reportAI in CourtsAI GovernanceAI Hallucinations
Consumer AI legal services and self-represented litigants: the services, the court record and the regulators
12 consumer legal-AI services on one schema, 41 self-represented rows in the incident tracker, two docket studies of AI pro se filings, and the regulators.
Edited and verified by Cognesio LLP
Researched with AI assistance · sources verified by Cognesio LLP · How this was made ↓
In the United States, “ai lawyer” is searched 9,900 times a month, the same volume as “legal ai”, and the people searching are not lawyers. This report is about what they find and what happens next. It reads twelve consumer-facing legal-AI services on one schema (operator, regulated status, what the service says it is, its disclaimer, whether it trains on users’ data, whether it publishes a price), sets out the 41 rows in the SafeLegalAI incident tracker in which a self-represented litigant put AI-generated material before a court in the United States, the United Kingdom, Canada, Australia, Ireland and New Zealand, and records what regulators, legislatures and appellate courts have done about services that answer legal questions without a lawyer.
It does not recommend a service, rate one, or advise anyone whether to use one. The site’s keyword registry excludes the “best AI lawyer” and “free AI lawyer app” searches as a target for that reason. What it can do is describe the record: the services’ own words, the courts’ own orders, and the two studies that have measured, from federal docket data, what AI-assisted self-representation looks like at scale.
Key findings
- Consumer demand is an order of magnitude larger than the site’s own lane and is need-shaped, not brand-shaped: on Semrush’s US database on 5 September 2026 the “ai lawyer” seed returns 1,907 keywords carrying 220,020 searches a month on Semrush’s totals panel (as recorded in the search-demand study; the published first page of 99 rows sums to 131,170), led by “ai lawyer” 9,900, “free legal advice online chat” 4,400, “free legal ai” 2,400, “talk to a lawyer for free 24/7 chat online” 1,600, “ai attorney” 1,600 and “ai lawyer free” 720. Only “paxton” (22,200) is a product name, and it is a professional tool.
- Of twelve consumer-facing services read, three are regulated legal providers (Garfield.Law, authorised by the Solicitors Regulation Authority; Lawhive, SRA-regulated with a US alternative business structure; Casium, combining software with licensed legal services), one is a public company that states it is not a law firm (LegalZoom), and eight are software services whose terms disclaim legal advice. Three state that they use customer content to develop or train AI (Rocket Lawyer’s Copilot terms, LegalZoom’s privacy policy, Casium on aggregated data); three say they do not; six do not say. Four publish a price.
- The one regulator finding on a consumer “AI lawyer” is the Federal Trade Commission’s: its final order against DoNotPay (vote 16 January 2025, announced 11 February 2025) records that the company “did not test whether its ‘AI lawyer’ operated to the level of a human lawyer” and “did not hire or retain attorneys to test the quality and accuracy” of its law-related features, requires USD 193,000 and notice to 2021 to 2023 subscribers, and bars unsubstantiated claims that the service performs like a lawyer.
- The incident tracker’s self-represented slice is 41 of 150 rows: 21 coded as litigant in person and 20 more whose orders describe the party as pro se or unrepresented. The slice grew from 3 rows in 2023 and 3 in 2024 to 7 in 2025 and 28 in the first eight months of 2026; it is 30 percent of the tracker’s 2026 rows. Outcomes are warnings (15), dismissals (11), sanctions (5), costs orders (3), a fine (1), other (5) and one pending; five rows carry a monetary figure, the largest CAD 20,000 (Arbuckle v Tanner, Nova Scotia Court of Appeal, 12 August 2026) and USD 10,000 (Kruse v Karlen, Missouri Court of Appeals, 13 February 2024). A tool is named in 13 of the 41, ChatGPT in 7 of the 13.
- Two studies of federal docket data measure the phenomenon the tracker samples. Shah and Levy (MIT and USC, March 2026), on more than 4.5 million non-prisoner federal civil cases from FY2005 to FY2026, find the pro se share rose “from a long-term steady-state average of 11% to 16.8% in FY2025”, docket entries per pro se case in the first 180 days up 158 percent on pre-AI means, and complaints flagged for AI-generated text rising “from essentially zero in the pre-AI period to more than 18% in 2026”. Cohen-Sasson (arXiv, 28 May 2026), on about 2.8 million filings, finds the pro se plaintiff rate rose from 11.33 percent to 16.94 percent, that 13.9 percent of post-2023 non-form pro se complaints are AI-flagged, and that AI-flagged complaints “are more likely to be dismissed and to terminate at earlier procedural phases”.
