Does AI Improve Access to Justice? | Rafael Research
Book a call
← Research<br>Applied AI Research
Does AI Improve Access to Justice?
What the controlled evidence shows about information,<br>representation, and outcomes — and why what people gained was<br>entry, not leverage.
Rafael Research · August 2026
There is a comfortable story about AI and access to justice.<br>Roughly 92% of the substantial civil legal problems of low-income<br>Americans get no or inadequate legal help.1<br>Lawyers are scarce and expensive. AI is abundant and nearly free.<br>Therefore AI closes the gap.
Every step in that argument is defensible. The conclusion still<br>does not follow, and we now have enough evidence to say why.
The evidence assembled here points somewhere more uncomfortable<br>than either the optimistic or the pessimistic version of this<br>debate. It is not that AI produces bad legal information —<br>increasingly it does not. It is that legal information is<br>not the only, and may not be the binding, constraint once a<br>litigant enters the system. The controlled interventions<br>we found that supply information have increased procedural<br>participation without producing detectable improvements in case<br>outcomes. And representation — the intervention everyone assumes<br>works — turns out to work conditionally, in ways suggesting that<br>lawyers are not primarily delivering legal knowledge either.
Meanwhile, two studies of federal court records document the same<br>pattern since generative AI became widely available:<br>substantially more self-represented filing, without evidence of<br>improved litigation outcomes.23, 24
Methodology
How these sources were selected. We ran a structured — not a<br>formal systematic-review-protocol — search across peer-reviewed<br>journals, working-paper repositories, research-institution<br>reports, court administrative data, and government statistics,<br>for evidence meeting three criteria: (1) it measures an outcome,<br>not a sentiment; (2) where it makes a causal claim, it uses a<br>design capable of supporting one, or the source itself disclaims<br>causal identification; (3) it is a primary source, or a faithful<br>report of one that we traced back to the original.
We separate findings into four evidence tiers, marked throughout:<br>primary source (government data, administrative<br>records, original survey instruments), peer-reviewed<br>research , working paper or preprint (not<br>yet peer-reviewed), and program or advocacy report<br>(typically no control group).
One further convention, which matters most in Parts 4 and 5:<br>claims that go beyond what a cited study states are marked<br>as interpretation. Everything else is a report of a source's own<br>finding. Where we draw an inference the underlying authors<br>did not draw, we say so and say why — and where an inference is<br>one of several readings the data supports, we give the<br>alternatives.
Two exclusions are worth stating. We excluded vendor benchmarks<br>and vendor-published performance claims entirely; they appear<br>nowhere in what follows. And we excluded program evaluations<br>reporting only volume served, satisfaction, or forms filed. Not<br>because those programs are unserious — because they do not answer<br>the question. That exclusion removed a substantial majority of the<br>published material we encountered.
Every figure in this article was verified against the primary<br>document, in its current version, in August 2026. Where a working<br>paper has been revised, we cite the current revision and say so.<br>Where a widely circulated figure could not be traced to a primary<br>source, we say that rather than repeating it — six such figures<br>appear in Part 1. Where we quote from a published abstract rather<br>than from body text, the source entry says so.
The funnel: what "improving access to justice" could mean
Almost every disagreement in this field is a disagreement about<br>which stage of a funnel is being measured. Setting the funnel out<br>first makes the rest of the evidence legible.
Problem → Recognition → Information → Action → Participation →<br>Legal position → Disposition → Real-world outcome → Welfare
Problem. A civil justice situation occurs.
Recognition. The person identifies it as something the legal system addresses.
Information. They learn what rules apply and what options exist.
Action. They file, answer, or otherwise engage.
Participation. They appear, respond to motions, meet deadlines.
Legal position. They have a defense, a claim, or a bargaining chip that carries weight.
Disposition. The case ends in a judgment, dismissal, or settlement.
Real-world outcome. They keep the housing, avoid the garnishment, obtain the status.
Welfare. They are better off, accounting for costs and any market response.
Representation, self-help materials, and AI are not stages — they<br>are interventions that enter the funnel at different points and<br>can only affect stages downstream of where they enter.
The funnel is an analytical decomposition, not a claim that every<br>case passes through these stages in sequence. Information can<br>produce...