The Rise of Shadow AI

Hey happy Sunday.

A lot to unpack in this week's edition.

Let's dig in.

New research from Access Legal, surveying 100 fee earners and 100 firm leaders across the UK—reveals a massive disconnect between leadership and the frontline when it comes to AI adoption.

The data highlights a growing trend of what is being called Shadow AI employees using unapproved, personal AI accounts to run work-related queries.

The facts from the report:

• 71% of paralegals and 57% of solicitors admit to using unapproved AI tools just to keep pace with their workload.

• Meanwhile, 68% of firm leaders believe they have full visibility and zero risk of unapproved AI touching client work.

Part of this shadow usage is driven directly by client pressure. The study found that 50% of clients now expect AI to be involved in their legal work. When the market demands AI-assisted efficiency but the firm hasn't provided an approved tool that actually does the job, fee earners are reaching for the closest available solution.

I think this raises two major questions:

1. Are firms actively equipping their teams with secure, enterprise-grade AI tools that are genuinely suitable for legal work?

2. Should firms outright block access to general consumer models like ChatGPT and Claude? Or is the better path to leave them open, trusting and training employees to experiment safely without inputting confidential data?

Berkeley Law goes the other way. Berkeley has adopted one of the most restrictive AI policies for law students. From this summer, students will be barred from using AI tools not only to draft papers, but also to brainstorm ideas, organize arguments, revise prose, correct grammar and translate written work into English. Uploading class materials into generative AI systems is also prohibited. Narrow exceptions remain for research and for courses explicitly designed to teach AI fluency. The bet is that thinking, not tooling, is what makes a lawyer.

Anthropic wants its accountants on the bleeding edge. Anthropic has posted a role for a “Head of Accounting AI & Innovation” in San Francisco. The successful candidate “will own the accounting AI program end-to-end: from roadmap to production deployment, from governance frameworks to community building, and from employee enablement to enterprise scale.” This is a sign they are heading vertical towards finance specific products. Could they be hiring for a legal product focused role next?

Harvey and DeepJudge, the Swiss-founded institutional intelligence platform, announced a partnership to bring an organisation’s past work, decisions, and expertise into Harvey’s workflows. The integration is framed as solving the “context tax” — the gap between what a firm collectively knows and what its AI tools can actually access in real time.

LawX raises €7.5m. Berlin-based LawX has secured €7.5 million in seed funding led by Motive Partners, with WENVEST Capital, xdeck and SIVentures participating, alongside angel backing from Flink co-founder Christoph Cordes and ex-Deutsche Bank board member Ralph Müller.

Lawve AI joins the Claude for Legal ecosystem. Lawve AI has collaborated with Anthropic to make community-built legal AI skills discoverable inside Claude, via a new connector that lets users search a curated public catalogue of legal Agent Skills created by practicing lawyers, in-house counsel and academics. Skills are searchable by jurisdiction, topic, language or workflow.

Recently Google users searching the word “disregard” on Google began receiving, instead of a dictionary definition, the reply: “Understood. Let me know whenever you have a new prompt or question.” The same happened with “ignore,” “dismiss” and “stop.” The AI Overview, it transpired, was treating the search query as a prompt instruction rather than a search term.

Google had actually rolled out a completely new Search experience earlier this week, foregrounding AI summaries and pushing the traditional ten blue links well down the page. A search engine that occasionally forgets it is a search engine is a useful reminder that putting an LLM at the front of any workflow is a trade-off, not a free upgrade.

That's everything for this week.

See you next Sunday,

George