Legal AI Citation Checkers Compared: What Each Method Actually Catches

The three ways a citation can be wrong

Fabrication, misgrounding, and bad law are independent failure modes. Passing one gate says nothing about the other two, which is why a single score cannot substitute for three separate confirmations.

A legal citation can fail in three distinct ways, and each verification method is built to catch a different one. Understanding these axes is the prerequisite for comparing tools honestly, because a checker that catches one failure mode may be blind to the other two. The three modes are fabrication, misgrounding, and bad law.

First, a citation can be fabricated: the case simply does not exist. Second, it can be misquoted or misgrounded: the case is real, but the quoted language is wrong or the case does not stand for the proposition cited. Third, it can be bad law: the case is real and accurately quoted, but it has been overruled, superseded, or otherwise stripped of authority. These are not hypothetical. Stanford RegLab found leading legal AI research tools hallucinate "1 in 6 or more," with reported rates above 17 percent for one tool and above 34 percent for another ^[1]. Damien Charlotin's database now catalogs more than 1,300 court proceedings (as of 2026) flagging suspected AI hallucinations ^[2]. A method that catches only fabrication leaves two live risks in every filing.

It helps to see these as three independent gates rather than one spectrum of quality. A citation can pass the existence gate and fail the quotation gate. It can pass both and fail the good-law gate. The gates do not correlate: a perfectly real, perfectly quoted case can be worthless authority because a later decision overruled it, and nothing about the citation string itself will reveal that. This independence is the reason a single confidence score, however high, cannot stand in for three separate confirmations. When a tool reports that a citation "checks out," the first question is always which of the three gates it actually opened.

• Fabrication: the cited case does not exist. This is the failure that first drew public attention, and it is the one existence checks are built to catch.
• Misgrounding: the case exists but the quotation is inaccurate, or the case does not support the proposition it is cited for. Reading is required, not matching.
• Bad law: the case exists and is quoted accurately, yet it has been overruled or superseded and no longer carries the authority claimed for it.

How each verification method works

The three method categories differ in architecture, not just accuracy. A deterministic database lookup matches a citation string against a structured corpus of known cases. A generative-AI review asks a model to read the citation and judge it. A verifiable certificate resolves each citation against live case law and issues court-verifiable proof. Each approach carries structural strengths and structural blind spots.

Deterministic lookup excels at existence: if the case is in the database, the string either matches or it does not, and fabrication surfaces cleanly. It says little about whether a real case was quoted accurately or remains good law. Generative-AI review can, in principle, read a quotation against an opinion and reason about propositions, but its judgments are inconsistent because the same generative behavior that drafts a citation can also misjudge one. Good-law status specifically requires a citator, the function performed by tools like KeyCite or Shepard's, because subsequent history is not visible in the citation string itself. A verifiable certificate is designed to bind all three checks together and seal the result to a tamper-evident log, producing evidence rather than a flag. Its honest limit: it is only as complete as the case-law corpus it can reach.

The architectural differences carry directly into the blind spots. A deterministic lookup is fast and decisive precisely because it does not read; it compares strings against records. That same design means it has no view into whether the sentence a lawyer quoted appears in the opinion at all, or whether a later court cut the holding out from under it. A generative-AI review reads, which is what makes it appealing for the quotation and proposition questions, but reading by a model is probabilistic rather than deterministic, so the same input can draw different judgments on different runs. Neither limitation is a defect to be patched away; it is a property of the approach. Recognizing the property is what keeps a legal team from expecting existence-grade certainty out of a reading task, or reading-grade nuance out of a string match.

Comparison: what each method catches

No method is universally superior; each maps to different failure modes and different evidentiary needs. The table below compares the three method categories against the axes that matter for a filing: whether each catches fabrication, misquotation, and bad law, whether it produces verifiable proof, and whether it is built to withstand a post-quantum threat model. Most tools on the market flag suspected problems; fewer certify and seal a result.

