Roderick Lucas is the founder and CEO of Moonlit, an Amsterdam legal publisher that gives lawyers and AI tools structured, daily-updated access to legislation and case law. He was previously AI Lead at Deloitte and lectures at Vrije Universiteit Amsterdam and Tilburg University. In this written Q&A with AI Time Journal, he explains how Moonlit's AI-drafted, expert-reviewed Commentaries are checked, what really causes hallucinations in legal research tools, what citation verification should look like after the SRA's August 2026 warning, and why paywalls around legal knowledge mean the market pays twice.
Moonlit’s new Commentaries are drafted by AI from primary sources and then reviewed by academic editors and practitioner-authors, with badges marking which passages an expert has validated. Where does the AI draft most often fall short, roughly how much of a draft changes in expert review, and what exactly must a reviewer confirm before a passage is badged?
The draft summarises a provision and the rulings that apply it accurately, because it can only cite documents we hold. What it does not know is which of those rulings practitioners rely on, where the tax authority reads a rule more strictly than the text suggests, and which exception comes up in every real case. It also treats every source as equally important. An editor rebalances that.
On review, the draft largely stands. The revisions often come down to refining the emphasis, and editing any claims that are not fully supported by the sources.
Before a passage is badged, the editor confirms three things: that the passage says what the commentary says it says, read in the original language; that the provision is quoted as it stood on the reference date, and no later ruling in the reference graph has narrowed or overturned the authority; and that the conclusion follows from those sources without a step the editor would not defend in front of a court. The badge means a named expert has done that check and put their name on the page. It does not mean the passage is beyond debate, but it substantiates that someone accountable stands behind it.
Your Data Layer is updated daily and flags new rulings to editors. How do you decide that a new judgment materially changes a commentary, and what is your target time from a ruling’s publication to an updated reference date?
The Data Layer handles the initial filtering, but deciding whether a ruling is truly material is always a judgment call for the editor. When a new ruling enters, the reference graph tells us whether it cites a provision a commentary covers or a case a commentary relies on. If it does, the editor gets a flag with the ruling and the passages it touches. Most of these flags are routine confirmations, such as a national court applying established European case law the way the commentary already describes. Those get noted and closed.
A ruling is material when it changes the practical legal advice or outcome. If the Court of Justice reinterprets an article of the VAT Directive, that clearly matters for every commentary covering that provision. On the other hand, a lower national court simply following established precedent does not change the underlying legal position, no matter how recent the judgment might be. The real nuance is in cases that create an exception, or narrow an existing one, and that is where the editor steps in to make the call, with the ruling linked next to the text.
Commercial commentary series update once or twice a year, so even a month would be a significant step up, and we aim well under that. Until the editor has finished, the reference date stays where it is, so a reader sees exactly which state of the law the text describes.
From building legal AI at Deloitte and now at Moonlit, what actually causes hallucinations in legal research tools: model behaviour, missing or stale source data, weak retrieval, or how lawyers frame their queries? Which cause do you think is most underestimated?
They all play a role in AI hallucinations, and they actually feed into each other and compound. Model behaviour is the one everyone talks about, where a model predicts what a citation should look like and simply invents one. Weak retrieval, on the other hand, produces a subtler failure. If the search pulls up a document merely close to the right authority, then the model reasons confidently from the wrong text. How lawyers frame their queries complicates this even further. For example, if a practitioner asks in English about a Belgian ruling and an English-only search step finds an incomplete summary or nothing at all, then the model is left to guess to fill the gap.
In my view, the most underestimated cause is missing data, which is a different problem from stale data. Most legal AI tools sit on a thin layer, typically the statute and whatever selection of case law was easy to scrape. The secondary material that explains how a rule works in practice sits behind publisher paywalls, and the model has never seen it in a reliable form. Ask it an operational question, and it does what any confident generalist would do. It extrapolates from the text to what the rule probably means.
That’s why the hallucination debate frustrates me. Everyone measures how often the model lies, and almost nobody asks what it was given to read. We built Moonlit around that question. When the model has to retrieve from a corpus where every record is real, and every citation resolves to a document, fabricated citations disappear. The errors that remain are misreadings, and those are the errors an expert can catch.
The SRA’s August 2026 warning notice makes clear that relying on AI is no defence and that every cited authority must be genuine and verifiable. What does a citation-verification workflow that would satisfy a regulator look like in practice, and how much of it can safely be automated versus left to the lawyer?
Every citation in a draft is resolved against an authoritative database before it is accepted. That is a lookup against records, so the citation either matches a document with a link to the official publication or it is rejected. The passage relied on is opened and read in the original language. The version is checked against the date that matters for the case, and subsequent treatment is checked to verify how later courts handled the ruling. Finally, the lawyer records that they did this, so a supervisor or a regulator can see the audit trail.
The mechanical side of that workflow automates remarkably well. Existence, link, version and citing decisions are database questions, and a machine answers them more reliably than, say, an associate manually checking references in the middle of the night. Our MCP tools take care of that directly, meaning a citation either points to a document in the Data Layer or it doesn’t appear at all. Our research instructions tell the assistant to open the passage before citing it and to say when the sources leave a gap.
What stays with the lawyer is the substantive judgment the SRA cares about, namely ensuring that the authority is relevant, that it supports the argument being made, and that the submission is properly arguable. No tool should pretend to do that. A tool that pretends to is more dangerous than one that does nothing, because it invites the lawyer to stop reading. The honest division is that machines guarantee the citation is real and current, and humans guarantee it is right.
You have described legal knowledge as ‘a clear case of paying twice.’ Where specifically do paywalls around commentary and case law slow innovation, and what pays for free Commentaries so that the model is sustainable?
