Article

2026-05-31 Cyber

[Reuters] At many law firms, calling a colleague at 2 a.m. is a last resort. At Ohio-based law firm Vorys, Sater, Seymour and Pease, the colleague is always awake — because it's AI.

The 375-attorney firm is rolling out one of the more eye-catching uses of legal AI that I’ve heard about recently.

Working with an AI-focused research lab at Stanford Law School, Vorys has developed “AI personas” of 19 of its partners, which can be embedded within generative AI tools to offer responses to questions and edit documents in the style of individual partners.

Seemingly every major law firm is experimenting with ‌AI right now, even as firm leaders wrestle with the unsettling question of how it could upend their entire business model. In conversations, firm leaders can’t yet answer how AI will reshape attorney training or the partnership model as the technology matures, but many acknowledge the bigger risk is falling behind.

Many firms are working with off-the-shelf AI products, though some are trying to get an edge by developing their own tools. Kirkland & Ellis, for instance, said this week it will invest $500 million to develop its own AI technology that will pull from Kirkland's internal intelligence.

BUILDING AI BRAINS
At Vorys, trying to stay ahead has meant partnering with Stanford along with separately developing other AI projects, like ones to help with labor-and-employment research for HR professionals and patent drafting.

When I first heard Megan Ma, the executive director of the Stanford Legal Innovation through Frontier Technology Lab, also known as liftlab, talk about what she dubbed “AI twins,” I pictured little online avatars crafting legal documents. ⁠The reality is slightly more nuanced but still intriguing.

The personas essentially work as an AI chatbot that spits out replies and edits documents in the mindset of the partner. The digital personalities are developed from hours-long interviews about the attorney’s approach to practicing law, values and experiences. They aren’t trained on work product from the partners, but instead seek to embody their individual thinking styles.

“It's almost like having a low resolution map of the person's brain,” said Nate Jedinak, Vorys’ senior director of software, data and innovation. He said the personas have been used by both associates looking to improve their work and partners using them to “be creative and be their best selves.”

Wayne Stacy, the executive director of the Berkeley Center for Law & Technology, said that while he sees AI personas as being novel, he doesn’t think they’ll fundamentally change the practice of law. He said a more impactful approach is building AI workflows around the specific skills needed to complete a task successfully, rather than mimicking one attorney.

“The problem with personas, at too granular of a level, it glorifies the individual personality of a lawyer” and how they think something should be done, Stacy said.

Scott Powell, a federal tax, aviation and transactional partner at Vorys, said the firm sees value in creating individual personas rather than an aggregate “best M&A lawyer” or “top litigator.” He said such aggregate personas work well with basic questions but that the nuance and skill of specific attorneys give better results. The ability to chat with personas based on actual partners also better reflects how people act ‌in real life, ⁠he said, when lawyers seek feedback from a trusted colleague down the hall.

Powell said he was a bit skeptical about how the answers he gave Ma to create his persona would lead to meaningful advice. Once he started using it, he realized, “It was presenting results to queries that were shockingly sounding a lot like me.”

The firm is rolling out the AI personas firmwide after testing them with smaller groups. Leadership has made clear that they are not a replacement for an attorney. Like all AI, the tools that use the personas run the risk of hallucinations, but Powell said they emphasize that human review to check for accuracy is a necessary step.

During the test phase, one Vorys litigator gathered viewpoints from three AI personas to strengthen the legal theories in a U.S. Supreme Court amicus brief. An intellectual property lawyer worked with personas to do ⁠a deep dive on a troublesome portion of a patent. With the editing function, associates can redline a document in the style of a partner to get feedback before turning in a draft.

Vorys M&A partner Bruce Paige said he's used litigator personas "to essentially pressure test" deal documents to look out for potential litigation risks, a step that used to be too time and cost-prohibitive on smaller deals.
# id) ?>

Posted by Skidmark 2026-05-31 00:00|| E-Mail|| Front Page|| [1867 views ] Top
File under: Lawfare

1 What does the AI lawyer charge per hour?

Posted by Seeking Cure For Ignorance 2026-05-31 02:40|| 2026-05-31 02:40|| Front Page Top

2 AI hallucinates and cites non-existent cases in its briefs. Already been caught several times by judges.

Posted by Procopius2k 2026-05-31 06:47|| 2026-05-31 06:47|| Front Page Top

3 These tools are only as good as the underlying programming. How many times have human lawyers been caught in court using false AI generated briefs etc. as P2K points out. How long until we get AI judges? Chat/GPT is already left leaning. The more you let computers do, the easier it is to game the system. My 42 years in the field tells me no good can come of this.

Posted by Warthog 2026-05-31 08:04|| 2026-05-31 08:04|| Front Page Top

4 "I Am The Law!"

Posted by swksvolFF 2026-05-31 10:54|| 2026-05-31 10:54|| Front Page Top

5 AI hallucinates and cites non-existent cases in its briefs.

That seems rather judgemental and robo-phobic. And who hasn't done questionable things in their underwear?

From the AI's point of view, hallucinations and "facts" are all the same thing - just a stream of tokens generated by the probability that this follows that. Human judgement is what determines if the AI is hallucinating. It simply cannot tell.

These tools are only as good as the underlying programming.

Once upon a time, that was strictly true. But the neural networks that form the basis for modern AI are really a new New Thing. Programming them is mostly about configuring the layers and interconnections of the network. The tricky part is training the network.

Garbage-in/Garbage out still applies, but now we are talking about the training data. A neural network actually learns, but what it learns is dependent on what you feed it. A popular way to get training data is to hoover the Internet scraping up every bit of text, image, and audio you can find.
tl;dr: If you train your neural network on stupid things, your network will be stupid too.

There is an interesting feedback loop in play here. Given that a lot of training data comes from the internet, as more AI-generated content appears online, that generated content will be sucked up as training data for a new round of AIs. A self-licking ice cream cone that eventually leads to the Butlerian Jihad.

Posted by SteveS 2026-05-31 21:24|| 2026-05-31 21:24|| Front Page Top

6 who hasn't done questionable things in their underwear?

I already paid the ransom for the ~~non-existent~~ videos! Quit spamming me!

Posted by Frank G 2026-05-31 21:47|| 2026-05-31 21:47|| Front Page Top

15:51 DarthVader
15:32 NoMoreBS
14:32 badanov
14:26 Silentbrick
14:10 chris
14:06 Skidmark
13:48 Bugs the Wicked6167
13:15 badanov
13:09 DarthVader
12:57 Grom the Reflective
12:06 Grom the Reflective
11:57 Bobby
11:55 Grom the Reflective
10:47 DarthVader
10:39 Silentbrick
10:14 Bobby
10:13 badanov
10:00 Snesing Schwarzeneggar8998
09:58 Besoeker
09:45 Bobby
09:34 Warthog
09:22 Besoeker
09:17 Mercutio
09:13 Glenmore
Merry-Go-Blog