Key takeaways
- Hallucination is a workflow problem with a workflow solution: no factual assertion enters work product without a source a reader can open in one click.
- The rule must apply identically to senior lawyers and to AI output. A double standard makes it unenforceable, and senior assertions are a real source of error too.
- Anchor to page and paragraph, not to a document name. 'See the contract' is not an anchor.
- An unanchored line should be treated as a failed preflight check - a draft that is not ready - rather than as a scandal.
- AI drafting inside a strict anchoring rule is generally safer than human drafting outside one, because the rule catches the error class that both produce.
What hallucination actually is, and what it is not
A hallucination is a confident, fluent, and false assertion produced by a language model - a case that does not exist, a holding stated backwards, a statutory section misnumbered, a quotation that appears nowhere in the source. The defining property is not that the model is wrong. Humans are wrong routinely. The defining property is that the output carries no internal signal of its own unreliability. It reads exactly like a correct answer.
This is why the intuitive defences fail. Reading more carefully does not help much, because the false output is stylistically indistinguishable from the true one. Using a better model reduces the frequency but does not change the category. Prompting the model to 'only cite real cases' produces a model that asserts more confidently that its citations are real.
The error class is not new to law. A lawyer misremembering a holding from a case they read three years ago produces exactly the same failure: a confident assertion, fluently expressed, unsupported. The profession has always managed that risk through verification. What AI changes is the volume and the speed, which means the verification step has to move from a habit to a rule.
The rule, stated in one sentence
No factual or legal assertion enters work product without a citation to a specific, checkable source that a reader can open in one click and verify at the level of a page and paragraph.
That is the whole mechanism. Everything below is implementation detail. The reason it works is that it does not attempt to detect hallucination at all - it makes unverified assertions structurally unable to ship, regardless of whether a human or a machine produced them.
Why it must apply to humans first
If you enforce anchoring only on AI output, you have created a two-tier standard that says senior judgment is exempt from verification. It will not be respected, it will not survive a partner's objection on a busy day, and it leaves the human error path completely open. Apply it to everyone, starting with the most senior person in the room.
Implementing it in a working practice
- 01
Write the rule down and circulate it
One sentence, in the practice's own document, with a named owner. Unwritten standards are preferences; written ones are process. This takes an afternoon and is the highest-leverage single action in this entire guide.
- 02
Apply it to human-drafted work for two weeks before touching AI
This does two things. It surfaces how much unanchored assertion is already in your work product - which is usually a genuinely uncomfortable number - and it establishes the rule as a practice standard rather than an anti-AI measure, which is how it survives politically.
- 03
Define what counts as an anchor
Specify the granularity. A document name is not an anchor. A pin cite to a page and paragraph, a clause number, a timestamped exhibit reference, or a direct link to the paragraph in your document store is. Write two or three examples of a compliant anchor so nobody has to guess.
- 04
Make the check a preflight step, not a review step
The anchoring check happens before substantive review, mechanically, on every draft. Reviewers should never be discovering unanchored assertions during a substantive read - by then their attention is on the argument, which is exactly where the misses happen.
- 05
Instruct AI systems to anchor inline as they draft
Every drafting prompt or agent configuration includes the anchoring requirement, and the agent is instructed to cite as it writes rather than to append citations afterwards. Post-hoc citation is where fabrication concentrates.
- 06
Treat a failure as a normal draft error
An unanchored line is a draft that is not ready, in exactly the same way a missing signature block is. Not a crisis, not a disciplinary matter, not evidence that AI cannot be trusted. Normalising this is what keeps people reporting misses rather than quietly fixing them.
What changes once the rule is in place
The first effect is immediate and slightly deflating: you discover the practice was already shipping unanchored assertions regularly, and always was. Most of them were correct, because experienced lawyers usually remember correctly. Some were not, and were caught late or by luck.
The second effect is that AI becomes deployable. Once every assertion carries an anchor by rule, the review burden on machine-drafted text collapses from 'reconstruct where each claim came from' to 'open the anchors and confirm'. That is a fast pass rather than a slow one, and it is the difference between AI saving time and AI creating a new checking job.
The third effect is durable and worth more than the first two. Work product becomes self-verifying. A colleague picking up a matter, a client's in-house counsel, an opposing party, or a court can follow every assertion to its source without asking anyone. That property compounds across the practice and is very hard for a competitor to retrofit.
The professional-responsibility position
Nothing about this rule changes who is responsible for the work. The lawyer reviewing and signing an output is responsible for it, whether the first draft came from a junior colleague, a precedent bank, or an agent. The anchoring rule does not distribute responsibility; it makes the responsibility discharargeable in reasonable time.
This is also the honest answer to clients and regulators asking how a practice using AI maintains standards. The answer is not 'we use a very good model'. The answer is 'every assertion in our work product is anchored to a checkable source, the standard applies identically to human and machine output, and a named lawyer reviews and takes responsibility for everything that ships'. That is a process answer, and process answers are the ones that hold up.
Frequently asked
Can AI hallucination be eliminated completely?
Not at the model level, and betting on that is the wrong strategy. What can be eliminated is unverified assertions reaching work product, which is the outcome that actually matters. A standing anchoring rule addresses the consequence rather than the cause, which is why it works regardless of which model you use or how it improves.
Does using AI for legal drafting create professional-responsibility risk?
It creates the same risk class as delegating a first draft to a junior lawyer: the reviewing lawyer remains responsible for the output. The risk becomes unmanaged when there is no verification step - which is a process failure, not a technology failure. Under an enforced anchoring rule with named human review, the responsibility position is straightforward and defensible.
What counts as a proper citation anchor?
Something a reader can open in one click and verify at paragraph granularity - a pin cite with page and paragraph, a clause reference, a linked exhibit with a timestamp, or a direct link into your document store. A document name, a case name without a pin cite, or 'per the agreement' are not anchors.
Should the anchoring rule apply to senior partners?
Yes, and this is not negotiable if you want the rule to function. A standard that exempts the most senior people signals that verification is a junior-staff formality, which guarantees it is quietly abandoned under time pressure. It also leaves open a real error path, since confident misremembering by experienced lawyers is a genuine source of mistakes.
How do I check AI-generated legal citations efficiently?
Require the agent to anchor inline as it drafts, then run the anchoring check as a mechanical preflight step before substantive review. Opening a list of pin cites and confirming each is a fast, bounded task. Reconstructing the provenance of an already-written draft is slow and is where checking fatigue causes misses.