How it works, in plain English.
Litigators are right to be careful with AI. These answers explain the technology, what makes ArsLex different from a general chatbot, and how we protect your confidential work — without the hype.
Understanding the technology
What is an LLM (large language model)?
An LLM is an AI trained on enormous amounts of text to predict the most likely next words in a sequence. That makes it remarkably fluent — it can draft, summarize, and rephrase — but it is fundamentally a language engine, not a fact database. It doesn't "look up" the law; it generates text that statistically resembles what it was trained on. Understanding that one distinction is the key to using AI safely in legal work.
Can an LLM actually "reason"?
It does something that looks like reasoning — it can follow a chain of inference and weigh considerations — but that ability is pattern-matching learned from text, not formal logic, and it can be confidently wrong. That is precisely why ArsLex never relies on the model for anything that has to be exact. The model drafts prose; the things that must be right — which authority is cited, how it is formatted, what page it sits on — are handled by deterministic, rule-based systems, not left to the model's judgment.
Why do AI tools "hallucinate" — and why has that gotten lawyers sanctioned?
Because an LLM produces text that is plausible rather than verified, it can generate a citation that looks completely real — a convincing case name, reporter, and court — for a case that does not exist. Courts have sanctioned attorneys who filed briefs containing these fabricated citations; the widely-reported Mata v. Avianca matter is the best-known example. ArsLex is built specifically so this cannot happen with your authorities: it drafts only from documents you provide, and it never asks the model to recall or compose a citation from memory.
What is RAG (retrieval-augmented generation)?
Instead of relying on the model's training memory, the system first retrieves the relevant passages from a known, trusted source — in your case, the documents you uploaded — and the model then drafts from those passages. "Retrieval" finds the right source; "augmented generation" writes grounded in it. It is the difference between an associate answering from memory and an associate answering with the actual record open in front of them.
What does "deterministic" mean, and why does it matter for citations?
Deterministic means rule-based and repeatable: the same input always yields the same output, with no guessing. It is the opposite of how an LLM works, which is probabilistic and can give different answers to the same prompt. ArsLex produces your citation formatting deterministically — by applying citation rules to a case's actual details — so a citation comes out correctly and consistently every time, rather than being improvised by the model.
How ArsLex is different from a general chatbot
How is using ArsLex different from using ChatGPT to write a brief?
Three ways that matter. Grounding: a general chatbot writes from its own memory and will invent authority; ArsLex drafts only from the record you give it. Citations: ArsLex anchors each citation to a real document at a real page and formats it by rule, rather than letting the model compose it. Confidentiality: your privileged material is never sent to a consumer AI service or used to train a model. It is the same broad class of technology, harnessed so it can't do the dangerous things.
Can ArsLex cite a case that doesn't exist?
It cites what you give it. ArsLex works from the documents you upload, so it will not produce a citation to authority you did not provide — it does not invent cases. As with any tool, you remain responsible for reviewing the finished work before it is filed.
How do I verify what ArsLex drafts?
Every cite-anchored passage traces back to the source you provided, so you can open the original document and check it directly — verification is built into the workflow, not an afterthought. We designed the product on the assumption that you will review everything, because a competent attorney always does.
Your role and professional responsibility
Does ArsLex give legal advice or "practice law"? Who is responsible for the work?
No. ArsLex is a drafting tool that the attorney directs — the same way a word processor or a research database is. It does not give legal advice and does not exercise legal judgment. The attorney remains fully responsible for the work product, for reviewing every output, and for the final filing.
What about my ethical duties — competence, confidentiality, disclosing AI to the court?
ArsLex is built to support them. On competence (ABA Model Rule 1.1, which is generally understood to include understanding the technology you use), every output is verifiable against a real source you provided. On confidentiality (Rule 1.6), your client data is never used to train AI, stays on infrastructure we control, is kept until you delete it, and can be deleted or placed under a litigation hold whenever your matter requires it. On disclosure: some courts have standing orders about disclosing AI assistance, and those rules vary by jurisdiction and judge — check the ones that govern your matter. None of this replaces your professional judgment; it is designed to help you stay on the right side of it.
What can ArsLex not do?
It will not do your legal research from scratch, it will not tell you which argument to make, and it will not replace your judgment. It accelerates the mechanical, citation-heavy parts of drafting over evidence you have already gathered — and it expects you to review and own the result. We would rather be clear about the limits than oversell.
Do you also handle eDiscovery and document review, or only drafting?