The slowdown happens at two points. The first is commentary. A developer building a VAT tool, for instance, can get the Directive from EUR-Lex and rulings from the Court of Justice. What they cannot get is the explanation of how one article is applied to a specific fact pattern in five member states, because that explanation is behind subscription services that prohibit automated access. So every legal AI product either builds without that layer and hallucinates where it is missing, or pays a steep per-seat licence that was priced for law firms.
The second is case law that is public in theory and unusable in practice. Rulings appear on national portals in inconsistent formats, often with no stable identifier and no machine-readable citations. The text is free but structuring it into reliable data costs engineering years. Commercial publishers did that work decades ago and simply sell the output to this day. Every new entrant ends up repeating that exact same heavy lifting. That is the second payment, and every team that wants to build pays it.
What pays for free Commentaries is the rest of Moonlit. Commentaries is integrated into the Platform that law firms, tax advisories, and public institutions subscribe to. It also reaches other legal AI products through our Data API and MCP connector, which commercial teams pay to query. Our editors are compensated directly from that enterprise revenue, while individual practitioners with a free account read the text without a paywall. The economic logic mirrors the Data Layer itself: the more systems that read from Commentaries, the more valuable that verified layer becomes. The cost of running an expert editorial workflow is spread across the entire infrastructure rather than charged to each reader.
Moonlit is adding Commentaries to its Data API and MCP connector, which already bring your legal data into Copilot, ChatGPT and Claude, where you no longer control how content is presented. How do you preserve provenance, reference dates and review status once your content is inside a general-purpose assistant, and what have you seen go wrong?
We do not control the presentation, so we embed the provenance into the data itself. Every result the MCP connector returns carries its source link to the official publication, and for Commentaries it will carry the reference date, the editor and the review status of each passage. When an AI assistant references a passage, it has the metadata needed to cite the reference date and the editor alongside it. Then, a reader who clicks through will land on the Moonlit page, where the badge and the editor’s name sit on the text.
Where things tend to go wrong with external assistants is citation stripping. General-purpose models often summarise aggressively and drop the reference date or the link entirely, which leaves the user without a way to verify the currency of the analysis. That is why building the link and metadata into the underlying tool payload is essential, ensuring the user always has a clear path back to the verified record.
I should note that while our primary legal Data Layer is live on the API and MCP today, Commentaries integration is still in the works, but it is coming soon.
The launch centres on EU indirect tax, where harmonised rules are applied differently by national courts. What does that teach you about structuring cross-jurisdiction legal data so machines can use it reliably, and which part has been hardest to get right?
VAT is the best test case in European law for exactly this reason, with it having one directive, but twenty-seven different implementations, and national courts that make different interpretations on the same article. The lesson is that you cannot flatten it. A model that treats the Directive and the Dutch VAT Act (Wet op de omzetbelasting 1968) as the same rule will be wrong in one of them.
So we keep every document distinct and make the relationships explicit. The directive article is one record. The national provision that implements it is another, linked where it cites the directive. A Court of Justice judgment interpreting the article is linked to the article, and a national ruling applying that judgment is linked to the judgment. The commentary sits on top of that structure and reads it in order, moving from the EU rule and Court of Justice interpretation, down to the national implementation and national case law. So if, for example, the Dutch Hoge Raad and the German Bundesfinanzhof diverge, the commentary shows both and the editor explains which position European case law actually supports.
The hardest part has been the citations. A German court cites a Court of Justice ruling using German formatting, a Dutch court uses Dutch conventions, Belgian courts switch between French and Dutch syntax, and all of them cite domestic provisions in formats that change by decade. Member states do not always specify which national article implements which directive article, so we resolve it from the texts and legal engineers verify the patterns.
Where has generative AI genuinely changed the day-to-day research workflow of the lawyers using Moonlit, and where has adoption stalled despite the tools being available?
The gathering phase has changed completely. For teams that have adopted Moonlit, the administrative slog of jumping between national portals, searching in multiple languages, and copying citations into a memo has effectively disappeared. Quantoz Payments is an example: their legal team connected our MCP server to Claude and built workflows for cross-border payments regulation that query every relevant jurisdiction at once and return the primary documents with official links. The first pass of a cross-border question now takes minutes, and the lawyer’s time goes into analysis and risk. Monitoring has changed in much the same way. An automated Monitor on a statute or a court, when set up on the Moonlit platform can replace the weekly ritual of checking what moved.
We’ve observed, though, that beyond the retrieval step, there is noticeably less adoption. Many lawyers will let a tool find the documents, but hesitate to let it read or synthesise them. Partly that caution is justified: the SRA notice and the court sanctions have made practitioners careful, and being careful is sensible. But part of it is also institutional habits. The firms where it stalled most are the ones that bought a general assistant without a verified data layer under it, saw it fabricate a citation in the first week, and decided the technology was not ready. Having written off the entire category because of experiences like that, they assume every legal AI tool has that same flaw, and end up missing out on the operational and productivity edge that properly structured, reliable systems can already deliver.
Which AI leader or practitioner, outside your own company, are you watching closely right now, and why?
Damien Charlotin, who built the database tracking court decisions involving AI hallucinations. He is a practitioner and an academic rather than an AI vendor, which essentially means he has no product to defend and no commercial incentive to flatter the technology. His database is arguably the only honest scoreboard the industry has. As of writing, it has already recorded over 2,100 decisions, more than 800 of them involving legal practitioners, where a court identified or penalised reliance on fabricated authorities.
While most AI companies measure hallucination on synthetic benchmarks they designed themselves, he tracks it at the point where it actually cost someone their professional reputation. The trajectory of that database is how I assess whether the legal tech field is truly solving this problem. So far those numbers are still climbing. When that curve starts to flatten, verified data layers will have played a central part in turning it.
