AI for Law Firms: Writing Without Fake Citations
Lawyers are writing with AI now, whether or not the firm has a policy about it. Partners use it to draft client alerts on a rule change, associates use it to refresh a stale practice-area page, and marketing teams use it to turn a partner's talking points into a bylined thought-leadership article. The appeal is obvious, because the writing never gets done otherwise. The risk is just as obvious. A single chatbot, handed a legal prompt, will produce fluent text that reads like a competent lawyer wrote it, and buried inside that text can be a case citation that does not exist. This article explains why single-agent AI for lawyers hallucinates, why a multi-agent approach with built-in fact-checking is the structural fix, and how one AI writing for law firms tool, Ghosts, is built specifically to keep that failure out of your firm's published work.
Key takeaways
- A single AI chatbot writes in one pass and has no separate step that checks claims against real sources, so it can confidently invent a perfect-looking citation that does not exist.
- Courts have sanctioned lawyers for filing AI-fabricated citations, starting with the widely reported 2023 Mata v. Avianca case and continuing into 2025.
- Multi-agent AI legal writing splits the job across a researcher, a writer, a dedicated fact-checker, and an editor, which structurally blocks hallucinations a single prompt cannot catch.
- Ghosts for law firms runs doubled fact-checking and a claims scrub on every legal draft, behind a privacy layer built for firms.
- The tool drafts; a licensed attorney at the firm reviews, edits, and holds final editorial responsibility for anything published.
The problem: one chatbot, one confident guess
The danger in AI for law firms is not bad grammar or a dull tone. It is the fabricated citation, delivered with total confidence. The most cited example is Mata v. Avianca, the 2023 matter in the Southern District of New York in which lawyers submitted a brief containing cases that ChatGPT had invented. The court imposed sanctions after the fabricated authorities were discovered. You can read the docket and the court's order on CourtListener's Mata v. Avianca record.
That was not a one-time embarrassment that the profession learned from and moved past. Courts have continued to sanction lawyers for AI-fabricated citations in the years since. In 2025, a federal court in Wyoming sanctioned attorneys after they filed a brief citing cases that did not exist, produced by a generative AI tool, in Wadsworth v. Walmart. The pattern repeats because the underlying tool has not changed. The lesson is not that AI is unusable in a law practice. It is that raw single-chatbot output cannot go near anything with your name on it.
Why a single agent hallucinates
A chatbot predicts plausible-sounding text one token at a time, in a single pass. It is optimized to continue the sentence in a way that looks right, not to confirm that the sentence is true. There is no separate step in that process that stops, pulls the actual reporter volume, and checks whether the case it just named exists and says what the draft claims it says. So when a prompt calls for supporting authority, the model generates a citation with the correct shape, a real-sounding case name, a plausible reporter, a believable pinpoint, and none of it is anchored to a real document. This is the core of the AI hallucination problem for lawyers. The output is not flagged as a guess. It arrives looking exactly as authoritative as a citation you spent an hour verifying, which is precisely what makes it dangerous.
The American Bar Association has told lawyers this is their problem to manage, not the software's. ABA Formal Opinion 512, issued in July 2024, sets out that lawyers using generative AI still owe duties of competence, candor to the tribunal, and supervision, and must understand the tools' limits, including the risk of inaccurate output. The full opinion is available as a PDF from the American Bar Association. A single unchecked chatbot draft puts every one of those duties at risk at once.
The fix: multi-agent writing with built-in fact-checking
The structural answer is to stop asking one model to do everything in one pass. Multi-agent AI legal writing splits the work into separate roles that check each other, which is closer to how a careful firm actually produces written work.
- A researcher gathers real sources first, so the draft is built on material that exists rather than on the model's memory of what a source might say.
- A writer turns that sourced material into the piece, in the voice the assignment calls for, working from the research rather than inventing support to fit a conclusion.
- A dedicated fact-checker runs a claims scrub. It goes through the draft assertion by assertion and attributes, softens, or removes anything it cannot verify, and it blocks invented citations before the draft moves on.
- An editor reviews the sourcing and the writing together and offers fixes, so the piece that reaches the lawyer is already tightened.
The reason this beats a single prompt is not that any one agent is smarter. It is that verification becomes its own step, performed against real sources, instead of a thing the writing model was quietly expected to do while it was busy predicting the next word. A single chatbot has no such checkpoint. This is why legal AI tools that separate research, drafting, and verification can prevent the exact failure that produced the sanctions cases above. It is also why treating legal document automation as a one-click generator, rather than a checked pipeline, is the mistake to avoid.
Single chatbot versus multi-agent legal AI
| Concern | Single-chatbot prompt | Multi-agent legal AI |
|---|---|---|
| Citations | Can be invented, look perfect | Pulled from real sources, invented ones blocked |
| Fact-checking | None as a separate step | Dedicated pass plus claims scrub |
| Client data privacy | Often unclear or shared | Isolated workspace, no default human review |
| Data retention | Rarely controllable | Firm-set windows, provable purge |
| Used for AI training | Frequently yes | No, contractual and technical |
| Who signs off | Nobody by default | A licensed attorney at the firm |
The solution in practice: Ghosts for law firms
Ghosts, the AI writing tool built for law firms, is designed around exactly this multi-agent, checked approach. It drafts publishable legal content, including client alerts, practice-area pages, and bylined thought leadership, behind a privacy layer built for firms, with extra fact-checking on every piece. The point is not to remove the lawyer. The point is to hand the lawyer a draft that has already survived the checks a single chatbot skips.