The pattern is clear: existence, quotation, and good-law status are three separate checks, and a defensible workflow either stacks methods that cover all three or uses a category built to bind them into one verifiable result. Flagging tells a lawyer where to look; a sealed certificate tells a court what was verified and when.

The table compares method categories, not brands. A given product may combine categories, and a workflow can stack several. Read each cell as a property of the approach, not a scorecard for any one vendor.

Reading the comparison one axis at a time

A comparison table is only useful if each row is read for what it actually says. The fabrication row shows why the earliest AI citation stories centered on invented cases: existence is the easiest failure to catch, and a deterministic lookup catches it cleanly. That is genuine value, and it is also the narrowest of the three checks. A tool that stops at existence has closed the most visible risk while leaving the two quieter ones open.

The misquotation row is where reading becomes unavoidable. Confirming that a quotation matches the opinion, or that a case supports the proposition it is cited for, requires comparing language to language, which a string match cannot do. Generative-AI review can attempt this, but the same inconsistency that lets a model draft a wrong citation can let it clear one, so its judgment on this row is uneven rather than dependable. The bad-law row is quieter still: an overruled case passes both the existence and quotation checks, and only a citator function surfaces the later history that stripped its authority. The verifiable-proof and post-quantum rows are different in kind. They do not ask what a tool catches; they ask what a tool leaves behind and whether that record is built to remain trustworthy over time. Those are evidentiary questions, and they separate tools that warn from tools that certify.

When to choose which method

Choose based on the failure mode you most need to close and the evidence you need to produce. If your only concern is whether a cited case exists, a deterministic lookup answers that quickly and cheaply, because string matching against a known corpus is decisive for fabrication. Match the method to the risk you are actually trying to retire.

If you need to know whether a real case was quoted accurately or stands for the cited proposition, add a review layer, because existence checks alone cannot read meaning. If you need to know whether a case is still good law, use a citator, because overruling and superseding history live outside the citation string. And if you need to hand a court or a client independent, tamper-evident proof that all three checks were run before a filing was signed, choose a verifiable certificate, because flags are not evidence and a sealed log is. Most legal teams facing a judge's standing order on AI use fall into that last category, where the deliverable is proof, not a dashboard warning.

1. Name the failure modeDecide whether your live risk is fabrication, misquotation, bad law, or all three. The answer sets the minimum method you need rather than the most expensive one.
2. Confirm existenceUse a deterministic lookup to retire fabrication. If the case is in the corpus, the match is decisive; if it is not, you have surfaced the clearest failure fast.
3. Read for meaningAdd a review layer to test quotation accuracy and proposition support, and treat its judgment as a prompt for attorney review rather than a final verdict.
4. Check good lawApply a citator function so overruled or superseded authorities do not pass simply because they exist and are quoted correctly.
5. Decide whether you need proofIf a court or client will ask what you verified and when, choose a method that seals the result to a tamper-evident log rather than one that only flags.

The proof gap: flagging versus certifying

The decisive difference between method categories is not accuracy percentages but what they leave behind. Most citation tools flag: they surface a suspected problem for a human to resolve, then produce nothing durable. A verifiable certificate certifies: it resolves each citation against live case law for existence, quotation, and good-law status, then seals the result to a tamper-evident log that a court can independently check.

This gap matters because a flag is a private, ephemeral signal, while a certificate is portable evidence. When a court asks what was done to verify authorities before signing, "our tool flagged nothing" is an assertion; a sealed certificate is a record. RankShield's category certifies which citations are real, accurately quoted, and good law, and proves it, rather than claiming any drafting process is free of error. The honest boundary remains the corpus: a certificate is only as complete as the case law it can reach. Within that boundary, the difference between a warning and verifiable proof is the difference a judge can see.

The distinction is easy to blur because both a flag and a certificate can arrive as a green checkmark on a screen. What separates them is durability and portability. A flag lives inside the tool and the moment; once the session closes, what remains is a memory of a screen, not a record a third party can inspect. A certificate is built to be handed off. It states which citations were resolved, on what axes, and against what it could reach, and it seals that statement so that a later reader can confirm the record has not changed. For a filing, the second form is the one that survives the question a court is most likely to ask months after the brief was signed.

