There are two conversations about AI and the law. The first is about legal reasoning: can a large language model interpret an ambiguous statute, assess the strength of competing legal arguments, predict what a court will decide? This is the interesting conversation. It is also, for this paper's purposes, mostly a distraction, because the answer is "sometimes, with caveats, in ways that require expert oversight to be reliable" — and because it crowds out a second category of legal AI work where the answer is already clearly yes.
The second conversation is about legal compliance checking: can a large language model read a document against a specific checklist and tell you whether the document satisfies the checklist? This is not a question of legal reasoning. It is a question of structured pattern matching, grounded in a document the model is given rather than one it must recall. The model does not need to exercise judgment about what the law means. It needs to determine whether the document in front of it contains language that satisfies a requirement whose terms are already specified.
For this task, the answer is yes — for a specific, bounded reason this paper now states precisely rather than by assertion: the task is document-grounded, not memory-dependent, and §5 explains why that distinction is the one doing the real work.
PBOMaster's constitution analyser is built on this distinction. Understanding what it does — and what it does not do — matters for the broader argument this paper makes about where AI creates genuine, checkable value in the legal compliance domain.
Not whether it can reason about law — whether it can read a document against a list. It can.
1. The grounded-task claim this paper makes, and what would break it
What it argues from. Published benchmark results on LLM performance on open-ended legal reasoning (bar exam, law school exams).12 A published benchmark on LLM hallucination rates for open-domain legal factual recall.3 A general finding from natural-language-processing research that grounding a model's output in retrieved or provided text reduces hallucination relative to generating from the model's own training-time memory.4 A description, from the inside, of the specific gap-patterns PBOMaster's constitution analyser has found across real Kenyan constitutions.
What it does not argue from. It does not argue from an independently audited accuracy benchmark of PBOMaster's own constitution analyser — no such external audit exists, and the "what the analyser finds" section in §3 describes patterns observed by the company that built the tool, not a third-party evaluation. It does not argue that the document-grounded/open-domain distinction in §5 makes document-grounded legal AI error-free; it argues the distinction explains why the error profile should differ, which is a narrower and more defensible claim.
What it does not claim. It does not claim large language models can replace legal judgment for the genuinely interpretive fraction of compliance work — §2 and §6 both state this fraction is real and is excluded from this paper's claim of reliability. It does not claim the checklist-matching task is immune to a different category of error (a document-grounded model can still misread or mis-summarise a clause that is actually present); §7 names this directly as a risk this paper's own evidence does not rule out.
What would refute it. Given in full in §9. The sharpest version: if a rigorous, independent study finds LLM-based constitution or document compliance-checking tools have an error rate on document-grounded, checklist-structured tasks that is not meaningfully lower than their error rate on open-domain legal recall tasks, the document-grounded/open-domain distinction this paper treats as doing the explanatory work would be shown to be the wrong explanation, and this paper's confidence in §2's claim would need to be withdrawn, not merely qualified.
2. The Checklist Problem
Return to the Second Schedule of the PBO Act. The requirements for a compliant constitution are, to use the technical term, enumerable. They constitute a list of specific things that must be present. Each item on the list can be checked: either the document contains language that satisfies the requirement, or it does not. Some items require checking the language itself — does the conflict-of-interest clause prohibit participation in conflicted decisions, or merely require disclosure? — but even these involve checking against a defined standard, not exercising open-ended legal judgment.
This is the task large language models are, on the published evidence, well-suited to. The academic literature on LLM performance in legal tasks has developed rapidly since 2023. Katz et al.'s evaluation of GPT-4 on the Multistate Bar Examination found performance at or above the passing threshold across all subject-matter categories tested.1 Choi et al., evaluating ChatGPT on law school examination questions across four University of Minnesota courses (95 multiple-choice and 12 essay questions), found it capable of passing with grades in the C+ range.2 Both results describe performance on the hard problem — the one involving legal reasoning, argument construction, and interpretation of ambiguous fact patterns. A C+ grade or a bar-exam pass is a genuinely impressive result for that category of task, and also, on its own terms, not a guarantee of reliability a compliance workflow could be built on without human review.
