Plain-Language Legal First Aid: Staff Tools Exist, Tested Public Advice Does Not

The problem

Low-income Americans received no or inadequate legal help for 92% of the civil legal problems that affected them substantially, according to the Legal Services Corporation (LSC), the US federal funder of civil legal aid. The figure comes from a household survey of 2021, published in April 2022, so it is more than three years old (survey, self-reported)¹. In the same study, 74% of low-income households had at least one such problem in the past year, and where people did not seek help, the belief that cost would rule it out was the commonest reason¹. LSC-funded organisations reported 1.9 million problems brought to them in 2021 and an estimated 1.4 million for which they could not provide any or enough help. For one-half of the eligible problems they could give no help at all (2021, grantee-reported)¹.

Starting point. The anchor paper in AI in Practice, published on 2 October 2026, examined generative AI as used by lawyers². It reported that an independent test of legal-specific research tools found hallucination rates of 17% to 33% in 2024 product versions (measured in evaluation)³. It also reported that regulators leave responsibility with the lawyer, and that the Solicitors Regulation Authority warns that tools may lack the safeguards needed for confidentiality⁴. In that setting a trained professional is expected to check the output.

This paper adds a different setting: the same capability aimed at a tenant or worker with no lawyer, who cannot check the output. The evidence found is almost all about US housing. No study of wage or employment problems was found, and non-US evidence comes only through the anchor.

What already works

Tools for staff. A field study gave 91 legal aid professionals up to two months of access to paid generative AI tools. Afterwards 90% reported increased productivity and 75% intended to keep using the tools (self-reported, abstract, baseline survey in fall 2023). Participants managed risk by keeping to lower-risk tasks: summarising documents, preliminary or confirmatory research, first drafts, and translation into plainer language⁵. In plain terms, the use that has taken hold sits with staff who can check the result.

Prototypes for housing. A Stanford studio report of August 2026 describes five prototypes built with three legal aid attorney partners, mostly on housing. They included a phone intake line, drafting of reasonable-accommodation letters, spotting defects in eviction notices, and help for tenants reopening default eviction judgments. Attorneys were reported to spend 30 to 60 minutes on each accommodation letter (partner-reported). The report's design lesson is to lock legal authority in attorney-reviewed templates, never model-generated, and to trace every factual claim to a fact the user supplied⁶.

Rules plus a language model. A preprint of February 2025 describes LogicLease, which uses a language model only to pull facts out of a plain description of a New York landlord-tenant dispute. A separate set of hand-coded rules, taken from the state Attorney General's tenants' rights guide, then decides the legal outcome. The authors report 100% accuracy on 10 cases, mixing condensed real cases and cases they wrote, and report that ChatGPT and Gemini reached the wrong conclusion in 3 of the 10 (measured without independent review)⁷. The sample is far too small to show reliability, but the design separates reading the story from applying the law.

Intake. A Stanford prototype built with the Legal Aid Society of San Bernardino asks eligibility questions and drafts a summary for staff. Its instructions state that it does not give legal advice or make final decisions, and it has not been deployed. The society reported assisting over 1,200 households with eviction prevention and landlord-tenant support in the first half of 2024 (operator-reported)⁸.

What does not exist. No public-facing tool for tenants or workers was found that has been evaluated against real users and outcomes.

The possibility

One way this could work is a service run by a legal aid group, court self-help centre, tenants' association or union. A person describes a notice or a dispute in their own words and language. The service reads attached documents, identifies the type of problem, and applies a library of rules written and approved by attorneys. It could then return a plain-language account of what the document means, which dates and papers matter, a draft letter from an approved template, and a route to a human adviser. It would offer information and triage, not individual legal advice.

Only part of this depends on AI. Language models could read untidy narratives and photographs of notices, translate, and rewrite legal wording plainly, which are the tasks the evidence above supports. The rules library, deadlines, templates, referral directory, funding and the human advisers are ordinary content, software and staffing. If the honest conclusion were that AI supplies the interface and little else, the design in S7 and S6 points the same way: the law sits outside the model.

What would have to be true

1. The legal content would be fixed in attorney-reviewed material, with the model limited to reading and rewording. This is the design lesson of the Stanford studio⁶.
2. Testing would use real, representative situations. The studio report states that real client data was largely unavailable and that synthetic personas were often too generic or non-representative⁶.
3. Silent failures would have to be caught. One prototype could drop a key fact, an emotional-support animal request, when a long conversation exceeded what the model could hold, and still produce fluent, professional output⁶.
4. Claims about the service would have to match its testing, which is the standard the FTC applied in the DoNotPay case⁹.
5. A human route would need capacity to absorb referrals. LSC-funded organisations already gave no help at all on one-half of eligible problems¹.
6. Evidence would have to extend to worker problems, where none was found.

What could stop it

What the evidence shows

In February 2025 the FTC finalised an order against DoNotPay, which had promoted itself as the world's first robot lawyer. The FTC said the company did not test whether its AI operated at the level of a human lawyer and did not hire or retain attorneys to test its accuracy. The order required $193,000 in monetary relief and bars such claims without sufficient evidence⁹.

A Stanford studio report of August 2026 states that none of its housing and eviction prototypes was ready for wide release⁶. The only accuracy figure it gives is for an early eviction-notice detector, which reached a ceiling of around 20% without a stated test set⁶. The tested LogicLease design rests on 10 cases, which the authors judged themselves⁷.

For lawyers' tools, independent testing found hallucination rates of 17% to 33% in 2024 versions³, and the Solicitors Regulation Authority warns that tools may lack safeguards for confidentiality⁴. The field study's authors reported that concerns remained among participants⁵.

Reasoning

The challenge question is who bears the risk. A lawyer who relies on a faulty tool faces professional consequences and can check. If the same tool were used by a tenant, the cost of an error would fall on the person least able to detect it, and a fluent but wrong answer could look as reliable as a correct one. The staff tools that work today suit the lower-risk pattern in S5 precisely because a professional stands between the model and the client.

Incentives could also work against caution. A provider that sells to funders or the public could gain from broad claims, which is the pattern in S9. A legal aid group short of capacity could gain from deflecting callers to a bot, and the benefit would depend on whether the human route behind it exists.

Simpler options might do much of the job. Plain-language guides, hotlines, form-filling wizards and more funded advisers do not depend on a model. No comparison between these and an AI service was found, so the added value of AI remains unmeasured.

Signs to watch

• A published pilot evaluation, with real users, of a public-facing tenant or worker tool. The Stanford studio produced draft pilot plans, which could lead to one⁶.
• An independent test of such a tool that measures outcomes, such as deadlines met or cases reopened, rather than answer accuracy alone.
• A larger test set and released code for LogicLease, which the authors list as future work⁷.
• A newer LSC Justice Gap measurement, to update the 2021 survey data¹.
• Further regulator action on claims made for AI legal services, following S9.