Claude for Legal and Access to Justice: The Good, the Bad, and the Unknown

Last week, when Anthropic released its biggest legal push to date — more than 20 MCP connectors, 12 practice-area plugins,a and integrations with Microsoft Word, Outlook, Excel and PowerPoint — much of the coverage focused on what the announcement means for law firms, in-house teams, and the legal tech ecosystem. That is where the bulk of the announcement was directed. It is also where the money is.

But a less talked about part of the announcement was something potentially more impactful – the access-to-justice component. Anthropic named the Justice Technology Association and the Free Law Project as access-to-justice partners. It put Courtroom5, BoardWise, Descrybe, and CourtListener inside Claude as connectors that anyone can turn on. It offered discounted pricing to qualifying legal aid clinics, public defenders, and nonprofit legal services groups through its Claude for Nonprofits program. And it framed all of this as “working to put legal help within reach of people who can’t currently access it.”

That is a remarkable commitment from a leading AI lab. JTA executive director Maya Markovich said as much when she called it “the first time that a leading AI company is explicitly naming access to justice as a foundational pillar.”

But promises are easy. Outcomes are hard. I have been writing about access to justice and legal tech long enough to have seen a lot of promising tools land in the laps of self-represented litigants without moving the needle. So, the question worth asking is not whether the new Claude integrations could help — undoubtedly, they could — but whether they actually will. And, if so, who are they likely to help, and do they pose any potential dangers to the people they are designed to serve?

Here are my initial thoughts on the good, the bad, and the unknown.

The Good

Begin with the size of the problem. The Legal Services Corporation’s most recent Justice Gap Study found that 92% of the civil legal problems substantially affecting low-income Americans got no help or inadequate help. Roughly three-quarters of low-income households experience at least one civil legal problem a year. And the problem is not confined to the poor: an LSC Harris Poll commissioned for the organization’s 50th anniversary found that 59% of all Americans who had a civil legal matter in the prior three years did not seek out a lawyer at all.

That is the backdrop against which to evaluate the access-to-justice piece of last week’s announcement. Even modest improvements matter, and there are several aspects of this news that seem promising in their potential impact.

First, MCP rewires the foundational layer to make AI more reliable in legal situations. MCP – short for Model Context Protocol, the open standard Anthropic developed and which has spread quickly across legal tech – lets an AI assistant pull reliable data from an authoritative database at the moment of the query, rather than relying solely on whatever the model absorbed during training. For a self-represented litigant, that distinction could be the difference between an AI confidently fabricating a case citation and an AI retrieving an actual opinion from CourtListener. Free Law Project itself put it well: “A response built on verified CourtListener data is categorically different from one built on even the best model alone.” That is not marketing hype. It is a qualitative improvement.

Second, the pricing makes these tools accessible. CourtListener’s MCP is free with a free CourtListener account. Anyone with an internet connection now has access to the same primary-law backbone – millions of federal and state opinions, PACER dockets, citation networks, oral arguments, judge profiles – that, until not so long ago, was effectively gated by pricey legal research subscriptions. Courtroom5, which serves the roughly three out of four civil cases involving a self-represented litigant, can now plug its case-assessment, deadline-calculation, and procedural-guidance tools directly into the chat interface that millions of people are already using. BoardWise gives licensed professionals facing state board matters something they almost never have – structured help. Descrybe offers access to some 300 million structured U.S. primary law records.

Third, that old saw of tech, that “infrastructure beats apps,” is, in this context, right on point. For most of the last decade, the tech response to the justice gap has been to build standalone self-help tools — court websites, forms portals, chatbots, plain-language guides. Courtroom5 argues persuasively that the problem has never been a shortage of legal information; it has been a shortage of structured preparation. Meeting people inside the AI tool they are already using, with a justice-tech tool plugged in behind it, is closer to where help is actually needed than a self-help website that not everyone will find.

Fourth, the legal aid sector is ready for this. In an Everlaw survey last year, in which this blog partnered along with NLADA and Paladin, 74% of legal aid organizations reported using AI in some form — roughly double the rate of the broader profession — and 88% said they believed AI could meaningfully help close the justice gap. As Scheree Gilchrist of Legal Aid of North Carolina put it, “We will never be able to ‘lawyer ourselves’ out of this access-to-justice crisis. AI is a force multiplier to scale our services.” A discounted-pricing program for nonprofit legal services is something that sector can benefit from enormously.

So the upside is clear. But is there also a downside?

The Bad (Potentially)

The most obvious hazard of putting a general-purpose AI assistant in front of someone who is about to file something in court is hallucination. The numbers on this are sobering. Damien Charlotin’s running database of court filings containing AI-fabricated citations passed 1,000 cases earlier this year and is climbing rapidly. Self-represented litigants account for more documented hallucination incidents than licensed attorneys — 304 versus 219 in 2025, according to an analysis of Charlotin’s database done by DISCO. The Sixth Circuit imposed $30,000 in sanctions in Whiting v. City of Athens in March. An assistant U.S. attorney was reportedly fired. State courts are amending local rules. Bar associations are rewriting ethics guidance.

Grounding AI in real legal data, as MCP does, helps with this. It does not eliminate it. A pro se litigant who asks Claude — even Claude with CourtListener attached — to “write me a brief in support of my motion to dismiss” is going to get a brief. Whether every proposition in that brief is supported, whether every citation is real and on point, whether the procedural arguments are even available in that jurisdiction at that stage — those are judgments that the litigant is now responsible for, but possibly not equipped to make.

