1. Generative AI can be wrong
Large language models generate text by predicting likely sequences of words. They do not know the law; they model the patterns of legal writing. That is why a model can produce a citation with a plausible case name, a correct-looking reporter volume, and a confident parenthetical, all describing a case that does not exist.
This is not a defect unique to any one product, and it is not solved by marketing language. Independent research from Stanford has documented hallucinations in specialist legal AI tools from the largest vendors in the industry, and the researchers specifically criticized categorical claims about eliminating hallucinations. Any vendor that tells you its output cannot be wrong is telling you something the evidence does not support.
Irys takes the opposite position. The platform is designed on the assumption that generated text can contain errors, which is why verification is built into the workflow rather than claimed away.
2. Attorney review is not optional
A lawyer's duty to verify the authorities they cite is non-delegable. It existed before AI, it applies to work produced with AI, and no software feature discharges it. Courts across the country have made this explicit, and bar guidance in every jurisdiction that has addressed the question says the same thing: the lawyer who signs the filing is responsible for what is in it.
That principle shapes how Irys presents its output. Research results link back to source opinions so they can be opened and read. Cite Check returns flags for review, not a certificate of correctness. Drafts arrive as tracked changes so every edit is visible and reviewable. The product is built to make the lawyer's review faster and better informed, never to replace it.
Nothing produced by Irys should be treated as filing-ready because it is well formatted, contains citations, or has passed through a verification feature. The responsible legal professional performs the review appropriate to the task before the work is relied on. Our Terms of Service say this in binding language, and we repeat it here in plain language.
3. How Irys layers its safeguards
Reducing error risk is an architecture problem, not a prompt. Irys approaches it in layers, each addressing a different failure mode:
Grounded research. Research runs against retrieved legal sources rather than relying on what a model remembers. The deeper research modes retrieve case law and other authority first, then synthesize from what was retrieved.
Source-linked results. Research output links back to the underlying opinions and documents, so cited authority can be opened and read rather than taken on faith.
Cite Check. Cite Check checks identified citations in a document against legal-source databases, flags citations it cannot match, and surfaces available treatment signals such as negative history. Flags are presented for attorney review.
Attorney review. The review itself is the layer the platform cannot perform. Irys is built to make it efficient: sources one click away, flags in one place, edits tracked.
Audit trail. Actions are logged, so a firm can see what was run, when, and in which mode.
The intended sequence is simple: research produces source-linked authority, Cite Check flags what needs attention, the attorney reviews and resolves the flags, and only then does work go out the door. Each layer reduces risk. No layer, alone or combined, eliminates it.
4. Using the right depth for the task
Irys offers different working depths because legal tasks differ in stakes. Using the right one matters.
Quick is for fast answers, short drafts, and orientation. It is deliberately lightweight: it does not retrieve and verify authority, and its answers are not cite-backed. Quick output is not a substitute for research and should never be the basis for authority cited in a filing.
Standard (Deep Research) produces memo-quality reports in one comprehensive pass, with aggregated citations. Because it gathers citations at the end rather than verifying each one against source as it writes, run Cite Check on the report before relying on its citations. Treat the two as a paired workflow.
Thorough is built for substantive legal research: it investigates step by step, checks case-law citations against source as it works, and threads them inline. When a question turns on case law and the answer has to hold up in front of a court, start here.
The rule of thumb we give every user: the closer the output is to a court or a client, the deeper the mode, and always Cite Check before filing. The full breakdown is in the modes documentation.
5. What Irys does not claim
Trust requires being precise about limits, so here are ours, stated as plainly as we can make them:
We do not claim Irys output is always accurate, and we do not claim any mode or feature makes output guaranteed correct.
We do not claim Cite Check verifies every citation. It checks identified citations against available legal-source data and flags what it cannot match or what shows treatment concerns. A flag review by the attorney is part of the design, not an afterthought.
We do not claim Irys is connected to, integrated with, or a substitute for any third-party citator or research subscription your firm may separately license. Irys is a standalone platform with its own research and citation-checking tools.
And we do not claim that using Irys satisfies a lawyer's professional obligations. It is a tool that supports the diligence the profession already requires.
6. What strong firm practice looks like
The firms that get the most from legal AI treat it the way they treat any capable junior contributor: real output, real supervision. In practice that means a written AI policy that says which tools are approved and for what; disclosure inside the team, so supervising attorneys know when AI was used in work they are reviewing; verification requirements that are explicit rather than assumed; and training, so every user knows what each tool does and does not do.
Irys supports this posture with audit logging, mode visibility, and documentation written to be read. If your firm is drafting or updating its AI policy, our team is glad to walk through how other firms have structured theirs. Talk to us.
