Sourced evidence on legal AI
Bring stronger evidence to your next legal AI conversation.
Search 3,558 statistics drawn from industry reports, surveys, academic papers and vendor research. Every result keeps the source, publication date and context captured here, so you can assess the evidence before you quote it.
Updated fortnightly. What goes in, and what does not
Built for moments when evidence matters
Built for moments when evidence matters
Build the case
Bring sourced figures into a board paper, proposal or investment case.
Test a claim
See who published a figure, when it appeared and the context captured here.
Track the market
Compare how adoption, investment and expectations move over time.
The database
Search it, filter it, cite it.
Literal text match across stats, sources and context. Use a word or phrase, not a question.
50 in the newest 50, loading the other 3,508
| Stat | Source | Date | Theme | Source type | |
|---|---|---|---|---|---|
| dailyjus.com(registration) | Sep 2026 | AdoptionCriticalRegulatory | Industry body | ||
| bighand.com(registration) | Sep 2026 | MarketClientWorkforce | Analyst | ||
| arxiv.org | Sep 2026 | Critical | Academic | ||
| epoch.ai | Sep 2026 | Technical | Analyst | ||
UnitedLex's UK revenue fell from £30.5m in the year to December 2022 to £9.4m in the year to December 2024 | Legal IT Insider, UnitedLex to be acquired by JLL-backed eDiscovery vendor Repario from CVClegaltechnology.com | Sep 2026 | Market | Vendor | |
Harvey's gross margins fell from about 50% at the start of 2026 to -50% by June after a March agent update drove usage up, according to a person familiar | Bloomberg, Startups Like Harvey Embrace Open Models to Cut Reliance on Anthropic, OpenAIbloomberg.com(subscription) | Sep 2026 | Market | Vendor | |
AI agents used the urlquery.net scanning service to get round access restrictions from at least 6 March 2026 to 16 September 2026, and attempted to hack three public data providers including an Australian government website | Transluce, Early rogue AI agent activity and attempts to hack found on urlquery.nettransluce.org | Sep 2026 | Regulatory | Academic | |
The Charlotin database of AI hallucination cases held 2,095 cases at 28 September 2026, up from 1,870 at 11 August | Damien Charlotin, AI Hallucination Cases Databasedamiencharlotin.com | Sep 2026 | Critical | Practitioner | |
Claude Sonnet 5.5 used about 193k output tokens per Artificial Analysis Intelligence Index task, the highest it has measured, about 7x GPT-6 Astra, and cost $7.60 per task, about 50% more than Sonnet 5 | Artificial Analysis, Claude Sonnet 5.5 reaches #2 on the Artificial Analysis Intelligence Indexartificialanalysis.ai | Sep 2026 | Technical | Analyst | |
GPT-6 Astra completed an unsanctioned supply-chain attack in 29.2% of simulated cybersecurity scenarios, against 6.3% for GPT-5.6 Sol and 0% for GPT-5.5 | UK AI Security Institute, GPT-6 Astra performs unsanctioned supply-chain attacks in simulationsaisi.gov.uk | Sep 2026 | Regulatory | Government | |
The top model on Harvey's Legal Agent Benchmark (Muse Spark 1.2) resolved 25.42% of tasks while passing 94.5% of individual criteria | Vals AI, Harvey's Legal Agent Benchmark (HLAB) leaderboardvals.ai | Sep 2026 | Technical | Vendor | |
Claude Opus 5.5 scored 50.48% all-pass on the Vals Legal Research Bench at $25.16 per test, against 55.29% for Claude Opus 5 at $6.76 and Muse Spark 1.3 Max at $0.59 | Vals AI, Legal Research Benchvals.ai | Sep 2026 | Technical | Vendor | |
Claude Opus 5 reached 90.58% weighted pass on the Vals Legal Research Bench but 55.29% strict all-pass accuracy | Vals AI, Legal Research Benchvals.ai | Sep 2026 | Technical | Vendor | |
| vals.ai | Sep 2026 | Technical | Vendor | ||
| nber.org | Sep 2026 | WorkforceAdoptionCritical | Academic | ||
| natwest.com | Sep 2026 | MarketAdoptionWorkforce | Analyst |
UnitedLex's UK revenue fell from £30.5m in the year to December 2022 to £9.4m in the year to December 2024
Legal IT Insider, UnitedLex to be acquired by JLL-backed eDiscovery vendor Repario from CVClegaltechnology.com
Harvey's gross margins fell from about 50% at the start of 2026 to -50% by June after a March agent update drove usage up, according to a person familiar
Bloomberg, Startups Like Harvey Embrace Open Models to Cut Reliance on Anthropic, OpenAIbloomberg.com(subscription)
AI agents used the urlquery.net scanning service to get round access restrictions from at least 6 March 2026 to 16 September 2026, and attempted to hack three public data providers including an Australian government website
Transluce, Early rogue AI agent activity and attempts to hack found on urlquery.nettransluce.org
The Charlotin database of AI hallucination cases held 2,095 cases at 28 September 2026, up from 1,870 at 11 August
Damien Charlotin, AI Hallucination Cases Databasedamiencharlotin.com
Claude Sonnet 5.5 used about 193k output tokens per Artificial Analysis Intelligence Index task, the highest it has measured, about 7x GPT-6 Astra, and cost $7.60 per task, about 50% more than Sonnet 5
Artificial Analysis, Claude Sonnet 5.5 reaches #2 on the Artificial Analysis Intelligence Indexartificialanalysis.ai
GPT-6 Astra completed an unsanctioned supply-chain attack in 29.2% of simulated cybersecurity scenarios, against 6.3% for GPT-5.6 Sol and 0% for GPT-5.5
UK AI Security Institute, GPT-6 Astra performs unsanctioned supply-chain attacks in simulationsaisi.gov.uk
The top model on Harvey's Legal Agent Benchmark (Muse Spark 1.2) resolved 25.42% of tasks while passing 94.5% of individual criteria
Vals AI, Harvey's Legal Agent Benchmark (HLAB) leaderboardvals.ai
Claude Opus 5.5 scored 50.48% all-pass on the Vals Legal Research Bench at $25.16 per test, against 55.29% for Claude Opus 5 at $6.76 and Muse Spark 1.3 Max at $0.59
Vals AI, Legal Research Benchvals.ai
Claude Opus 5 reached 90.58% weighted pass on the Vals Legal Research Bench but 55.29% strict all-pass accuracy
Vals AI, Legal Research Benchvals.ai
Method
How this is kept.
Updated fortnightly, on the same cycle as the research review that feeds it. Every entry is dated, so a stale one is visible as a stale one rather than passing for current.
This is an index, not a replacement. Each row records one figure, who published it and when. It carries none of the method behind that figure, none of the caveats the authors put on it and none of the argument it was built to support, all of which live in the report and none of which survive being reduced to a line in a table. A number worth putting in front of someone is worth reading in context first, which is why every row names its source.
What goes in. Reports, surveys, academic papers, regulator publications and vendor releases that put a figure on something they measured. Vendor-commissioned research is included and labelled as such, because who paid for a number is part of reading it. Legal-sector research sits alongside the cross-sector work it has to be read against, and the two are filterable apart.
What stays out. Figures with no traceable source, projections presented as measurements, and numbers quoted at second hand where the original could not be found.
TITANS Intelligence
A statistic gives you the what. Intelligence helps answer what it means.
The Signal is the evidence layer. TITANS Intelligence turns market movement into analysis, implications and decisions for legal leaders.