Skip to content

explainerAI GovernanceAI HallucinationsAI in Courts

Must lawyers verify AI output before relying on it?

Across 130 country records, 50 have a binding, guidance, proposed or case-law answer on verification of AI output.

Edited and verified by Cognesio LLP

Researched with AI assistance · sources verified by Cognesio LLP · How this was made ↓

Yes, in many jurisdictions lawyers must check AI output before they file it, cite it or rely on it. As of 4 September 2026, the verification-duty category records 10 binding country answers, 28 guidance answers, 11 case-law answers and one proposal. The remaining 80 country records are either none or unclear; 54 of those are provisional.

How was the count made?

The count comes from the country YAML files in src/content/countries/, one country record at a time. The script read each country’s verification-duty cell and grouped the status as binding, guidance, proposed, case-law, none or unclear. Sub-national rules, such as Australian state practice notes and United States standing orders, support the country cell but do not create extra country counts.

That matters because a country marked binding can still have a limited rule. The United States entry is not a single national statute; it is a combination of Rule 11 certification, ethics opinions, court orders and sector rules. Australia is similar. The map records the operative answer for a legal-risk reader, then the country page gives the limits.

Where is it binding?

The country map marks verification as binding in Australia, Canada, Ireland, Philippines, Poland, Qatar, Singapore, Slovakia, Spain and the United States.

The binding group is mixed. Some answers are court practice directions, such as Ireland’s HC 142 and Singapore’s Registrar’s Circular. Some are professional rules or ethics opinions, such as the United States position combining Rule 11 certification, state rules and ABA Formal Opinion 512. Australia is binding because several courts now attach verification duties to filed material.

Where is it guidance or case law?

Guidance records are United Arab Emirates, Austria, Belgium, Brazil, Switzerland, Czechia, Germany, Estonia, the European Union, Finland, France, the United Kingdom, Hong Kong SAR, Indonesia, Israel, International bodies, Japan, South Korea, Luxembourg, Malaysia, Nigeria, the Netherlands, Norway, New Zealand, Sweden, Türkiye, Taiwan and Vietnam.

Case-law answers are Argentina, Chile, China, Colombia, Denmark, India, Italy, Kenya, Pakistan, Portugal and South Africa. Ecuador is proposed.

What do the documents actually say?

The tracker records the duty in each document’s key-provision field. The primary texts support those records. The SRA warning notice says solicitors should be assured that named case-law submissions are “genuine, relevant, have a verifiable citation, and advance the arguments that are being put forward in your documents.”

The Irish High Court’s Practice Direction HC 142 says a person responsible for a court document “must ensure that any material generated wholly or partly with the assistance of a GenAI Tool has been appropriately and independently verified before it is used in proceedings.” It also says: “It is not sufficient to use the same or other GenAI Tools to confirm that information is accurate.”

The Singapore Courts guide says: “As a Court User, you are fully responsible for the content in all your Court Documents.” It also tells users to verify that AI-generated references “actually exist and stand for the legal positions that are attributed to them.”

The Supreme Court of India judgment in Gadiya quotes its earlier rule that decisions are to be set aside “even if an iota of fake or hallucinated material enters the decision-making process.”

What pattern do the rules follow?

The duty usually has two parts. First, a lawyer remains responsible for the filed document or advice even if software produced the first draft. Second, verification must use authoritative sources, not another AI answer. Ireland states that point expressly. Singapore reaches the same result by making the filer responsible for every court document, while the SRA frames it through the duty not to mislead the court.

The case-law category reaches the same place after a failure. India’s Gadiya is an example: the Supreme Court set aside a customs penalty because the adjudicator relied on AI-generated authorities that did not exist or did not support the propositions used. That is why the category covers both lawyer conduct and adjudicator reliance.

Where has nothing been found?

The script counted 77 country records marked none and three marked unclear for this category. Most of the none records are not editor-graduated yet: 54 are provisional and should be treated as researched, not finally checked. The category page is the canonical list; this explainer uses the counts, not a claim that absence is permanent.

What does it not cover?

A verification duty is not the same as a disclosure duty. Some jurisdictions require disclosure of AI use in filings; others, such as Singapore, do not require a pre-emptive declaration unless the court asks. It is also not a client-consent rule, a confidentiality rule or a vendor-procurement rule, although those duties often appear in the same documents.

Sources