Free Law Project, to its credit, says so explicitly: “The MCP server is infrastructure. It connects a powerful AI model to a high-quality legal data source. … Claude is not a lawyer. Nothing produced through this integration is legal advice, and human judgment remains essential, especially for high-stakes decisions.”

Good on them for saying that. But how many users will get that message?

The second concern is that the appearance of help can obscure the substance of help. There is documented evidence that Courtroom5’s structured-preparation model produces results — the company reports that 73% of members who resolve their cases win or settle, and over 90% in eviction proceedings. Those numbers come from a curated platform with a curriculum, peer community, and human coaching. It is not a generic AI chat. The risk in the new architecture is that a user finds the brand-name AI assistant first, asks it a legal question, gets a confident-sounding answer, and never makes it to the structured tools — whether they be Courtroom5 or a lawyer-for-the-day program at the courthouse — that might have produced a better outcome.

Third, none of this resolves the unauthorized-practice-of-law question, which has hung over AI legal help since DoNotPay’s “robot lawyer” era. Anthropic’s approach — partnering with justice-tech tools rather than holding itself out as a legal service — is a sensible one. But the line between “information” and “advice” is fuzzier in a long, personalized chat than it ever was in a self-help website. State bar regulators might have a thing or two to say about that. Will that produce sensible regulatory sandboxes — as Utah, Arizona, and others have experimented with — or a defensive crackdown that protects the guild more than the public?

Fourth, there is a minor but real affordability question. While the CourtListener piece is genuinely free, the Anthropic piece on top of it is not. Use of these MCPs requires a Claude Pro subscription. Access to Claude for Nonprofits requires qualifying as a nonprofit. Even though a Claude Pro subscription is just $20 a month, even that might be a stretch for a self-represented litigant in a foreclosure or custody case. And there is always the risk of a two-tiered AI justice system, with better-resourced organizations running on Claude Enterprise, while unrepresented litigants rely on the free tier of whatever chatbot a friend told them about — which may or may not be the one with the verified primary-law connector turned on.

The Unknown

Beyond the potential upsides and downsides, there are also several open questions, the answers to which remain to be seen.

Will the integrations be used? Anthropic put the tools in the directory. That does not mean self-represented litigants will find them, turn them on, or know to use the CourtListener connector instead of asking Claude cold. Discovery and onboarding are real problems. For legal aid organizations, there may be a bandwidth challenge in actually deploying these in their workflows. Even if the infrastructure is there, the human and institutional plumbing may not be.

Do outcomes improve? We do not yet have data on whether self-represented litigants using grounded, MCP-connected AI assistants do measurably better in court, or with other legal problems, than those using ungrounded chatbots, or those using nothing. We need these studies, ideally conducted as collaborations among the justice-tech sector, legal aid and the courts themselves. Until we have them, any claims around access to justice are largely just hypotheses.

How will courts respond? Some judges are already imposing standing orders requiring litigants to disclose AI use. Some are amending local rules. A handful are imposing hallucination-related sanctions on pro se litigants in addition to attorneys. Some are openly embracing AI use – and are even using it themselves. The judiciary’s posture toward AI-assisted self-representation will shape whether the new tools actually expand access or set up users for an even harder fall.

Where is Anthropic in five years? Anthropic is not a justice-tech company. It is a frontier AI lab with an aggressive growth trajectory – including, we now know, in BigLaw and corporate legal. Its access-to-justice partnerships are only a small piece of a much larger commercial strategy. Whether the public-service aspect remains a foundational pillar may depend on what the company’s growth looks like over the next few years, who at the company champions this work, and whether the justice-tech partners are able to do their parts to keep the commitments alive.

What happens to the legal aid funding picture? Even the most capable AI tools are not a substitute for human lawyers handling the most serious civil matters — domestic violence, eviction defense, immigration, disability benefits. LSC is requesting $2.1 billion for FY2027, an increase over current levels. If the political response to, “AI can help with the justice gap,” is, “We can therefore cut legal aid funding,” then an unintended consequence of all this technology could be even less access for low-income Americans. Needless to say, if legislators start pushing that argument, then the access-to-justice community needs to push back on it hard.

Where That Leaves Me

The bottom line is that the access-to-justice component of last week’s announcement is unquestionably substantive, and that the infrastructure — open protocols, primary law accessible for free, justice-tech tools as first-class citizens of the Claude ecosystem — is genuinely promising. Maya Markovich is right that this is the first time a leading AI company has named access to justice as a foundational pillar, and that is worth marking.

But infrastructure is not outcomes. The same technology that lets CourtListener feed Claude verified case law could also let a self-represented litigant produce a beautifully formatted brief riddled with errors no judge will excuse. The potential good and the potential bad both turn on the same infrastructure.

Going forward, what I would like to see are three things:

• Data on outcomes for self-represented litigants using grounded AI versus everything else.
• Serious and open-minded engagement from courts and bar regulators on what is permitted and what is not.
• A sustained commitment from Anthropic to keep the access-to-justice partners on equal footing with the enterprise side of the house – and similar commitments from other major AI companies.

If those things happen, then we will someday look back on this announcement as the moment the access-to-justice talk in legal tech turned tangible. If they do not, then the announcement may just become a forgotten footnote in a much larger commercial story.

One thing for sure: That someday will be sooner rather than later.