TL;DR: As of mid-2026, hallucinations remain the defining risk of legal AI — but the profession has learned to manage them. Courts in the US, UK, Canada, Australia, and India have confronted fabricated citations and answered with sanctions, disclosure requirements, and practice guidance that all point the same way: the lawyer who files it owns it. Meanwhile the technology split in two — general-purpose chatbots still fabricate freely, while grounded legal platforms cite page and passage so errors are caught in seconds. The firms that thrive treat every AI answer as a draft and every citation as unverified until opened.
When a New York court sanctioned two lawyers in 2023 for filing a brief full of cases that ChatGPT had invented, it looked like a cautionary tale that would end the practice within months. It did not. Fabricated authorities kept appearing in filings around the world through 2024 and 2025, and courts kept catching them. What changed by mid-2026 is not that the risk vanished — it is that the profession now understands it, courts have built a consistent response to it, and the tooling divide between grounded and ungrounded AI has become impossible to ignore.
What a hallucination is — and is not
A hallucination is output that is fluent, confident, and unsupported by any real source: a case that does not exist, a quote that appears nowhere in the judgment, a statutory provision with the wrong effect. It is worth being precise about what it is not. An authority that exists but is weaker than suggested, or good law cited for a proposition it only partly supports, is a research-quality problem — serious, but different. The hallmark of a true hallucination is that there is nothing underneath: the citation resolves to no document at all.
The distinction matters because the two failures have different fixes. Fabrication is attacked by architecture — grounding the answer in retrieved documents. Overstated support is attacked by verification — a lawyer reading the cited passage in context. A defensible workflow needs both, which is why our verification guide treats them as separate checks.
How courts responded, 2023 to mid-2026
The judicial response began with Mata v. Avianca in 2023 and spread with remarkable consistency. Courts in the United States imposed monetary sanctions and referred lawyers to disciplinary bodies. English courts confronted fabricated citations in 2025 and made clear that submitting invented authority can engage serious professional consequences, including referral to regulators. Canadian and Australian courts reached similar conclusions, and Indian courts flagged AI-fabricated material in filings as well. The details differ by jurisdiction; the principle does not: a lawyer is responsible for every authority cited, however it was found.
Just as telling is what courts did not do. No major jurisdiction banned lawyers from using AI. The consistent line — visible in judicial guidance in England and Wales, in US standing orders, and in bar association opinions — is that AI use is permitted, and in many contexts unremarkable, provided the output is verified by a human who takes responsibility for it.
Disclosure orders, sanctions, and practice directions
Three instruments now recur across jurisdictions as of mid-2026:
- Certification and disclosure requirements. Some judges require parties to disclose or certify the use of generative AI in drafting, or to certify that every citation has been human-verified.
- Sanctions and costs orders. Fabricated citations have drawn fines, adverse costs, struck filings, and referrals to disciplinary authorities — with penalties trending upward as courts lose patience with repeat patterns.
- Judicial and bar guidance. Judiciary-issued guidance for judges and court users, and bar opinions such as the American Bar Association's Formal Opinion 512, frame verification as part of the existing duty of competence rather than a new rule. Our summary of bar ethics opinions tracks the pattern.
Why hallucinations persist in 2026
Three years of model improvements did not eliminate fabrication, because fabrication is not a bug in a specific model — it is a property of ungrounded generation. A language model produces the most plausible continuation of the text. Ask it for authority it does not have, and the most plausible continuation is citation-shaped text: plausible party names, a plausible year, a plausible reporter. Larger models produce more plausible fabrications, which are harder to spot on their face, not easier.
The persistent real-world driver is simpler still: lawyers under time pressure using general-purpose chatbots — free, fluent, and always available — for tasks those tools were never designed to do, then skipping verification because the output looked finished. The failure is a workflow failure as much as a technology one, which is why firm policy (below) matters as much as tool choice. For the underlying mechanics, see our explainer on how legal LLMs work.
What actually improved by mid-2026
The genuine progress came from architecture and workflow rather than raw model quality:
- Retrieval-grounded answers became the standard for serious legal tools. Purpose-built platforms now draft from retrieved primary-law passages rather than model memory, which sharply reduces fabrication.
- Citations became verifiable objects. The better tools cite to the exact page and quoted passage and link straight to the source, turning verification from an afternoon into seconds.
- Deterministic citation labels. Leading platforms stopped letting the model write citation strings; labels are copied from source metadata, so a real authority can no longer be mangled into a fake-looking one.
- Source permanence. Archiving web sources as permanent snapshots at retrieval time means a citation cannot rot between research and filing.
- Honest uncertainty. Well-designed systems now confirm the jurisdiction, ask clarifying questions instead of guessing, and flag low-confidence answers rather than papering over them.
The numbers: what studies measured
The most-cited measurement remains the Stanford RegLab and HAI study, which tested leading proprietary legal research tools and found hallucination rates between 17% and 33% — despite marketing that suggested the problem was solved — with general-purpose models performing worse on the same legal queries. Whatever the current figures for any given tool, the study's structural lesson has not aged: marketing claims and measured behaviour are different things, and only architecture you can inspect — grounding, citations, source links — deserves weight in a procurement decision. Our 2026 statistics roundup collects the adoption and accuracy data in one place.
Rules of engagement for firms in 2026
The firms that use AI heavily and safely converge on a short set of working rules:
| Rule | What it means in practice |
|---|---|
| Every AI answer is a draft | Nothing AI-generated goes to a court or client without human review and sign-off. |
| Every citation gets opened | Existence, support, status, and weight are checked against the source — no exceptions for "obvious" cites. |
| Confidential work stays on approved tools | Client documents never go into consumer chatbots; approved platforms must not train on your data. |
| Grounded tools for legal questions | Research happens on platforms that cite page and passage, not general assistants. |
| Know your court's requirements | Check standing orders and practice directions on AI disclosure in every forum you file in. |
Codifying these into a one-page policy — and training everyone from partners to paralegals on it — is the highest-return governance step a firm can take. Our guide to AI governance for law firms includes a fuller framework.
How Judicio approaches the hallucination problem
Judicio attacks fabrication with architecture rather than assurances. Research answers are drafted from sources retrieved across 33 dedicated legal databases spanning 100+ jurisdictions; every proposition carries a formal citation to the exact page and passage; citation labels are deterministic — drawn from the source, never written by the model; and each web source is archived as a permanent PDF snapshot so cited material cannot disappear. The same discipline runs through Document Review and the Review Matrix, where every finding and every cell quotes the clause and page it came from, with confidence flags that admit ambiguity instead of hiding it. The full design is on our methodology page — and because we assume verification, not obedience, the workflow is built for a lawyer to check everything fast. Try it against your own matters with a 7-day free trial: 500 credits, no credit card.
Judicio outputs are for research and informational purposes and are not legal advice. Statements about court responses are general as of mid-2026 — check the current rules and orders of the forum you practise in.