What AI is genuinely good at in legal work
The honest starting point: general AI tools are useful in a law practice, and pretending otherwise loses to the firms that use them well. They are strong at transformation tasks, where the source material is in front of them:
Hand it these
- Summarizing a deposition transcript or a long opinion you provide.
- Restructuring your own draft: tightening, reordering, adjusting tone.
- First-draft frameworks for routine documents you then rewrite.
- Explaining unfamiliar doctrine well enough to know what to research properly.
- Brainstorming counterarguments to stress-test a position.
Never hand it these
- Finding the case law your argument will stand on.
- Stating what a case holds without reading it yourself.
- Deciding whether authority is still good law.
- Anything containing privileged or confidential client facts.
- The final say on any citation that will appear in a filing.
The common thread: on the left, you supply the facts and the authority, and you evaluate the output. On the right, the tool would be the source of record, and that is exactly where it fails.
Where it fails, and why the failures are invisible
Language models generate plausible text, not verified facts. Legal research is precisely the task where that distinction bites, because the output looks identical either way. A fabricated case and a real one read the same on the page.
| Task | Risk | The safeguard |
|---|---|---|
| Finding case law | Severe: fabricated cases, blended holdings, invented quotes. | Treat every AI-suggested authority as a lead, never a source. Pull and read each case in a real database before it touches a draft. |
| Stating what a case holds | High: real case, misstated holding or misattributed quote. | Read the opinion. Confirm the holding at the cited page yourself. |
| Summarizing provided documents | Moderate: omissions and subtle mischaracterizations. | Spot-check against the source, especially anything you will rely on or quote. |
| Checking if law is current | Severe: training cutoffs and no citator. The model cannot know about last month’s reversal. | Citator, always. No AI answer about current validity substitutes for it. |
| Confidential facts | High: consumer tools may retain or train on inputs. | Only use tools whose terms prohibit training on your content, and scrub privileged detail regardless. |
The ethics rules that apply
- Competence. The duty of competence now includes understanding the benefits and risks of the technology you use. Bar guidance on generative AI has been consistent: you must understand that these tools fabricate.
- Candor and reasonable inquiry. Rule 11 and the duty of candor apply to the filing, not to how it was drafted. Signing a brief certifies the legal contentions were checked, whoever or whatever produced the first draft.
- Confidentiality. Pasting client facts into a consumer chatbot can put confidential information on third-party servers under terms you have not read. Several bar opinions treat this as a live confidentiality problem.
- Supervision. The duty to supervise non-lawyer assistance extends to software. An unreviewed AI draft is unsupervised work product going out under your name.
A workflow that keeps the speed and drops the risk
The pattern that works treats AI as a drafting accelerant inside a verification loop, never as a research source:
- Research. AI suggests doctrines, search terms, and possible lines of authority. All actual authority comes from a real legal database, read by a lawyer.
- Drafting. AI produces structure and prose from your outline and your verified authority. Supply the citations yourself; never let the model add authority.
- Editing. AI tightens, reorders, and polishes. You confirm no citation or quotation changed during the rewrite.
- Verification. No AI role. Every citation is checked for existence, quote accuracy, pincite support, and current status before filing. This stage exists because of the other three.
Stage 4 is where CiteScreen lives. Upload the finished draft and every citation is extracted and verified against its source: existence confirmed, quotes compared word for word, pincites checked against the cited page, with plain-English reasons for anything flagged. It is the systematic backstop that makes the first three stages safe to speed up.
Frequently asked questions
No jurisdiction prohibits it outright. The obligations attach to what you do with the output: verify authority before filing, protect client confidences, supervise the work, and comply with any judge-specific disclosure orders.
They fabricate less than general chatbots because they are grounded in real databases, but misstated holdings and misattributed quotes have been documented across tools. The verification duty is unchanged: read the case, check the page.
It depends on the court. Some judges have standing orders requiring disclosure or certification; most jurisdictions have no general disclosure duty. What every court requires is that the filing be accurate, which is the verification duty.
Run every citation through the four checks: existence, quote accuracy, pincite support, and citator status. Done manually it takes several minutes per citation; CiteScreen automates the first three and hands you the short list that needs judgment.