The compliance-checking problem this paper is about is structurally easier, for a reason §5 makes precise rather than asserting by analogy: it does not require the model to reason about what the law means or to recall a fact from its training data. It requires the model to determine whether a specific provision is present in a document supplied directly in the prompt, and whether the language of that provision satisfies a defined standard also supplied directly. The model does not need to be a lawyer. It needs to be a thorough reader with a complete checklist and the source text in front of it.
3. What the Analyser Finds
In the course of building PBOMaster's constitution analyser and testing it against Kenyan NGO and CBO constitutions, a pattern emerged that was consistent enough to be worth documenting here. It is Flux's own observation from its own tool, not an independently audited finding. The following gaps appear across constitutions of different ages, from organisations of different sizes and types.
Conflict-of-interest provisions are absent or insufficient. This is the most consistent gap. Most constitutions older than 2013 were drafted before conflict-of-interest governance became a standard expectation in Kenya's civil society sector. Many constitutions that do have a conflict-of-interest clause have one that requires disclosure but does not specify the consequence: board members are expected to declare interests, but the constitution does not prohibit participation in conflicted decisions. The PBO Act requires both.
Dissolution clauses contain the wrong asset-reversion language. Many constitutions have dissolution clauses but specify that assets revert to "members" or are "distributed according to board resolution" rather than reverting to another public benefit organisation or to a public fund. This was acceptable under the old framework. Under the PBO Act, it is a disqualifying defect.
Board term limits are absent or non-compliant. The Act's governance requirements include provisions designed to prevent board capture — indefinite board tenure by founding members being a particular concern in long-running CBOs. Constitutions drafted before the Act often have no term limits, or term limits written in a way that allows indefinite renewal.
Meeting provisions use ambiguous language. Quorum requirements specified as "a majority of members" are not specific enough — the Act requires numerical clarity. Meeting frequency specified as "as required" rather than at defined intervals does not satisfy the minimum meeting-frequency requirements.
The public-benefit-objectives clause is too general. Many constitutions state objectives at a level of generality — "to promote community development and welfare" — that does not map clearly onto the First Schedule's categories of qualifying public-benefit activities. The constitution needs to be specific enough to make the regulatory case.
These are not obscure defects. They are structural, common, and invisible to an organisation reading its own constitution without a compliance checklist in hand — which is precisely the condition this paper's broader argument, in Africa's regulatory failure is not a law problem, treats as the translation-layer gap.
4. The Two Kinds of Legal AI
Legal AI is often discussed as though "can AI do legal work?" were a single question with a single answer. It is not. The question has very different answers depending on which legal task is being asked about, and §5 argues the relevant variable is not "how legal is the task" but "where does the needed information live."
At one end: can AI predict how a judge will rule on a novel question of constitutional interpretation? No, not reliably. This involves reasoning about ambiguous text in the context of a specific factual record, anticipating how a particular judicial temperament will weigh competing considerations, and exercising the kind of integrative judgment that comes from years of legal practice. AI can assist with components of this task. On the published evidence, it cannot be trusted to do the task unsupervised.
At the other end: can AI determine whether an organisation's constitution contains a conflict-of-interest clause that prohibits participation in conflicted decisions? On the evidence in §2 and §5, yes, reliably enough to be useful as a first pass. This involves finding text that is present in the document supplied, reading it against a defined standard also supplied, and reporting the result. It does not require legal judgment. It requires thoroughness, a good checklist, and the source text in context — which, §5 argues, is the condition under which language models are measurably more reliable.
PBOMaster is an instance of the second kind of legal AI. The constitution analyser does not reason about the law. It reads a document against a checklist the legislation itself provides. The value it creates is not in the interpretation — the interpretation is already done, by the Act and its schedules. The value is in the systematic application of that interpretation to every document that needs checking, faster and, this paper argues without an external audit to confirm it, more completely than most individual reviewers would manage unaided.
