← Back to HAQQ Blog

The Maker Cannot Be the Checker: A Verification Protocol for Legal AI

By Stephane Boghossian · · 8 min read · Guides

Regulators in London, Abu Dhabi, Dubai and Doha now say the same thing: AI doesn't move your responsibility. A six-step protocol for checking AI work before it leaves the firm.

Banks learned this decades ago. The person who books a trade doesn't approve it. Law firms are learning it again with AI.

A model that drafts a memo cannot also be the proof that the memo is right. Asking the same tool "are you sure?" is not a check. It's a second draft.

The short answer

What the regulators actually require

Regulator or courtWhat it says about checkingSource
SRA (England and Wales), 17 Aug 2026Be assured that all named case law authorities you submit are genuine. AI does not diminish or transfer your responsibilities.
Law Society (England and Wales), Oct 2026AI is not a legal entity. Solicitors must keep the ability to carry out and scrutinise work without AI.
DIFC Courts, PGN 2/2023Don't rely on AI content without first verifying its accuracy, and check the sources the tool relied on.
ADGM Courts, *Arabyads v Alam*Failing to verify AI research was reckless and a breach of the Rules of Conduct.
QICDRC (Qatar), PD 1/2026All citations must be checked against primary sources or recognised legal databases.

Five bodies, three continents, one rule. Check against the primary source, and own the result.

The six-step protocol

1. Separate maker and checker

The person who prompted the tool should not sign off alone. On small teams, at least rotate: whoever drafted with AI reviews someone else's AI-drafted work.

2. Check every authority at the source

Open the judgment, the law or the regulation itself. A summary from the same tool, or from another chatbot, is not a source. This is the one step every regulator in the table names.

3. Check the pin cite, not just the case

Fake cases get caught. Real cases cited for a point they don't make get through. Arabyads involved both: authorities that didn't exist, and others that didn't support the point.

4. Ask the same question three times

Run the key question again, in a fresh session, with different wording. If the answers disagree on the law, the number or the outcome, treat all of them as unverified. Consistency doesn't prove an answer right. Inconsistency proves you can't trust it yet.

5. Check the Arabic against the Arabic

For Gulf and Levant work, verify the Arabic text of the article you cite, and its current version after amendments. A correct English summary of an outdated Arabic article is still wrong.

6. Keep a record of what you checked

Log the tool, the task, the sources you opened and who verified them. In DIFC you may need to declare AI use. Anywhere, you may need to explain it. A record written only by the AI proves only that the AI wrote it.

A one-page checklist

Before any AI-assisted document leaves the firm:

HAQQ Legal AI is designed to make step 2 faster. It cross-checks verified sources and shows where each answer came from, so you can open the source in one click. It speeds up checking. It doesn't replace the checker.

Frequently asked questions

Can I use a second AI model to check the first?

As a filter, yes. As proof, no. Disagreement between models is a useful warning. Agreement isn't verification, because both can draw on the same wrong summary.

Do I have to tell the court I used AI?

In DIFC, the guidance expects early declaration. In Qatar, only if the court asks. In onshore UAE and Saudi courts, there is no specific rule yet. See The Three AI Rulebooks a UAE Lawyer Works Under.

What did the lawyers in Arabyads get wrong?

They filed a defence relying on authorities that didn't exist or didn't support the point, without verifying the AI-assisted research.

Does this slow us down?

Steps 2 and 3 take minutes per authority. Arabyads cost AED 282,508.

Sources and further reading

FAQ

Can I use a second AI model to check the first?

As a filter, yes. As proof, no. Disagreement between models is a useful warning. Agreement isn't verification, because both can draw on the same wrong summary.

Do I have to tell the court I used AI?

In DIFC, the guidance expects early declaration. In Qatar, only if the court asks. In onshore UAE and Saudi courts, there is no specific rule yet. See [The Three AI Rulebooks a UAE Lawyer Works Under](/blog/uae-ai-rules-for-lawyers-onshore-difc-adgm).

What did the lawyers in Arabyads get wrong?

They filed a defence relying on authorities that didn't exist or didn't support the point, without verifying the AI-assisted research.

Does this slow us down?

Steps 2 and 3 take minutes per authority. *Arabyads* cost AED 282,508.