ChatGPT for Lawyers: Real Risks and Safer Alternatives
ChatGPT can help a lawyer draft, summarize, and brainstorm, and it can also leak client facts into a consumer system and invent case law. This post covers what lawyers actually use it for, the real risks, safer usage rules, and where purpose-built legal AI differs.
Try it before you read on
Run the live agent on a fictional sample matter: pick a case, pick a task, and watch it produce a case brief, a deadline timeline, or a drafted response.
What lawyers actually use ChatGPT for
Start with the honest observation: a large share of lawyers have already tried a general-purpose chatbot for work, whether or not their firm has a policy about it. Bar association technology surveys have tracked generative AI moving from novelty to routine tool in a couple of years. The common uses are sensible ones:
- Drafting from a blank page. A first cut at a letter, an email to opposing counsel, a policy, a set of interrogatories to react against.
- Rewriting and summarizing. Turning a dense clause into plain English for a client, tightening a paragraph, summarizing a long document the lawyer pastes in (which is exactly where the confidentiality problem starts).
- Brainstorming. "What arguments would the other side make?" "What questions should I ask this witness?" The model is genuinely useful as a sparring partner with no ego.
- Explaining unfamiliar territory. A quick orientation to an unfamiliar area of law before doing real research in a real database.
- Non-legal work. Marketing copy, job postings, spreadsheets formulas, meeting summaries: the office work around the practice.
Used this way, with nothing confidential going in and everything factual being verified on the way out, a general chatbot is a legitimate productivity tool. The trouble is that the convenient path and the safe path diverge fast, and the convenient path is one paste away.
The four risks that actually matter
1. Confidentiality: pasting client facts into a consumer product
The fastest way to get value from a chatbot is to paste in the document or the fact pattern, and that is precisely the problem. A consumer chatbot is a third-party service; depending on the plan and settings, prompts may be retained, reviewed for abuse, or used to improve the service. A lawyer's duty of confidentiality does not have a "but the tool was convenient" exception, and several bar ethics opinions on generative AI converge on the same conclusion: lawyers must understand where prompt data goes before client information touches the tool, and consumer-grade settings generally do not clear that bar. The exposure is invisible, which makes it worse; nothing appears to go wrong at the moment the duty is breached.
2. Hallucinated citations
A general chatbot writes plausible text, and a fabricated citation is plausible text in its most dangerous form. In widely reported cases over the past few years, courts have sanctioned lawyers who filed briefs citing cases a chatbot invented: correct citation format, real-sounding party names, no underlying decision. The pattern repeats across jurisdictions, and it repeats for the same reason: the output read like research, so it was filed like research. A chatbot with no retrieval layer is not searching a legal database at all; it is composing text that resembles the answer. Every citation it produces must be treated as unverified until you pull the case yourself, as we detail in our legal research software approach and our AI research verification guide.
3. No audit trail
Legal work needs a record. When a draft comes out of a personal chatbot session, there is no matter-linked history of what was asked, what was produced, and what the attorney changed before it went out. If a client, a court, or your malpractice carrier ever asks how a document was prepared, "an associate's personal chat history, since deleted" is not an answer you want to give. Chat logs also live outside the case file, so the work product is unfindable six months later even for your own team. That is the gap legal document management software exists to close: every version filed against the matter, with who changed what and when.
4. Terms of service and privilege
Consumer terms of service are written for consumers, not for privileged material: they typically disclaim confidentiality obligations of the kind lawyers need, and they change. There is also a live, unsettled question about whether sharing privileged facts with a consumer AI service could support a waiver argument, since privilege depends on keeping communications within a protected circle. No lawyer wants to litigate that question about their own file. Purpose-built legal tools answer it with contracts: business terms, data processing agreements, and explicit no-training commitments.
If you use ChatGPT anyway: the safety rules
Many lawyers will keep using general chatbots for the lightweight work, and that is defensible if the guardrails are real. A minimal set:
- No client identifiers, ever. No names, no case numbers, no unique fact patterns that identify a matter. If you would not put it in a public filing, do not put it in a consumer prompt. Abstract the question: "a delivery driver rear-ends a stopped vehicle" carries no client data.
- Verify everything factual. Every citation pulled and read, every quote checked, every legal claim confirmed in a real database before it influences advice or a filing. Treat the output as a draft from an unlicensed stranger, because that is what it is.
- Turn off training and history where possible. Use the settings or the business tier that excludes your prompts from model improvement. It narrows the confidentiality gap; it does not close it.
- Put it in a firm policy. One page: approved tools, prohibited inputs, mandatory verification, and who to ask. The sanctioned-lawyer cases are, almost uniformly, stories of individuals improvising without one.
- Disclose where required. A growing number of judges have standing orders about AI-assisted filings. Know yours before you file.
A note on the business tiers
The team and enterprise tiers of the major chatbots are a real improvement over the consumer product: prompts are excluded from training by default, admin controls exist, and the terms are written for organizations. If a general chatbot is going to be part of your practice at all, one of these tiers should be the floor, and several bar technology guides say as much. But upgrading the terms of service does not upgrade the tool's nature. A business-tier chatbot still has no legal database behind its answers, still fabricates citations with the same fluency, still knows nothing about your matter that you do not paste in, and still leaves the work product outside the case file. The tier solves the data-handling question; it leaves every other row of the comparison below untouched.
A useful mental model: the business tier makes the chatbot a defensible place to think. It does not make it a defensible place to research, and it does not make it a system of record. Those two jobs need tools that were designed for them, which in practice means comparing legal case management software on where the work product actually lands.
How purpose-built legal AI differs
The difference between a general chatbot and purpose-built legal AI is not intelligence; the underlying models are often similar. The difference is everything wrapped around the model: where the data goes, what the output is grounded in, and where the work lands. This is the reason the category of AI for lawyers exists at all.
| Dimension | Consumer ChatGPT | Purpose-built legal AI |
|---|---|---|
| Data handling | Consumer terms; retention and training depend on plan and settings | Business terms, DPA, no-training commitments, encryption designed for client data |
| Citations | Generated from patterns; can fabricate authority | Grounded in a legal corpus; flagged for attorney verification |
| Matter context | None; you re-paste facts every session | Works from the case file: facts, documents, deadlines already in context |
| Audit trail | Personal chat history, outside the file | Work product saved to the matter with review history |
| Workflow | Copy-paste in, copy-paste out | Drafts, chronologies, and deadlines land where the work happens |
| Cost | Free to ~$20–$60/user/mo | Higher; priced against paralegal hours, not chat subscriptions |
The matter-context row deserves a moment, because it is the difference lawyers feel first. A chatbot starts every session at zero: you re-explain the parties, re-paste the facts, re-describe the procedural posture, and the moment the session ends that context evaporates. A purpose-built system that lives inside the matter already knows the file, so "draft a status update for the client" or "what deadlines does this order create" are one-line requests instead of twenty minutes of setup. The productivity gap between the two approaches is mostly this gap.
Caseagent sits at the far end of that right-hand column: it is legal case management software with the AI agent inside the case file, so the drafting, research memos, and deadline tracking happen in the matter, with every authority flagged for verification and every draft waiting on attorney review. It is in early access, which we state plainly: no customers yet, a working demo on the homepage, and a signup list for firms that want the agent-in-the-file model instead of the paste-into-a-chatbot model. Either way, the rule that keeps you safe is the same: the attorney reviews everything, and nothing a model writes is legal advice.
Legal AI that lives inside the case file
No pasting client facts into a consumer chatbot. Caseagent works the matter from inside it, with verification flags and attorney review built in. Join the early-access list and we'll email you when your spot opens.