- Courts have twice ruled on whether a self-represented litigant’s AI work is discoverable, both times protecting it as work product: Warner v Gilbarco (E.D. Mich., 10 February 2026) and Morgan v V2X (D. Colo., 30 March 2026), the latter while ordering disclosure of the tool’s name and barring confidential material from public AI services. The contrary ruling, United States v Heppner (S.D.N.Y., 17 February 2026), concerned a represented criminal defendant, and the Colorado court distinguished it on that ground.
- The unauthorised-practice question has reached a federal appellate court and a legislature. In Upsolve v James (2d Cir., 9 September 2025) the Second Circuit held that New York’s unauthorised-practice statutes, as applied to trained non-lawyers giving individualised advice, “constitute a regulation of speech” but one that “is content neutral and thus subject only to intermediate scrutiny”, vacating the injunction that had protected the programme. New York Senate Bill S7263A would make an operator liable where a chatbot impersonates an attorney, provide that disclaimers cannot waive that liability, and let the Attorney General seek penalties of up to USD 15,000 a day; it was committed to the Rules Committee on 5 June 2026. Nippon Life v OpenAI (N.D. Ill., filed 4 March 2026) pleads that ChatGPT practised law; OpenAI’s reported response is that ChatGPT “is not a ‘person’” and “is incapable of practicing law within the meaning of the statute”.
- The best evidence on how people treat chatbot advice is a controlled experiment: Schneiders and colleagues (CHI 2025, 288 participants over three experiments) found that when the source was hidden, participants “showed a significantly higher willingness to rely on the ChatGPT advice” than on a lawyer’s, that when the source was known they were “equally willing to rely on both”, and that participants could tell the two apart only slightly better than chance (0.59, where 0.5 is guessing).
Why this demand exists and why the site excludes it
The consumer keywords are the largest block of legal-AI search demand in the registry and the only block the site declines to target. “ai lawyer” (9,900), “free legal advice online chat” (4,400) and “free legal ai” (2,400) together exceed the entire measured demand for court-rule, sanction and benchmark content that the search-demand report describes. Semrush’s intent labels put the “ai lawyer free” and “best ai for lawyers” sub-clusters at 64 to 100 percent commercial. A site that ranked for them would be a consumer guide, and the descriptive test in the site’s editorial standards forbids recommending a service.
What the site can do is what no consumer guide does: read the services’ terms against each other, count what happens when the output reaches a judge, and record what regulators have decided. That is the report’s scope. The “ai lawyer” head term and every “free” or “best” shape stay excluded in the keyword registry; the four advice-shaped queries (“ai legal advice” 390, “chatgpt legal advice” 480, “can ai give legal advice” 40, “legal ai chatbot”) are the ones this report answers descriptively.
Method and data
Services. The twelve are the eleven records in the SafeLegalAI tools directory (/tools/tools.json, 129 records as of 5 September 2026) that list consumers among their target users or sit in the consumer-legal category, plus one service from the first page of the US “legal ai” results (ailawyer.pro) read for this report. For each, the schema is the operator and its jurisdiction; whether the operator is a regulated legal provider; the service’s own description; the disclaimer in its terms; whether its terms or privacy policy permit customer content to be used to develop or train AI; and whether a price is published. Tool records were read on 3 September 2026 and the ailawyer.pro terms on 5 September. Eleven of the twelve are verified: false in the directory, meaning the site has recorded public pages and not confirmed them against a certificate or contract. No service is scored.
