When an AI tool answers a legal question without identifying its source, the lawyer who receives that answer faces a specific problem. The answer may be correct, but the lawyer has no way to know that without doing the verification work independently. The tool has saved time on retrieval. It has created work on verification. Whether that is a net saving depends on how long the verification takes and whether the lawyer has the depth to do it efficiently. Often the answer is: not much time saved, and more risk loaded onto the lawyer's shoulders than was there before.
The professional duty that makes sourcing mandatory
Legal professional standards in most jurisdictions require counsel to be able to explain the basis of any advice they give. In the UAE context, practitioners are subject to Federal Law No. 23 of 1991 on Legal Professions and its amendments, as well as the rules of the relevant bar association or professional body. These frameworks do not address AI specifically, but they do not need to. The underlying duty is the same regardless of how the research was conducted: the lawyer must be able to stand behind the answer.
An AI response without a citation is not itself an answer in the professional sense. It is a summary of what the AI believes to be the position, without the documentation that would allow the lawyer to verify it and vouch for it. The lawyer who relies on an uncited AI response and incorporates it into advice is assuming responsibility for the accuracy of that response without the means to have checked it. If the response is wrong, the lawyer is exposed; the AI tool is not.
This is not an abstract risk. Language model outputs on legal questions can be plausibly worded and confidently phrased while being substantively wrong. The model may produce a provision number that sounds right but does not exist, or cite a rule that applied under a prior statutory version now superseded. Without access to the source, the lawyer cannot detect the error before relying on it.
The audit trail dimension
Beyond individual advice, legal teams working in regulated environments face an audit trail requirement. When a dispute arises and the team's advice is reviewed, the question is not only what they said but what they relied on. An advice file that includes research notes with citations provides demonstrable evidence of a diligent process. An advice file that includes AI-generated summaries with no sources is harder to defend under scrutiny, even if the underlying advice happened to be correct.
In-house teams are particularly exposed here. Outside counsel can point to their firm's review process. An in-house team that has adopted an AI tool and produced a stream of uncited research notes has created a documentation gap that could become significant if internal decisions are challenged. The General Counsel who adopted the tool is accountable for what the tool's outputs looked like.
This consideration is separate from whether the AI tool was accurate. An accurate uncited output provides less audit trail protection than an accurate cited output. The citation is not just about accuracy assurance; it is about demonstrable process.
When tool designers choose not to cite
Not all AI legal research tools omit citations because of technical limitations. Some do it by design. The design argument is that citations distract from the answer, that lawyers want a clean prose response rather than a cluttered list of footnotes, and that the citation layer can be accessed separately if the lawyer wants to dig in.
There is a version of this argument that is defensible for certain use cases, for example, rapid internal orientation on a topic where the lawyer will then do independent research. For that use case, a clean prose summary is a useful starting point. The problem arises when that summary is the end point, when the lawyer uses it to form a view rather than as a prompt for further research.
The professional context of legal advice does not readily support the orientation use case as the primary workflow. Lawyers are not generally asking questions to orient themselves before doing separate research; they are asking questions to get answers they can use. If the tool is positioned as a time-saving research assistant rather than as a preliminary overview generator, the absence of citations means the tool is saving the retrieval step and creating a separate verification step. That is not the same as saving research time overall.
The hallucination problem in legal context
Legal AI hallucinations are a known problem. A language model produces an output that looks like a legal citation but refers to a provision that does not exist, or an article number from the wrong statute, or a case name that is a composite of real and invented details. The model does not know it is hallucinating; the output is generated from patterns in training data that happen not to correspond to an actual source.
The hallucination problem is not solved by requiring citations, but citation requirements contain the damage in a useful way. When a tool is designed to cite, and that citation can be verified against the original source text, a hallucinated citation is detectable. The lawyer follows the citation, finds it does not exist or does not say what the tool claimed, and knows to discard the response. This detection is only possible when the citation is present and specific.
A tool that provides prose answers without citations has hallucinations that are not detectable at the point of use. The lawyer cannot check a source that was not provided. The error is detectable only if the lawyer independently researches the same question and reaches a different answer, which is precisely the work the tool was supposed to save.
What verifiable sourcing looks like
For legal research purposes, a citation is verifiable if it includes enough information to locate the exact passage in the primary source. "UAE Commercial Companies Law" is not a verifiable citation. "Federal Law No. 32 of 2021, Article 88, paragraph 2" is verifiable. The difference is that the second citation tells the lawyer exactly where to look, and the lawyer can confirm within seconds whether the cited passage says what the tool claimed.
This level of specificity is achievable by a research tool designed around it. The technical challenge is indexing source material at paragraph level and associating each piece of retrieved text with its precise provenance. This is an engineering investment, not an impossibility. The choice to build this layer or not is a product decision, and it reflects a view about who bears the verification burden.
At Qanooni, we made a specific product choice to front-load the citation work rather than defer it to the lawyer. Every response includes article-level citations with direct links to the source material. We did this because we built the tool for professional legal use from the start, and professional legal use has a non-negotiable sourcing requirement. A tool that does not meet that requirement is not suitable for that context, regardless of how fast or accurate the underlying model might be.
The appropriate scope of this critique
We are not arguing that all AI legal tools without citations are worthless or professionally irresponsible as a categorical matter. There are use cases, orientation, competitive intelligence, general background research, where an uncited summary is a reasonable output for the purpose. The critique is narrower: tools positioned as research tools for professional legal practice, where the outputs are intended to inform advice or decisions, must provide citations for those outputs to be professionally usable. The verification gap is not a minor convenience issue. It is a structural gap between what the tool provides and what the professional context requires.
