term, or any other factual error — please report it.
How to report: Email terachen@berkeley.edu with a description of the error and, if possible, the document type and language. You can also use the feedback button on any results page.
What happens: All error reports are reviewed personally. Confirmed errors are used to improve the system prompt, the translation quality, or the UI warnings that appear in similar situations. We aim to respond to all reports within 48 hours.
We do not have a mechanism to correct an analysis that a specific user has already acted upon. This is a known gap. If a time-sensitive error is reported, we will prioritize it and respond the same day.
We want to be direct about this: VeriLex does not currently have a supervising attorney. No licensed attorney reviews outputs before users see them, and no attorney has formally endorsed the tool's methodology.
We believe this limitation should be named explicitly rather than obscured by disclaimers.
What we have done to compensate:
What we are working toward: We are actively pursuing a review of VeriLex's methodology by a faculty member or clinical attorney at UC Berkeley School of Law. The goal is a written endorsement of the tool's information/advice framing — not review of individual outputs, but attestation that the approach is legally sound. We expect to have more to report on this within the coming months.
We are also open to feedback from legal aid organizations about what oversight or process changes would make VeriLex appropriate to recommend to clients.
VeriLex was built with your clients in mind. The populations most likely to receive a legal document they cannot afford to have reviewed are the same populations you serve: low-income tenants, immigrant communities, workers without employer legal benefits.
We are not asking organizations to endorse VeriLex's accuracy or take on any accountability for its outputs. We are asking to be listed as a supplementary resource — a tool that helps a client understand what their document says before they come to you, or helps them understand a document when they cannot reach you.
If you choose to list VeriLex as a resource, the following language is designed to protect your organization while clearly describing what the tool does:
You are welcome to adapt this language. The key elements to preserve are: (1) "legal information only, not legal advice," (2) no affiliation with your organization, and (3) a pointer back to your own services for advice.
For partnership inquiries, error reports, or questions about listing VeriLex as a resource, contact:
Tera Chen
UC Berkeley, pre-law
terachen@berkeley.edu — subject line: "VeriLex Partnership Inquiry"
We welcome feedback from practitioners. Hearing about real cases where VeriLex helped — or fell short — is the most direct path to improving it.
Last updated: June 2026