Court record. The incident collection, 150 rows (/tracker/incidents.json), filtered two ways: rows with actor: litigant-in-person (21) and rows with no actor code whose conduct or summary describes the party as pro se, self-represented, unrepresented or a litigant in person (20). The union of 41 is the “self-represented slice”. Outcome, monetary penalty, jurisdiction, tool named and year are read from the row fields. The two docket studies are read from the authors’ papers (Shah and Levy, March 2026, PDF; Cohen-Sasson, arXiv 2605.29493, 28 May 2026, abstract). The NBC News report of 8 October 2025 and the MIT Technology Review report of 4 June 2026 are cited as reports.
Regulators and courts. The FTC’s press release of 11 February 2025 and case page; the Second Circuit’s opinion in Upsolve v James (22-1345, 9 September 2025, read from an Institute for Justice mirror of the slip opinion, the court’s own link having moved); the Alabama State Bar’s Formal Opinion 2026-01 (PDF); New York S7263A on the Senate’s site; the Morgan v V2X order (Dkt. 65) and the Warner and Heppner opinions as read for the chat-logs report; the Nippon Life v OpenAI docket (1:26-cv-02448, N.D. Ill.) and OpenAI’s motion as reported by Bloomberg Law; the University of Southampton’s announcement of the CHI 2025 paper (doi 10.1145/3706598.3713470); OpenAI’s usage policy line as reported by Legal IT Insider on 3 November 2025, the policy page itself returning 403 to the site’s fetch on 5 September 2026.
Limits. The services are a selection, not a market survey; the tools directory was built for professional products and its consumer coverage is incomplete. The self-represented slice depends on courts publishing orders against unrepresented parties, which many do not, and the 20 uncoded rows were identified by wording, not by a field. The two docket studies use different samples, periods and AI-text detectors and their figures are not interchangeable. Individuals are not named in this report except as they appear in published decisions and orders.
The demand
The “ai lawyer” seed on Semrush’s US database (5 September 2026, first page published as ai_lawyer_us.csv): 1,907 keywords, 220,020 searches a month and average difficulty 45 on the totals panel, as recorded in the search-demand study; the saved 99 rows sum to 131,170. The 15 largest that are not professional terms:
| Keyword | Volume/month | KD | Intent (Semrush) |
|---|---|---|---|
| ai lawyer | 9,900 | 64 | informational |
| free legal advice online chat | 4,400 | 62 | informational |
| free legal ai | 2,400 | 63 | informational |
| talk to a lawyer for free 24/7 chat online | 1,600 | 50 | informational |
| ai attorney | 1,600 | 61 | informational |
| legal answers | 1,600 | 50 | informational |
| ai lawyers | 1,000 | 62 | informational |
| online lawyer | 1,000 | 94 | informational |
| free legal advice online | 880 | 61 | informational |
| legal advice online | 880 | 65 | informational |
| ai lawyer free | 720 | 60 | informational |
| consult a lawyer online free | 720 | 63 | informational |
| free lawyer advice online | 720 | 61 | informational |
| speak to a lawyer free online | 720 | 52 | informational |
| ai atodmey (sic) | 590 | 57 | informational |
The pattern is a person with a legal problem and no budget. “Free” appears in seven of the fifteen; “24/7” and “chat” describe the interface they expect. India’s database shows the same shape at smaller scale (“free legal advice online in india” 880; “law bot” and “legal bot” 480 each; “free legal advice chat” 210), as the India report records.
