Datasets
Every table we maintain, published as data
SafeLegalAI is a growing set of maintained datasets, not three pages. Each one below answers a question, links every record to its official source, states when it was last checked, and downloads under CC BY 4.0 for reuse and citation. Every table is mirrored on Hugging Face and GitHub (synced on every content update; this site stays canonical), and the pipelines that collect the newer datasets are open source in the same GitHub repositories. More are on the way — the next candidates are listed at the bottom.
Hugging Faceas of 2026-10-07 · researched with AI assistance · verified by SafeLegalAI · published by SafeLegalAI
What went wrong when AI met the courtroom?
Legal AI Incident Tracker
173incidents · 15 jurisdictions · 51 regulatory outcomes
Every court case where AI misuse reached a judgment: court, date, conduct, outcome, penalty, and what the regulator did next.
What do the rules say, country by country?
Regulatory map
130countries and entities · 102 with at least one rule · 20 categories
Countries scored across 20 categories of rule (disclosure, verification, judges' use, evidence, client data…), each cell with status, note, sources and the date last checked.
Which official texts exist, and how have they changed?
Official documents, versioned
522documents · 573 versions tracked
Every court rule, practice direction, ethics opinion, statute and guidance document on AI in legal practice, with its full version history and the date we last checked it.
What does the EU AI Act require, of whom, from when?
EU AI Act, structured
677coded obligations · 119 articles · 198 reach legal practice
Regulation (EU) 2024/1689 as data: every article as a page, every duty coded by actor, risk tier, application date and fine tier, with the provisions that reach law firms, courts and legal-AI vendors marked. Text © EU; coding CC BY 4.0.
What does each legal tech tool actually document?
Legal tech tools
130tools · 77 publish a no-training commitment
Governance facts per product — training commitments, certifications, data residency, accuracy evidence, incidents on the record — each linked to the vendor page that states it. Never a ranking.
Which US judges and courts have rules on AI in filings, and what do they require?
US court AI orders
222court instruments · 42 states · 131 courts · 155 judge-specific
Every standing order, judge-requirements page, local rule, administrative order and judiciary policy on AI in proceedings before a US court — 42 states and territories, 131 courts, 155 judge-specific — coded from the primary text for disclosure, certification, verification and prohibition. New records enter the regulation tracker as unverified until the editor re-opens each court page.
What does each vendor's own website actually state, and where?
Vendor documentation audit
3480documented facts · 130 tools · audited 2026-09-06
One row per fact a legal-AI vendor publishes about pricing, contract terms, security certifications, training and retention, residency and accuracy — with the quote, the page, an archived copy and the date read. 'Not found' is recorded as a fact. Feeds the tool records and the side-by-side comparisons.
What have US courts actually said about generative AI, beyond fabricated citations?
US court decisions on generative AI
566decisions · 59 beyond fabricated citations · 9 dockets
Every court-authored US document since November 2022 that substantively discusses generative AI — opinions, orders, concurrences, dissents — read from the court's own copy, coded into eleven topics (fabricated citations, the court's own AI use, AI-generated evidence, privilege, unauthorized practice, rules adopted by opinion…) and quoted: US court text is public domain. Plus the legal-AI litigation dockets with dated milestones.
When AI-made or AI-altered evidence reaches a courtroom, what do courts do — and what do the rules say?
AI-generated evidence and deepfakes in court
60rulings · 13 countries · 10 rules and proposals
Every ruling we can find, in any jurisdiction, on AI-generated, AI-enhanced or allegedly deepfaked evidence and on deepfake defences — evidence type, who offered it, the challenge and the ruling in the court's own words — beside the rules and proposals with dated status, including proposed Federal Rule of Evidence 707. No media is hosted or linked.
What did each legal-AI vendor announce, and when?
Legal-AI release and disclosure timeline
341dated events · 75 vendors · 44 firm deployments named
Dated first-party statements from vendors' own newsrooms, blogs, status pages, filings and terms: launches, features, the underlying model when disclosed, pricing, named law-firm deployments, partnerships, acquisitions, funding as announced, certifications, incidents and terms changes. One page per month, per vendor and per event; a firm × tool table of announced deployments.
Where do courts, prosecutors, defenders and justice ministries use AI, on the record?
Justice-sector AI deployments and procurement
595official records · 17 countries · 153 bodies
Official records of AI used, piloted or procured by justice bodies worldwide: government AI-use inventories, algorithmic-transparency records, procurement notices and awards, court announcements, judgments in which a court discloses its own AI use, and judicial-conduct opinions on judges' AI use. Status words are the source's own.