5. Why the Distinction Holds: Grounding vs. Recall
This paper's central distinction rests on something more rigorous than an appeal to intuition, because "checking a document is easier than reasoning about the law" risks sounding like a convenient claim rather than a tested one.
A 2024 study found that large language models hallucinate — produce confident, false factual claims — on a striking proportion of specific, verifiable legal questions: 58% of the time for GPT-4 and 88% of the time for Llama 2, when asked about specific federal court cases.3 This is a serious, well-documented finding, and it would seem, on its face, to undercut this paper's confidence in LLM-based legal tooling generally. It does not, for a specific reason worth stating precisely rather than waving at: the tasks in that study were open-domain factual recall — asking the model to retrieve, from its own training-time memory, a fact about a specific court case it was not given in the prompt. The model had to "remember" whether a case existed, what it held, and what court decided it, with nothing but the question itself to work from.
The compliance-checking task this paper is about is structurally different in exactly the dimension that matters. PBOMaster's constitution analyser is not asked to recall anything from memory. It is given the organisation's actual constitution, in full, in the prompt, alongside the actual text of the Second Schedule's requirements. The question is not "what do you know about PBO Act compliance in general" but "does the text in front of you, right now, contain language satisfying the text also in front of you, right now." This is a grounded task, not a recall task, and the distinction is not this paper's own invention: a 2021 study found that retrieval-augmented generation — providing a model with the relevant source text rather than relying on what it learned during training — measurably reduces hallucination relative to ungrounded generation, across multiple benchmark conversation and question-answering tasks.4
It establishes that there is a real, independently documented mechanism (grounding versus recall) that predicts the checklist-matching task should have a meaningfully different, and probably better, error profile than the open-domain legal-recall task Dahl et al. studied. It does not establish that PBOMaster's own analyser achieves any specific accuracy rate — no external audit of it exists — and it does not establish that grounded tasks are error-free. A model reading a provided document can still misread a clause, summarise it incorrectly, or miss a provision that uses unusual phrasing. That is a different, and on its face more checkable, kind of error than inventing a court case that never existed — a human reviewer can re-read the same provided document and verify the model's claim directly, which is categorically harder to do for a fabricated case citation drawn from nowhere the reviewer can see. But "more checkable" is not "absent," and §7 returns to this directly.
This has a direct, practical implication for how a tool like PBOMaster's analyser should be built, and it is worth stating because it is a falsifiable design commitment rather than a background assumption. If grounding is what is doing the reliability work, the analyser's design should never ask the model to answer from what it "knows" about the PBO Act in general — it should supply the Act's actual Second Schedule text in the prompt alongside the organisation's actual constitution, every time, rather than relying on the model having memorised the Act's requirements during training. A version of the same tool that instead asked the model "does this organisation's constitution comply with the PBO Act?" without supplying the Act's text directly would be closer, in structure, to Dahl et al.'s open-domain recall condition than to the grounded condition this paper's confidence depends on — and this paper's argument would apply to it much more weakly, if at all. This is a testable claim about PBOMaster's own architecture, not a general property of "AI checking documents" as a category.
6. The Generalisation
The broader argument is this: the discourse around AI and professional services has been captured by the hard cases — the ones that require genuine reasoning, judgment, and professional accountability. These are real and interesting. But they have obscured the category of work where AI is already more reliable, already more scalable, and already underused: the work of checking documents and processes against specific, enumerable standards supplied directly to the model, not recalled from memory.
Regulatory compliance is disproportionately populated with this kind of work. Checking a constitution against the PBO Act's Second Schedule. Checking a financial report against the Companies Act's reporting requirements. Checking a board composition against the governance rules of a specific funding framework. For every one of these tasks, the compliance requirement is already specified, and — crucially, on §5's argument — both the requirement and the document being checked can be supplied directly to the model rather than left to recall. The work of encoding the requirement has already been done by the regulator. The remaining work — the work this paper argues AI is well-suited to — is applying that encoding systematically, and with the source text visibly in front of both the model and any human reviewing its output.