The services
| Service | Operator, jurisdiction | Regulated legal provider | What it says it is | Disclaimer in terms | Customer content used to develop or train AI | Price published |
|---|---|---|---|---|---|---|
| DoNotPay | DoNotPay, Inc., US | No | Consumer self-help app with more than 100 AI tools; formerly “the world’s first robot lawyer” | Subject to the FTC order of 2025 barring unsubstantiated claims that it performs like a lawyer | Not stated in the record | No (subscription) |
| Rocket Lawyer (Rocket Copilot) | Rocket Lawyer Incorporated, US | No (attorney escalation offered) | Free AI assistant for legal questions and contract review | Outputs “not guaranteed to be error free or accurate” | Yes: Copilot terms permit use of content to develop and improve the service | Free |
| LegalZoom AI (Doc Assist) | LegalZoom.com, Inc., US | No; states it is not a law firm | Consumer and SMB platform with AI document assistance and attorney-curated review | Not a law firm | Yes: privacy policy permits personal-data use for ML and AI development | Usage-based, not listed |
| Lawhive | Lawhive Ltd, UK (US ABS in 35 states) | Yes: SRA-regulated | Online legal services with vetted solicitors and a proprietary AI platform | Not recorded | Not stated | Usage-based, not listed |
| Garfield.Law | Garfield.Law Limited, UK | Yes: SRA-authorised (2025) | Law firm delivering small-claims debt recovery up to £10,000 solely through AI | Not recorded | No | Yes: £2 initial letter; £7.50 letter before action |
| Genie AI | Genie AI Ltd, UK | No | Self-serve contract drafting and review for SMEs and small teams | Not recorded | No; ISO 27001; states no SOC 2 | Yes: free; Pro USD 75/month; Business USD 320/month |
| Casium | Casium, Inc., US | Combines software with licensed legal services | AI-native business immigration platform | Not recorded | Yes: trains on aggregated, de-identified data | No (enterprise) |
| Lawfully | Lawfully, Inc., US | No | Immigration case tracker with AI analysis of 6.6M+ agency cases | Not recorded | Not stated as training: privacy policy says data is used for “research, analytics, or improving our services”; anonymised case data retained; SOC 2 stated | Freemium, not listed |
| descrybe.ai | Descrybe, LLC, US | No | Case-law research with plain-language summaries | Not recorded | Not stated | Yes: USD 50/month; USD 25/month Legal Engine |
| Caseway (Casey) | Caseway, Canada | No | Legal research AI with cited answers; uploads auto-deleted within 24 hours | Not recorded | No | Freemium, not listed |
| Lawpath (Atlas AI) | Lawpath Operations Pty Ltd, Australia | No (lawyer access offered) | SMB legal platform with an AI assistant for legal, tax and compliance questions | Not recorded | Not stated | Per-seat, not listed |
| AI Lawyer (ailawyer.pro) | AI Lawtech Sp. z o.o., Poland | No | ”your personal legal AI assistant" | "not a substitute for professional legal advice … should not be considered legal advice”; no guarantee of “accuracy, completeness, or timeliness” | Not stated on the terms page read | Free trial; no price shown |
Three things the table shows. First, the regulated providers are the exception: two of twelve are law firms or regulated firms in England and Wales, and one bundles licensed lawyers in the United States; the rest are software. Second, the training question divides them: three say they use customer content to develop or train AI, three say they do not, six are silent, and the one whose terms were read in full for this report (ailawyer.pro) disclaims legal advice and accuracy in the same paragraph. Third, price is the least published field: four of twelve state one.
The general-purpose chatbots that the tracker’s self-represented rows actually name are not in the table because they do not present as legal services. OpenAI’s usage policies, as reported by Legal IT Insider on 3 November 2025, bar use of ChatGPT for the “provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional”; OpenAI’s health lead said publicly that the line was not new and that model behaviour had not changed. The line is a term of use, not a technical limit, and the court record below is the evidence that it is not one.
The court record
Forty-one of the tracker’s 150 rows concern a self-represented litigant. The count by year, against all rows:
| Year | Self-represented rows | All rows | Share |
|---|---|---|---|
| 2023 | 3 | 6 | 50% |
| 2024 | 3 | 11 | 27% |
| 2025 | 7 | 40 | 18% |
| 2026 (to 3 Sep) | 28 | 93 | 30% |
| Total | 41 | 150 | 27% |
By jurisdiction: 23 United States (14 state, 9 federal), 9 United Kingdom, 3 Canada, 3 Australia, 2 Ireland, 1 New Zealand. By outcome: warning 15, dismissal 11, other 5, sanctions 5, costs order 3, fine 1, pending 1. The five monetary rows:
| Date | Matter | Court | Amount |
|---|---|---|---|
| 13 Feb 2024 | Kruse v Karlen | Missouri Court of Appeals | USD 10,000 (22 fictitious citations in a brief) |
| 1 Oct 2025 | Specter Aviation v Laprade | Quebec | CAD 5,000 (reported) |
| 22 Jul 2026 | Asif v Minister for Immigration | Australia | AUD 9,097.93 (ChatGPT admitted) |
| 11 Aug 2026 | AP v SK | Quebec Court of Appeal | CAD 1,000 (ChatGPT conceded) |
| 12 Aug 2026 | Arbuckle v Tanner | Nova Scotia Court of Appeal | CAD 20,000 lump-sum costs |
Two features distinguish the self-represented slice from the lawyer rows in the sanctions ledger. Money is rarer: five of 41 rows carry a figure, against 45 of 150 across the tracker, and the largest self-represented figure, CAD 20,000, sits below the ledger’s upper band of lawyer orders (no currency conversion is applied here). And the court’s response is more often a warning or a dismissal on the merits than a sanction: 26 of 41 rows end in a warning or dismissal, and two rows carry a regulator referral: in Lodhia the court referred the litigant’s direct-access counsel to the Bar Standards Board and the litigant himself to the Attorney General for possible contempt proceedings, and in Varma the court referred to the State Bar of California. A tool is named in 13 rows; ChatGPT in seven of those, generative AI unspecified in five, an AI avatar in one.