Both. Drafting is ArsLex Associate. The ArsLex Discovery Suite is a separate, full eDiscovery platform your firm can run itself: ingest native files and load-file productions, review and code for responsiveness and privilege, redact, Bates-stamp, log privilege, and produce, with per-matter custody that keeps the working copy of your production live and searchable. It runs on the same zero-retention foundation described below, and is priced by usage, quoted per matter.
Can it code documents in bulk? What stops it from coding 5,000 of them wrong?
It can, and the answer is that you see the damage before it happens rather than after. Select a search result and ask to code it, and nothing is written yet. You get a preview first: how many documents actually change versus already carry that call, which family members are pulled along, which families would end up split with a parent produced and a child withheld, and a list of every document where the AI's read of the record disagrees with the call you are about to apply. That last list is advisory. It never blocks you, and the lawyer always wins the argument. If the gesture still turns out to be wrong, one action reverses it, and anything a colleague touched in the meantime is left alone rather than stamped over. Two things are deliberately kept out of bulk: removing a privilege mark is a one-document decision, because over-withholding shows up on a log and under-withholding is a waiver.
Do I have to read every email in a thread?
No. ArsLex works out which messages actually carry the conversation, so a reply that quotes everything above it makes the earlier copies redundant. You review the messages that carry content, and the rest of the thread follows your decision with a record of where that decision came from, rather than being quietly filled in. Attachments stay with their parent email throughout, including inside one continuous Bates range.
Your data and your confidentiality
Where does my data go — and is it used to train AI?
No. Your documents are never used to train any AI model — ours or anyone else's, and no model provider retains them. Text generation runs on Amazon Bedrock under AWS's zero-data-retention and zero-operator-access terms. Embeddings never go to a third party at all: wherever ArsLex uses them, including search by concept and grouping a production by subject in the Discovery Suite, they are computed on infrastructure we control and never leave it. Everything is encrypted in transit and at rest, isolated per tenant, and deletable at any time. See how we handle your data for the full detail.
How long do you keep my documents? Is there an auto-delete?
We keep your work until you delete it. Nothing is on a deletion timer unless you put it there. If your firm wants one, an administrator can turn on auto-deletion and pick a window — anywhere from 1 day to 365 days — either for the whole account or for a single matter, and change or remove it whenever you like. Firms that set a window before August 2026 keep it running until they change it; the setting and its current value are shown in the admin console.
When you do delete something, it is gone: the record, the search index, and the file are removed from live systems immediately, and the data leaves our encrypted backups within 30 days. There is no trash can and no restore — so keep your own copy of anything you cannot afford to lose. And if a matter is under a litigation hold, nothing in it can be deleted at all until your firm releases the hold.
What happens to my chat conversations with the assistant?
They are never saved. Conversations are held only in server memory so your signed-in devices (say, Word and your phone) can share one live thread, and they are wiped when your session ends, when you log out, or when you clear them: always within hours, never written to a database or disk, and never included in backups. Your documents and drafts are a different thing — those we keep until you delete them; the chat itself simply evaporates.
Who holds the encryption key, and can you prove a deletion actually happened?
Your documents are encrypted at rest as a matter of course. Beyond that, a firm can ask to have its documents in our durable store sealed under an encryption key that belongs to that firm alone, held in AWS Key Management Service and used for nothing and nobody else. It matters because of what it makes possible: destroying that key makes the firm's stored documents unreadable everywhere at once, which is a stronger and faster statement than walking a delete through every copy.
That is a serious thing to do, so it cannot be done quietly or quickly. Every administrator on the account is warned in the app and by email. Someone has to confirm it explicitly, and a confirmation typed against a stale warning does not count. A waiting period runs before anything is destroyed, the destruction itself takes deliberate human action from us rather than a scheduled job, and if we cannot reach an administrator at all, the request simply cannot go forward. Ordinary deletion is unchanged and is not on any of these clocks: delete a document and it leaves live systems immediately, and our encrypted backups within 30 days.
Can other firms see my data? What about deployment inside our own walls?
Each firm's data is walled off from every other tenant — isolation we enforce and test continuously. Access requires authenticated login with multi-factor authentication, and the database is never exposed to the public internet. For firms that must run the application inside their own environment, ArsLex is also available for on-premises deployment — your documents, database, and search index stay on infrastructure you control, while AI generation still calls our zero-retention model provider.
Still have a question? Email admin@arslex.ai — we respond within one business day.