This is the category of work PBOMaster is built for. It is also, this paper argues, the category of work with the highest return on AI investment relative to risk: not because the AI is doing something sophisticated, but because the task's structure — grounded, bounded, checkable — is close to the condition under which the published evidence says language models perform best.
7. Objections I take seriously
"Flux has a commercial interest in this conclusion." Flux builds PBOMaster, the product this paper uses throughout as its working example of reliable, document-grounded legal AI, and sells software premised on the claim this paper argues for. This is a direct conflict. The honest mitigation is narrow: §5's grounding evidence (Dahl et al.'s hallucination findings, Shuster et al.'s retrieval-augmentation findings) is independent of Flux and of this paper's conclusions, and establishes that the grounding/recall distinction is real elsewhere. It does not establish that PBOMaster specifically achieves any particular accuracy rate — no external audit of PBOMaster exists, and this paper does not claim otherwise.
"Document-grounded tasks are not hallucination-proof, and this paper's framing risks implying more confidence than the evidence supports." This is fair, and §5 already concedes it rather than asserting a stronger claim: a model can misread or mis-summarise a provision that is genuinely present in the supplied document. The honest distinction is that this is a different, more checkable failure mode than inventing a case citation from nothing — a human reviewer re-reading the same supplied document can verify the model's claim against visible source text, which is not possible for Dahl et al.'s open-domain recall failures, where the "source" the model drew on cannot be inspected at all. But "more checkable" still requires the checking to actually happen, and §9's falsification conditions take this seriously rather than assuming it away.
"The six gap-patterns in §3 are Flux's own unaudited observations, not a peer-reviewed or third-party-verified finding." This is accurate and is stated directly in §3 itself. No independent researcher has audited PBOMaster's constitution analyser against a representative sample of Kenyan CBO constitutions with ground-truth labels. The patterns described are consistent with what the company building the tool has observed, which is evidence, but weaker evidence than an independent study would be.
"Even a reliable checklist-matching tool does not resolve the genuinely interpretive fraction of compliance work, and emphasising the reliable fraction risks understating how much of real compliance work is not this kind of task." §2 and §6 both state the interpretive fraction is real and excluded from this paper's reliability claim, but the objection is worth stating plainly rather than leaving implicit: this paper has not measured what share of total PBO Act compliance burden is procedural versus interpretive, beyond the qualitative observation (made in Africa's regulatory failure is not a law problem) that the procedural fraction appears large. That measurement is registered as open in Appendix B rather than assumed.
8. Where this paper's confidence stops
PBOMaster's constitution analyser has not been independently audited for accuracy; no such audit exists, and §3 and §7 both state directly that the gap-pattern findings are Flux's own unaudited observations.
It does not claim document-grounded compliance-checking is immune to hallucination or misreading; §5 and §7 both state a model can still misread a clause that is genuinely present, and this paper does not have a measured error rate for that specific failure mode.
It does not claim large language models can perform the genuinely interpretive fraction of legal compliance work reliably; §2, §4, and §6 all exclude that fraction from this paper's claim.
It does not claim to have measured what share of total PBO Act compliance work is procedural (checklist-amenable) versus interpretive; that is asserted qualitatively, not measured, and is registered as open in Appendix B.
And it does not claim the grounding/recall distinction in §5 guarantees any specific accuracy level for any specific tool; it explains why the error profile should differ, which is a narrower claim than a guarantee of reliability.
9. What would falsify this
Stated as commitments, in decreasing order of how much damage each would do to this paper's claims.
A rigorous, independent study finds LLM-based document compliance-checking tools have an error rate on document-grounded, checklist-structured tasks that is not meaningfully lower than the open-domain legal-recall hallucination rates Dahl et al. report (58-88%). That would show the grounding/recall distinction this paper treats as the explanatory mechanism is not actually predictive of real-world performance, and this paper's central claim would need to be withdrawn, not softened.
An independent audit of PBOMaster's constitution analyser, run against a representative sample of real Kenyan constitutions with expert-verified ground-truth labels, finds an error rate that would not be acceptable in a compliance workflow without full human re-review of every clause — i.e., that the specific tool this paper uses as its working example does not, in fact, deliver the reliability this paper's general argument predicts it should. That would not falsify the general grounding/recall mechanism (§5's evidence is independent of PBOMaster), but it would falsify the implicit claim that PBOMaster is a working instance of it.