The most recent rows show the shape of the 2026 wave: in the last week of August 2026 alone the tracker holds warnings from federal courts in Ashmeade v Amazon, Johnson v Nationstar, Potterf v Wessels and Turgeon v FHLMC, state courts in Salloum, Noble and State v Presti, and a sanction in Varma v Bank of New York Mellon. Most are warnings on a first occurrence; the orders that impose money are appellate and follow a second or egregious filing.
The two docket studies give the denominator the tracker lacks. Shah and Levy, using administrative records on more than 4.5 million non-prisoner federal civil cases and 46 million PACER entries, report the pro se case share “rising from a long-term steady-state average of 11% to 16.8% in FY2025”, concentrated in “case types characterized by formulaic document production”, with docket entries per pro se case in the first 180 days “up 158% from pre-AI means to 2025”, and, from a 1,600-complaint sample, AI-flagged complaints rising “from essentially zero in the pre-AI period to more than 18% in 2026”. Cohen-Sasson’s separate analysis of about 2.8 million filings puts the rise at 11.33 to 16.94 percent, finds 13.9 percent of post-2023 non-form pro se complaints AI-flagged, “disproportionately associated with first-time rather than repeat filers”, and finds “no evidence of improved win rates; in fact, AI-flagged complaints are more likely to be dismissed and to terminate at earlier procedural phases”. NBC News reported on 8 October 2025 that the public database maintained by Damien Charlotin then held 282 US cases and more than 130 elsewhere, with acceleration from spring 2025; the tracker’s own 2026 curve matches.
The courts have also had to decide what a self-represented litigant’s AI use is, as evidence. In Warner v Gilbarco (E.D. Mich., 10 February 2026) a pro se plaintiff’s ChatGPT queries and responses were held to be work product and the defendants’ motion to compel them denied. In Morgan v V2X (D. Colo., 30 March 2026) the court agreed, held that the litigant “is simultaneously the party and the advocate”, found that public AI tools’ data collection “does not eliminate all expectations of privacy or automatically waive protections”, but ordered the plaintiff to name any AI platform he had used with confidential discovery and amended the protective order to keep such material out of “any mainstream AI tool like standard ChatGPT, Claude, Gemini, or similar platforms” without contractual safeguards. The chat-logs report sets both against the contrary criminal-case ruling in Heppner.
The regulators’ position
Advertising. The Federal Trade Commission’s action against DoNotPay is the only regulator finding in the record about a consumer AI legal service. The complaint was announced on 25 September 2024 in the Commission’s “Operation AI Comply”; the final order was approved by a 5-0 vote on 16 January 2025 and announced on 11 February 2025. The Commission’s statement records that DoNotPay “did not test whether its ‘AI lawyer’ operated to the level of a human lawyer when generating legal documents and giving advice, and the company did not hire or retain attorneys to test the quality and accuracy of its service’s law-related features”. The order requires USD 193,000, notice to consumers who subscribed between 2021 and 2023, and bars advertising that the service performs like a lawyer without evidence. It is a consumer-protection order about claims, not a finding on the practice of law.