Which bills and laws reach the courtroom or legal practice?
US legislation on AI in courts, evidence and legal services
41bills and laws · 10 enacted · 13 jurisdictions
A lane-limited register: state and federal bills since 2023 whose operative sections address AI-generated evidence, AI in courts, AI legal services and unauthorized practice, attorneys' AI duties, AI police reports or legal aid — with status, last action, operative sections and the public-domain text. General AI-regulation bills are excluded by rule.
What has each Member State done: law, authorities, sandbox, penalties?
EU AI Act — national implementation
147national instruments · 32 states
One page per Member State (plus EEA, Switzerland and the UK): the national implementing act, designated market-surveillance, single-point-of-contact and notifying authorities, the Article 99 penalty regime, Article 57 sandbox status (deadline 2 August 2027 after the Digital Omnibus), guidance that reaches legal practice, and any stated position on the Omnibus — dated, sourced to the gazette or authority, titles quoted in the original.
What have regulators and courts done about AI legal services offered to the public?
Consumer legal-AI enforcement and authorisation register
43official actions · 36 entities · 4 countries
Every official action on AI-enabled legal services for the public: regulator enforcement, unauthorized-practice litigation and opinions, sandbox and licence authorisations the regulator itself describes as AI-based, and courts' own AI tools for self-represented people — dated and sourced to the regulator or court. An entity appears only through an official action; nothing here reviews or rates a service.
When did we last check, and what changed?
Content log
57log entries · 840 records re-opened on record
Every content run as data: what was re-opened against its sources, what was added, updated, corrected or held, and the decision behind it. The audit trail for every last-checked date on the site.
Coming next
Datasets in preparation
Each becomes a page and a JSON file once the SafeLegalAI editorial team has verified the first release. Tell us what you need at hello@safelegalai.com.
Regulator enforcement on lawyers' AI misuse
Disciplinary and regulatory dispositions (SRA, BSB, state bars, law societies) as their own table, linked back to incidents.
planned
Guidance version diffs
Machine-readable diffs between versions of the same guidance document.
planned
load_dataset. Each row links back to its record here; the cards carry the same licence and disclaimer as this site.Embed a widget →
The latest incidents or rules — worldwide or one jurisdiction — on your blog, library guide or intranet. Refreshes when the record changes.
Syndicate the weekly block →
A ready-to-paste "This week in legal AI" section for other newsletters, global or per jurisdiction, in HTML, Markdown or JSON.
Watch a jurisdiction or tool →
Scoped RSS feeds and email digests that send only when something on the record changes for that scope.
Terms that travel with the data
Provided as is, without warranty of any kind. Not legal advice; no lawyer–client relationship. SafeLegalAI records findings made by courts, regulators and vendors' own published pages; the linked official documents are the record. Third-party names and marks identify their owners' products or bodies and imply no affiliation or endorsement. Every JSON export carries this notice in a notice field; every dataset card, feed, widget and chart repeats it. Full terms, including the right of reply for anyone named in a record, notice-and-takedown, and the CC BY 4.0 warranty exclusion: /disclaimer.
Load the data
Every table is one load_dataset call on Hugging Face, one file on GitHub, or the JSON this site serves. The url field on each row is its record page here; the notice field carries the terms.
from datasets import load_dataset # any table: "safelegalaidata/<repo>"incidents = load_dataset( "safelegalaidata/legal-ai-incidents", split="train")courts = load_dataset( "safelegalaidata/us-genai-court-opinions", split="train") print(incidents[0]["url"]) # its record page here# canonical exports, CC BY 4.0, each with a notice fieldcurl -s https://safelegalai.com/tracker/incidents.json \ | jq '.incidents | length'curl -s https://safelegalai.com/courts/opinions.json \ | jq '.decisions[0] | {case_name, court, url}' # the same tables as CSV/JSONL on GitHubgit clone https://github.com/SafeLegalAI/legal-ai-incidentsFiles by group
- Legal AI Incident Tracker JSON
- US court decisions on generative AI JSON
- AI-generated evidence and deepfakes in court JSON
- Regulatory map JSON
- Official documents, versioned JSON
- EU AI Act, structured JSON
- EU AI Act — national implementation CSV
- US court AI orders CSV
- US legislation on AI in courts, evidence and legal services JSON
- Consumer legal-AI enforcement and authorisation register JSON
- Justice-sector AI deployments and procurement JSON
How to cite
SafeLegalAI (published by SafeLegalAI), “[dataset name]”, safelegalai.com, accessed 2026-10-07. Data: CC BY 4.0. Every record page carries its own citation block with the record's last-checked date.