A study finds the "misread a present clause" failure mode (§7, second objection) occurs at a rate high enough that the "more checkable" distinction from Dahl et al.'s fabricated-citation failures is not, in practice, actually easier for human reviewers to catch — e.g., reviewers given AI-flagged compliance gaps fail to notice misreadings at a rate comparable to how often they would catch a fabricated citation. That would weaken §5's core distinction between grounded and ungrounded error, even if the underlying mechanism (grounding reduces hallucination rate) remained true.
The procedural-versus-interpretive split this paper assumes (§6, §7) is shown, by a rigorous audit of the PBO Act's actual requirements, to be much narrower than this paper's qualitative impression — i.e., most of what looks like checklist-matching on the surface actually requires interpretive judgment PBOMaster's analyser is not equipped to exercise. That would shrink the practical scope of this paper's argument considerably, even without falsifying its technical claims about grounded-task reliability.
Appendix A — Evidence table
| # | Claim used in this paper | Value / finding | Source |
|---|---|---|---|
| 1 | GPT-4 performance on the Multistate Bar Examination | At or above passing threshold across all subject categories | Katz et al. 20241 |
| 2 | ChatGPT performance on four University of Minnesota law school exams | C+ average across 95 MCQ + 12 essay questions | Choi et al. 20232 |
| 3 | LLM hallucination rate on specific, verifiable open-domain legal factual questions | 58% (GPT-4), 88% (Llama 2) | Dahl et al. 20243 |
| 4 | Retrieval-augmented/grounded generation reduces hallucination relative to ungrounded generation | Measured across multiple conversation/QA benchmarks | Shuster et al. 20214 |
| 5 | Six recurring constitution-gap patterns found by PBOMaster's analyser | Internal, unaudited | Flux internal observation |
| 6 | Independent accuracy audit of PBOMaster's constitution analyser against ground-truth-labelled constitutions | — | — |
| 7 | Error rate of "misreads a clause genuinely present in a document" specifically, as distinct from fabrication | — | — |
| 8 | Share of total PBO Act compliance burden that is procedural (checklist-amenable) vs. interpretive | — | — |
Appendix B — Citation register
CITATION-NEEDED — independent PBOMaster accuracy audit. Appendix A row 6. §3, §7, and §9 all name this directly: no external party has audited the constitution analyser's accuracy against ground-truth-labelled real constitutions.
CITATION-NEEDED — misreading-of-present-clause error rate. Appendix A row 7. §5, §7, and §9 all depend on the claim that this failure mode is "more checkable" than fabrication, but no study measures how often it actually occurs or how reliably human reviewers catch it.
CITATION-NEEDED — procedural vs. interpretive share of compliance burden. Appendix A row 8. §6 and §7 both state the procedural fraction is qualitatively large based on PBOMaster's build experience; no rigorous audit of the PBO Act's requirements quantifies this split.
Daniel Martin Katz, Michael James Bommarito, Shang Gao, and Pablo Arredondo, "GPT-4 passes the bar exam," Philosophical Transactions of the Royal Society A 382, no. 2270 (2024): 20230254, doi:10.1098/rsta.2023.0254.↩
Jonathan H. Choi, Kristin E. Hickman, Amy B. Monahan, and Daniel Schwarcz, "ChatGPT Goes to Law School," Journal of Legal Education 71, no. 3 (2023): 387–400.↩
Matthew Dahl, Varun Magesh, Mirac Suzgun, and Daniel E. Ho, "Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models," Journal of Legal Analysis 16, no. 1 (2024): 64–93, doi:10.1093/jla/laae003.↩
Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston, "Retrieval Augmentation Reduces Hallucination in Conversation," in Findings of the Association for Computational Linguistics: EMNLP 2021 (2021), 3784–3803, doi:10.18653/v1/2021.findings-emnlp.320.↩