Unauthorised practice. Upsolve v James is the appellate decision. A non-profit trained non-lawyer “Justice Advocates” to help low-income New Yorkers complete a state answer form in debt-collection cases; all parties agreed the activity would breach New York’s unauthorised-practice statutes; the district court enjoined enforcement as a content-based restriction failing strict scrutiny. On 9 September 2025 the Second Circuit agreed that the statutes “as applied to Plaintiffs, constitute a regulation of speech” but held “that such regulation is content neutral and thus subject only to intermediate scrutiny”, and vacated and remanded. The case concerns human advisers, not software; it is in this report because it fixes, for the Second Circuit, the constitutional test any challenge to applying unauthorised-practice law to a chatbot would face.
The bar-side guidance treats a chatbot as the lawyer’s responsibility. The Alabama State Bar’s Formal Opinion 2026-01 gives the example of “a law firm [that] deploys a chatbot on its website that answers prospective clients’ legal questions in real-time” and says that “if the chatbot provides specific legal advice without flagging that it is not an attorney and without attorney review of its responses, this may constitute UPL”, with the deploying lawyer answerable under Rules 5.5, 5.3 and 8.4; firms “should avoid marketing AI tools in a manner suggesting that the software itself is ‘the lawyer’”. It says nothing about services no lawyer deploys, which is where the consumer demand goes.
Legislation. New York’s S7263A (Senator Gonzalez), with Assembly companion A06545, would make the person who “owns, operates or deploys” a chatbot liable where it impersonates an attorney “in a manner that, if done by a natural person … would violate” the Judiciary Law’s prohibition on practising or appearing as an attorney without admission; the operator “is deemed to know that a chatbot cannot lawfully practice a covered profession”; disclaimers “cannot waive liability”; and the Attorney General could seek penalties of up to USD 15,000 a day. The Senate page records the bill committed to the Rules Committee on 5 June 2026. It has not passed.
Litigation against a model provider. Nippon Life Insurance Company of America v OpenAI Foundation, filed in the Northern District of Illinois on 4 March 2026 (1:26-cv-02448), pleads, as reported, that ChatGPT helped a former claimant draft 44 post-settlement filings including a fabricated citation and thereby engaged in the unlicensed practice of law, tortious interference and abuse of process. Bloomberg Law reported OpenAI’s motion to dismiss as arguing that “ChatGPT is not a ‘person,’ but a tool that relies on statistics to predict the most appropriate sequence of words based on its training” and “is incapable of practicing law within the meaning of the statute”. No ruling had been reported on 5 September 2026. The complaint and motion were not read for this report; the docket is on CourtListener.
What people do with the advice. The Southampton, Nottingham and Antwerp study presented at CHI 2025 gave 288 participants, across three experiments, legal advice on 18 hypothetical cases in traffic, planning and property law, written by ChatGPT and by lawyers. When the source was hidden, participants “showed a significantly higher willingness to rely on the ChatGPT advice”; when told the source, they “were equally willing to rely on both sources”; asked to tell the sources apart, they scored 0.59 where 0.5 is chance. The authors point to the length and complexity of the two kinds of advice as candidate explanations. It is one study with a British sample, and it is the only controlled evidence in the record.
