Every legal AI tool on the market promises accurate citations. Most of them are still being caught hallucinating in filed briefs. The difference is architectural — not marketing. This is a plain explanation of how Ark's citation verification works, why hallucinated citations are impossible under ordinary operation, and what happens when an authority genuinely cannot be verified.
Paste any filing. Ark checks every citation against 7.4 million court opinions in seconds. Free, no signup.
Run a Free Citation CheckLarge language models produce plausible text, not true text. A general-purpose model asked to cite a case will confidently generate a plausible-sounding name, a realistic reporter citation, a plausible pin cite, and a convincing quotation — none of which needs to be real. The tool has no built-in notion that a citation refers to something external. It is optimizing for the next token, not the next fact.
Since 2023, over 1,000 filed cases have involved AI-hallucinated authority. Mata v. Avianca, the first case to go to sanctions, was an avoidable outcome — but it was avoidable only if the lawyer understood how the tool worked. Most tools, even specialized legal ones, still hallucinate at rates between 10 and 33 percent on adversarial queries.
The standard fix — "just check the citations yourself" — works for one brief but does not scale. A 40-page appellate brief with 80 citations takes hours to verify manually. A matter-resident AI that produces hundreds of propositions across months of work cannot be manually verified in any meaningful sense. The fix has to be architectural.
Ark is built on a legal database, not the other way around. The AI does not produce citations and then ask whether they are real. The AI is constrained, at generation time, to reference only authorities that exist in the database. Then every generated proposition is cross-checked against the cited authority for substantive match. Four layers, running in order:
Before the AI writes anything, the relevant authorities are retrieved from Ark's database — 7.4 million opinions, 50-state statutes, 142,000+ court rules, state administrative codes, federal regulations. The AI's context window contains only real primary sources. It cannot cite what it has not retrieved.
Every case name, reporter cite, statute, and rule the AI produces is resolved back to a specific record in the database. If the resolution fails, the citation is stripped and the passage is regenerated or flagged for human review. Unresolved citations do not reach the user. Fabrication is not a possible output.
Real citations that are used to support false or misgrounded propositions are the subtler failure mode. Ark extracts every opinion's salient points — the discrete legal and factual propositions the case actually stands for — at indexing time. When the AI attributes a proposition to a cited case, the proposition is matched against the salient-point extraction of that case. If it falls outside, the passage is flagged as unsupported.
Overruled, superseded, and distinguished authorities are flagged during generation. A case that was good law in 2015 may have been reversed in 2023; Ark's indexing captures subsequent treatment and surfaces any negative history at the point of citation.
Most legal AI products layer a language model on top of a legal database and hope the model behaves. Some add a post-hoc verification step, rejecting outputs that do not verify. Ark's architecture makes verification a constraint on generation, not a check after the fact. The model cannot produce a citation that does not resolve to a real record, because the generation is conditioned on the retrieval.
The practical difference: post-hoc verification catches many hallucinations but not all, and users often see the unverified output before the system rejects it. Retrieval-grounded generation prevents the hallucination from being produced in the first place. A compared hallucination rate of 10 percent versus 0.01 percent is the difference between "check everything twice" and "check as a sanity measure."
For the tool-by-tool comparison, see Ark vs ChatGPT for legal research, Ark vs Casetext / CoCounsel, and Ark vs Westlaw.
The failure mode matters as much as the success mode. When Ark encounters a question it cannot fully ground in primary authority, three things happen:
A legal AI that admits what it does not know is dramatically more useful than one that confidently fabricates. Silence on a hard question is the right behavior. Hallucination is not.
Zero is an engineering target, not a marketing claim. Ark's architecture prevents fabricated citations under ordinary operation. Edge cases — unusual citation formats, authorities in specialty reporters the database does not index, recent opinions not yet incorporated — can still produce unsupported output, which Ark flags rather than hides. The duty to verify before filing is still on the attorney; Ark's job is to make that verification trivial rather than a failure point.
Yes — that's what the free citation checker is for. Paste or upload any document. Ark extracts every citation and resolves it against the primary-source database, flagging hallucinated or questionable cites. This works regardless of what tool drafted the document, including ChatGPT, Claude, Harvey, or a human associate.
All 50 states and the federal courts. Ark indexes state supreme-court and appellate opinions, federal district and circuit opinions, Supreme Court opinions, state statutes, federal statutes, state administrative codes, federal regulations, and court rules. Coverage varies by completeness and recency — newer opinions typically appear within days of publication for federal and major-state courts, and within weeks for smaller jurisdictions.
Ark reports that it cannot find the authority and does not attempt to generate a plausible citation as a substitute. Users can submit the citation; the coverage gap is logged and addressed. Transparency about coverage is treated as core to trust, rather than a feature to minimize.
Related but not identical. Cite-checking traditionally means formatting citations to Bluebook standards — correct italicization, pin cites, signals, short forms. Citation verification checks that the cited authority actually exists and says what you claim it says. Ark does both: validates existence and substantive accuracy, and can also format to Bluebook, ALWA, or state-specific citation rules.
Sign up for Ark and every citation your legal research produces is already verified. No post-hoc review. No hallucinated authority. Just primary sources that exist.
Start Free TrialDisclaimer: Verification architecture reduces but does not eliminate the attorney's duty to review every citation before filing. Professional responsibility rules apply to AI-assisted work the same as to work produced by human associates.
Related: AI legal writing without court sanctions • Ark vs ChatGPT for legal research • Free citation checker