Why stacking methods still leaves a seam

A common and reasonable response to all of this is to stack methods: run a deterministic lookup for existence, a review layer for quotation, and a citator for good law. That covers the three failure gates, and for many workflows it is a sound approach. It is worth being honest, though, about what stacking does and does not solve. It solves coverage. It does not, by itself, solve evidence.

When three separate tools each produce a separate flag, a legal team ends up with three private signals and no single record binding them together. The lawyer knows the checks were run, but the artifact that could show a court all three were run, before signing, in one tamper-evident form, does not automatically exist. That is the seam a verifiable certificate is designed to close: not by catching a failure the stack misses, but by binding the results the stack produces into one sealed, portable record. Stacking and certifying are not rivals so much as different jobs. The stack decides what is true; the certificate makes what was checked provable after the fact. A team that only needs internal confidence may stop at the stack. A team that will need to demonstrate diligence to an outside reader has a reason to bind the result.

What a court is actually asking for

Standing orders on AI use vary, but they tend to converge on a single underlying question: what did you do to verify the authorities you cited, and can you show it. That is a question about process and record, not about which product a firm licenses. A tool that catches every failure mode but leaves no durable trace answers the first half and struggles with the second. A record that shows existence, quotation, and good-law status were each resolved before signing answers both.

This is why the evidentiary rows in the comparison matter as much as the detection rows. A judge assessing an AI-use certification is not grading a vendor; the judge is asking whether the filer can substantiate a representation made to the court. Detection without a record forces the filer back to assertion. A sealed certificate lets the filer point to something an independent party can check. Neither one changes the underlying professional obligation, and neither one is a shield against a citation that was wrong for a reason no tool could see. What the record changes is the filer's ability to demonstrate that reasonable verification steps were taken, in a form that outlasts the drafting session.

The corpus limit, stated plainly

Every method in this comparison shares one boundary, and it should be named rather than glossed. A checker's existence check is only as complete as its corpus. If a case is not in the database a deterministic tool searches, the tool cannot confirm the case exists, and depending on its design it may report the case as unknown rather than as real. The same limit applies to a verifiable certificate: it can resolve a citation only against the case law it can reach, and a citation that falls outside that reach is outside what any certificate can honestly attest.

This is not a reason to distrust the methods; it is a reason to read their outputs precisely. A clean existence result means the citation matched a known record, not that the universe of law was searched. A good-law confirmation reflects the citator's coverage, not an omniscient view of every subsequent decision. Naming the boundary is part of using the tools well, because it tells a legal team where automated verification ends and attorney judgment must begin. No honest account of these methods claims to eliminate the corpus limit, and no product in this category is complete beyond the case law it can reach.

This site is published by a technology vendor, not a law firm, and nothing here is legal advice. Verification tools support professional judgment; they do not replace it.

Attorney review remains the backstop

Whichever method or combination a team chooses, the same conclusion holds: a citation checker is support for professional judgment, not a substitute for it. Tools retire specific, nameable risks. A deterministic lookup retires fabrication for cases in its corpus. A review layer surfaces likely quotation and proposition problems for a human to weigh. A citator retires the specific risk that an overruled case is treated as good law. A verifiable certificate binds those results into a record a court can check. None of them reads a brief for whether the authority actually fits the argument, and none of them carries the professional responsibility that a signature does.

The most defensible posture treats these methods as instruments that make attorney review faster and more focused, not as a reason to review less. Match the method to the failure mode, stack methods when coverage requires it, and bind the result when you will need to prove diligence to an outside reader. Then read the citations yourself, because the checks tell you what a tool could confirm and the argument tells you what the case is actually being cited to do. The difference between a warning and verifiable proof is the difference a judge can see; the difference between verification and judgment is the part that still belongs to the lawyer.