The Signal

Sourced evidence on legal AI

Bring stronger evidence to your next legal AI conversation.

Search 3,240 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.

3,240Statistics
870Reports
20202026Covering
18 Aug 2026Last updated

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.

Recently added

Literal text match across stats, sources and context. Use a word or phrase, not a question.

Try
Scope
Theme
Source
Year

50 in the newest 50, loading the other 3,190

TechnicalVendor
17 Aug 2026

GPT-5.6 Sol, asked to assess a simple static personal website, found 13 security issues in about 15 minutes and remediated them over roughly an hour

Greg Brockman (OpenAI), "The Defender's Window"

AdoptionVendor
17 Aug 2026

At OpenAI, "almost all of our initial security alerts are triaged by intelligence before humans are looped in"

Greg Brockman (OpenAI), "The Defender's Window"

damiencharlotin.com
TechnicalVendor
7 Aug 2026

OpenAI states it "cannot rule out critical cyber capabilities" for its upcoming model Astra under its Preparedness Framework; previous models including GPT-5.6-Sol were assessed at High rather than Critical

OpenAI, "Responding to the next frontier of critical cyber capabilities"

RegulatoryGovernment
4 Aug 2026

$3,500 Rule 11 sanction plus referral to the Grievance Committee against defence counsel for AI-assisted filings in which no case was fabricated

*Barteca Holdings LLC v. TacoBarn Newtown LLC*, No. 3:26-cv-00250-VDO (D. Conn.), Order Imposing Sanctions ECF No. 43

RegulatoryGovernment
4 Aug 2026

US monetary sanctions for AI hallucinations "in most cases" range from $1,500 to $15,000

Judge Vernon D. Oliver's own survey in *Barteca v. TacoBarn* (D. Conn.)

RegulatoryGovernment
4 Aug 2026

Sanctioned counsel used three AI tools on the offending brief: Open Law (paid subscription), Claude and ChatGPT, on his first ever AI-assisted matter

*Barteca v. TacoBarn* (D. Conn.), show-cause hearing record

redis.io
CriticalPractitioner
Aug 2026

Monthly questions recorded on Stack Overflow fell 99.3%, from a peak of 207,204 in March 2014 to 1,442 in July 2026

Stack Exchange Data Explorer via Daniel Lockyerx.com

MarketNews and media
Aug 2026

Norm Law, the AI-native law firm affiliated with Norm Ai, was operating with 25 lawyers across 8 practice areas seven months after launch

Legal Technology (Reena SenGupta)legaltechnology.com

MarketNews and media
Aug 2026

Norm Law named Blackstone, Goldman Sachs and New York Life as clients; Blackstone, Vanguard and New York Life are also Norm Ai investors

Legal Technology (Reena SenGupta)legaltechnology.com

AdoptionAcademic
Aug 2026

ChatGPT Enterprise aggregate output tokens grew roughly sevenfold between June 2025 and March 2026

OpenAI / Chatterji et al, How Organizations Use AIcdn.openai.com

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.

Explore TITANS Intelligence

Provenance

Don’t trust. Verify.

Every stat here carries a source and a date, but we have not independently re-checked them all against the original documents. The honest position is that this is a starting point extracted from reports, not a final answer.

Before a number goes into a client deck or a newsletter, we run the verify skill below against the specific stats that matter. It finds the original source, cross-references it against independent reporting, and flags anything that does not match. It uses parallel research where the tool supports it.

Claude Code: copy the skill, drop it into ~/.claude/skills/verify/SKILL.md and run /verify on whatever you are about to cite.

ChatGPT: open Plugins > Skills > Create > Create with editor, paste the skill and install it. In a chat, mention @verify followed by the claim or stats you want checked.

The vault file paths and changelog step in the skill only apply when the tool can access that vault. Pasted claims work in either tool.