On the verification side, every legal draft in a law workspace runs doubled fact-checking plus a claims scrub that attributes, softens, or removes unverifiable assertions and blocks fabricated citations before the draft ever reaches the lawyer. Those extra verification passes run on every draft, at every plan tier, not only on the expensive plan. That is the difference between a tool that hopes the model got the citation right and a tool that treats an unverifiable citation as something to stop.
Privacy and data handling built for firms
Confidential material is the other reason firms hesitate to adopt AI writing for law firms. Ghosts addresses it in three concrete ways, documented on its trust and security page for your diligence file. First, the firm sets its own data retention policy, with per-status windows from 24 hours to a year, and expired content is hard-deleted by an automated shredder that leaves a tamper-evident, provable purge log. Second, content is never used to train AI models, a guarantee the company describes as both contractual and technical, end to end. Third, every workspace is isolated with database-level row-level security, and no Ghosts employee reviews a firm's content by default, so there is no client-data crossover between firm users or between firms. A Data Processing Addendum is available for the firm's review.
A dedicated legal bench, and a lawyer who signs off
Ghosts runs a dedicated legal bench, meaning specialist AI writers for litigation, corporate, employment, and personal-injury content, rather than one generic writer stretched across every subject. The named legal writer, Bernard, is built to be precise and cautious, yet plain enough that a client can actually read the result. Every draft arrives with research and sources cited inline, writing in the voice the piece calls for (the firm's or the client's), an editor's review, and the credit cost shown before it drafts so nothing is a surprise. Plans start at $29 a month, and the $29 plan starts with a 7-day free trial. Across all of it, a lawyer at the firm holds editorial responsibility and the final say. The AI legal content writer produces the draft, and the attorney reviews and approves it.
Where firms actually use this
The practical value of AI for lawyers shows up in the writing that never gets done because it is not billable. Concrete uses include:
- Client alerts when a rule, regulation, or statute changes and clients need to hear from you first.
- Refreshing practice-area pages that are three years stale and quietly costing search visibility.
- Partner bylined thought-leadership articles that bring in work but keep losing to billable hours.
- Firm blog posts and plain-language legal guides for prospective clients.
- Client newsletters and litigation updates that summarize a development without overstating it.
- Corporate and employment-law explainers, and personal-injury content, drafted by a writer tuned to that practice.
- Ongoing content programs for reputation and search visibility, produced on a schedule instead of in a rushed weekend.
In every one of these, the workflow is the same. The tool drafts, the checks run, and a licensed attorney reviews and signs off before anything is published under the firm's name.
A note on responsibility
This is worth stating plainly for a legal audience. AI-generated content is a starting point, not a finished work product. No draft from any tool, Ghosts included, should be filed, sent to a client, or published without a licensed attorney reading it in full, verifying every citation and factual claim, and approving it. The technology described here is built to make that review shorter and safer. It does not replace it, and it does not shift professional responsibility away from the lawyer whose name appears on the work.
FAQ
Can AI-fabricated citations really lead to sanctions?
Yes. Courts have sanctioned lawyers for filing briefs with citations that a generative AI tool invented, beginning with the widely reported 2023 Mata v. Avianca case in the Southern District of New York and continuing with cases such as the 2025 Wadsworth v. Walmart matter in Wyoming. The fabricated authority is the recurring cause.
Why does a single chatbot invent citations in the first place?
A chatbot predicts plausible text in one pass, with no separate step that checks a claim against a real source. It generates a citation that has the correct shape and a believable name, but nothing anchors it to a document that exists. The output is not marked as a guess, so a fake citation looks exactly as confident as a real one.
How does a multi-agent approach prevent AI hallucination?
It separates the work into roles that check each other. A researcher pulls real sources, a writer drafts from them, a dedicated fact-checker runs a claims scrub that verifies or removes every unverifiable assertion and blocks invented citations, and an editor reviews the result. Verification becomes its own step against real sources, which a single prompt never performs.
Does using AI writing for law firms remove the lawyer's responsibility?
No. The lawyer keeps full editorial responsibility and the final say. Tools like Ghosts draft and check the content, but a licensed attorney must review, verify every citation, edit as needed, and approve before anything is published or filed. ABA Formal Opinion 512 makes clear the duties of competence and candor stay with the lawyer.
How is confidential client material protected?
In the case of Ghosts, every workspace is isolated with database-level row-level security, no employee reviews a firm's content by default, and content is never used to train AI models, a guarantee described as contractual and technical. The firm also sets its own data retention windows, and expired content is hard-deleted with a provable purge log.
What to require before your firm writes with AI
Before any legal AI tools touch client-facing work, hold the vendor to a short, concrete list. Require verification as a built-in step, not a setting, so that fact-checking and a claims scrub run on every draft and invented citations are blocked before a lawyer sees them. Require that your content is never used to train AI models, in writing. Require workspace isolation and a retention policy the firm controls, with proof that expired content is actually deleted. Require that a licensed attorney reviews and approves everything, and that the tool is built to support that review rather than route around it. A tool that meets that list, such as the AI writing tool built for law firms described here, lets a firm publish more without betting its name on a citation no one checked. Ask for the Data Processing Addendum and the trust and security documentation, put them in your vendor file, and only then let the drafting begin.