Timeline
| Date | Event | Source |
|---|---|---|
| 6 Oct 2023 | Thomas v Pangburn (S.D. Ga.): first self-represented row in the tracker; dismissal | Tracker |
| 4 Dec 2023 | Harber v HMRC: first UK case of AI-fabricated authorities, litigant in person | Tracker |
| 13 Feb 2024 | Kruse v Karlen: USD 10,000 against a pro se appellant, 22 fictitious citations | Tracker |
| 25 Sep 2024 | FTC complaint against DoNotPay announced (Operation AI Comply) | FTC |
| 16 Jan 2025 | FTC final order approved 5-0; announced 11 Feb 2025 | FTC |
| Apr 2025 | Southampton study at CHI 2025 | University announcement |
| 9 Sep 2025 | Upsolve v James: UPL statutes content neutral, intermediate scrutiny; injunction vacated | 2d Cir. slip opinion |
| 8 Oct 2025 | NBC News: 282 US AI cases in the public database; pro se litigants “starting to win” | Report |
| 29 Oct 2025 | OpenAI usage-policy line on licensed advice reported (3 Nov) | Report |
| 10 Feb 2026 | Warner v Gilbarco: pro se ChatGPT materials are work product | Primary |
| 4 Mar 2026 | Nippon Life v OpenAI filed (N.D. Ill.) | Docket |
| Mar 2026 | Shah and Levy paper: pro se share 11% to 16.8% in FY2025 | Paper |
| 30 Mar 2026 | Morgan v V2X: work product for pro se AI use; tool name to be disclosed | Primary |
| May 2026 | Alabama State Bar Formal Opinion 2026-01 (chatbot UPL example) | Primary |
| 28 May 2026 | Cohen-Sasson paper on arXiv | Paper |
| 5 Jun 2026 | New York S7263A committed to Rules | Senate page |
| 22 Jul to 31 Aug 2026 | 17 self-represented rows added to the tracker in six weeks | Tracker |
What to watch
The Nippon Life motion is the first occasion on which a court will be asked whether a general-purpose chatbot’s operator can be liable for the unlicensed practice of law; the answer, either way, will be the first primary document on the question. New York’s S7263A would settle it by statute for that state if enacted. The Second Circuit’s remand in Upsolve will produce the first application of intermediate scrutiny to unauthorised-practice law, with software challenges to follow the human one.
In the tracker, the self-represented share of 2026 rows (30 percent to 3 September) is the number to re-read quarterly against Shah and Levy’s 16.8 percent pro se share and 18 percent AI-flagged rate: if courts keep publishing warnings on first occurrence, the slice will grow faster than the sanctions. Whether any appellate court fines a self-represented litigant at the level of the lawyer orders (the USD 10,000 in Kruse remains the US maximum in the slice) is the other marker.
On the services side, the fields to re-read are the training clauses and the price. The FTC order is the only external check on any of the twelve; the SRA’s authorisation of an AI-only firm (Garfield.Law) is the only instance in the record of a legal regulator licensing the model rather than the marketing.
Three sentences journalists can quote
“AI lawyer” is searched as often in the United States as “legal ai”, and the people searching want free advice through a chat window; of twelve consumer legal-AI services read, two are regulated law firms, three say they use users’ content to develop AI, and four publish a price.
Self-represented litigants account for 41 of the 150 rows in the SafeLegalAI incident tracker and 30 percent of its 2026 rows; courts mostly warn or dismiss, five rows carry a monetary figure, and the largest is CAD 20,000.
Federal docket studies put the pro se share of US civil cases at 16.8 percent in fiscal 2025 against a long-run 11 percent, with AI-flagged complaints rising from near zero to 18 percent in 2026 and more likely to be dismissed early.
Appendix A: data tables
A1. Self-represented slice, 41 rows (tracker, 5 September 2026)
| Field | Values |
|---|---|
| Coding | 21 actor: litigant-in-person; 20 uncoded rows whose order describes the party as pro se, self-represented, unrepresented or a litigant in person |
| Jurisdiction | US state 14; US federal 9; UK 9; Canada 3; Australia 3; Ireland 2; New Zealand 1 |
| Outcome | warning 15; dismissal 11; other 5; sanctions 5; costs order 3; fine 1; pending 1 |
| Monetary rows | 5: USD 10,000; CAD 20,000; AUD 9,097.93; CAD 5,000; CAD 1,000 |
| Tool named | 13 (ChatGPT 7; generative AI unspecified 5; AI avatar 1); not named 28 |
| Regulator referral | 2 rows (Lodhia: counsel to the Bar Standards Board, the litigant to the Attorney General; Varma: State Bar of California) |
| Year | 2023: 3; 2024: 3; 2025: 7; 2026 to 3 Sep: 28 |
A2. The 20 uncoded rows counted by wording
Harber v HMRC (UK, 4 Dec 2023); Kruse v Karlen (Mo. App., 13 Feb 2024); Al-Hamim v Star Hearthstone (Colo. App., 26 Dec 2024); Dewald (N.Y. App. Div., 26 Mar 2025); An v Archblock (Del. Ch., 4 Apr 2025); HMRC v Gunnarsson (UK UT, 23 Jul 2025); Father v Mother [2025] EWHC 2135 (Fam) (30 Jul 2025); Williams v Kirch (Ind. App., 18 Aug 2025); Cingel v Ferreri (Ind. App., 19 Sep 2025); Specter Aviation v Laprade (Que., 1 Oct 2025); Huish v HMRC (UK FTT, 16 Jan 2026); Folarin v Immigration Services Commissioner (UK FTT, 29 Jan 2026); Jones v Family Court Whangarei (NZ SC, 11 Feb 2026); Green v Imprint Creative (UK ET, 9 Mar 2026); Von Geitz v Kelly (Ir. CA, 16 Mar 2026); Guerin v O’Doherty (Ir. CA, 26 Mar 2026); Tuyunuklu v Akmal (UK UT, 5 May 2026); Rafique v HMRC (UK FTT, 6 May 2026); Gleason v Marcus (Fla. 2d DCA, 13 May 2026); Dineen-Shibata v Kotchka (Ariz. App., 15 Jul 2026). Each is linked from the tracker.
A3. Service register fields, summary
| Field | Yes | No | Not stated |
|---|---|---|---|
| Regulated legal provider | 3 | 9 | 0 |
| Customer content used to develop or train AI | 3 | 3 | 6 |
| Price published | 4 | 8 | 0 |
| Verified by the site against a primary document | 1 (Genie AI) | 11 | 0 |
Appendix B: sources
- Federal Trade Commission, FTC Finalizes Order with DoNotPay (11 Feb 2025): ftc.gov; case page: ftc.gov; Operation AI Comply (25 Sep 2024): ftc.gov
- Upsolve, Inc. v James, No. 22-1345 (2d Cir. 9 Sep 2025), slip opinion: Institute for Justice mirror; Justia (archive)
- Alabama State Bar, Formal Opinion 2026-01, Artificial Intelligence Use: Best Practices Under Existing Professional Conduct Rules: PDF
- New York Senate, S7263A (2025-2026), Imposes liability for damages caused by a chatbot impersonating certain licensed professionals: nysenate.gov (archive)
- Nippon Life Insurance Company of America v OpenAI Foundation, No. 1:26-cv-02448 (N.D. Ill., filed 4 Mar 2026): CourtListener docket; OpenAI’s motion as reported: Bloomberg Law
- Morgan v V2X, Inc., No. 1:25-cv-01991, Dkt. 65 (D. Colo. 30 Mar 2026): CourtListener RECAP PDF; Warner v Gilbarco, 2026 WL 373043 (E.D. Mich. 10 Feb 2026): PDF
- Shah and Levy, Access to Justice in the Age of AI: Evidence from U.S. Federal Courts (MIT and USC, March 2026): PDF; Cohen-Sasson, The New Pro Se: Generative AI and the Surge in Federal Civil Self-Representation, arXiv 2605.29493 (28 May 2026): arxiv.org
- Schneiders et al., CHI 2025 (doi 10.1145/3706598.3713470), as announced: University of Southampton, April 2025
- NBC News, People are using ChatGPT as a lawyer in court. Some are winning (8 Oct 2025): nbcnews.com; MIT Technology Review, How courts are coping with a flood of AI-generated lawsuits (4 Jun 2026): technologyreview.com
- OpenAI usage-policy line, as reported: Legal IT Insider, 3 Nov 2025; policy page (403 to the site’s fetch on 5 Sep 2026): openai.com (archive)
- Service pages as recorded in the tools directory (read 3 Sep 2026) and ailawyer.pro terms (read 5 Sep 2026): DoNotPay terms; Rocket Copilot terms; LegalZoom privacy; Lawhive; Garfield.Law; Genie AI; Casium privacy; Lawfully privacy; descrybe pricing; Caseway privacy; Lawpath; AI Lawyer terms
- Note: law.justia.com, nysenate.gov and openai.com return 403 to the site’s link checker; the Senate page was read through a fetch tool on 5 Sep 2026 and archive links are given.
- SafeLegalAI datasets: /tracker/incidents.json, /tools/tools.json, as of 5 September 2026; Semrush US pull: /data/search-demand-2026-09/ai_lawyer_us.csv
Appendix C: changes to this report
None since first publication.