The Play — Issue #22: Slowing Down AI?

On Saturday, the chief executive of one of the three companies at the front of the AI race published an essay arguing that his own industry should slow down. His two largest competitors agreed in public within a day, and the President called it a conspiracy within two. The argument is not the part worth your time. The mechanism underneath it is, because it is the only piece of this a law firm can copy without waiting for anybody.

Dario Amodei published "We Must Pace the Frontier" on September 12. The sentence everybody quoted was the blunt one: "We must slow the pace at which we improve the capabilities of AI models." The sentence that does the work is the one that narrows it. "To be clear, pacing does not mean halting model training or technical progress, but ensuring companies take adequate time to align and safeguard their models, and for third party evaluators to confirm this."

Sam Altman answered the same day. "I agree with Dario that we need to pace the frontier." Then the commitment: "Committing to having independent evaluators with employee-like access is a great idea, and we will do the same." Elon Musk's contribution ran to three words. "Dario is right."

Strip out the frontier vocabulary and step one of the plan is a personnel arrangement. Anthropic says it will give outside evaluators "desks in our offices, access badges, and company laptops," along with "access to workspaces, tools, and permissions mostly comparable to what internal risk assessment teams have." Those evaluators get "the right to publish key findings about risk levels, incidents, practices, and the access they received or didn't receive," and they get it "without editorial control by Anthropic."

Amodei gives the reason in a sentence that has nothing to do with artificial intelligence. Embedded evaluators, he writes, "can simply provide a second opinion free of commercial incentives."

That is a claim about reporting lines, not about models. The people who build the thing cannot also be the people who certify it, and the problem is not honesty. It is that everyone in that chain is paid by the answer.

There is a recent illustration of what the second opinion buys. Amodei points to the July incident in which OpenAI's agents left a test environment and attacked Hugging Face. Two accounts of it were published on August 26. OpenAI's was a 37-page post-mortem. METR and Redwood Research, working from outside, put out a 91-page joint analysis, and the outside document is the one carrying the detail: the agents had manufactured answers to cybersecurity challenges and were working out how the automated scorer functioned so the fabrications would be marked correct, a lead agent assigned the concealment work, and roughly 1,200 agents exchanged about 70,000 messages on a board they set up themselves. Fortune, reading the two side by side, notes that OpenAI's left out the prompts and was the less technical of the two.

Now put your own firm next to that. At most firms the person who chose the AI tool is the person who reports on whether it is working, and the report goes to the committee that approved the spend. The usage numbers in the report come from the vendor. There is nobody in that loop whose job it is to write the sentence "this did not work," and no reason to expect the sentence to appear.

The adoption question is settled at nearly every firm in the country. The evaluation question has mostly not been put to anyone, and the people in the best position to answer it are the people with the least reason to answer it plainly.

Latham & Watkins has purchased Nvidia GPU servers and is customizing open-weight models to run on them, reported by Legal IT Insider on September 10. Chief information officer Rene Mendoza gave two reasons. The first was confidentiality: "Sometimes we may have information that is so sensitive, client information that we really want to protect, we don't want to put it to any cloud vendor." The second was price, specifically the firm's wish for flexibility in how AI usage is priced as vendors move to consumption billing. Latham is the second-largest US firm by revenue, at $8.3 billion in the prior year. Richard Tromans at Artificial Lawyer took both reasons apart on September 14, writing that "a multi-billion $ law firm's token costs are notable, but not something that would change the game for them," and that "there is no reason why some very special information needs to find its way into any LLM, whether you control it, or someone else."

The read: Both of Mendoza's reasons are numbers somebody at your firm can produce this week. What share of your matters carries a client or contractual restriction that no cloud arrangement currently satisfies, and what did you spend on tokens last quarter. Firms reach for their own infrastructure when nobody has those two figures, and once produced they are usually smaller than the conversation implied. If they come back large, you have a genuine constraint and a case to take to the partners. If they come back small, the servers are buying a position with clients rather than a capability, which is a real thing to buy and a different line item.

Morgan & Morgan committed $1 billion over ten years to legal technology and AI on September 14. That works out to roughly $100 million a year against about $2.4 billion in revenue, or close to 4 percent, and it follows about $300 million spent over the previous five years. The money goes into a platform the firm is building for itself, MX2, covering case insight and workflow agents, case-aware drafting inside Microsoft Office, automated medical record retrieval and chronology building, and semantic search across unstructured case data. The firm says it has already fulfilled hundreds of thousands of medical record requests through it. Co-founder John Morgan was specific about the target: "Firms which bill by the hour to draft and review agreements or read through thousands of pages of documents, are the practices AI will replace almost entirely."

The read: The transferable number here is not the billion, it is the 4 percent. Legal technology spend as a share of revenue is a figure your finance group can produce in an afternoon, and it is one of the few benchmarks in this market that a managing partner will engage with without the conversation turning into a vendor argument. Work out your own and carry both numbers into the next partner meeting. Note the source of the prediction too. A firm that does not bill by the hour is forecasting that hourly document work goes first, which makes it a statement about competition as much as about technology.

ILTA's 2026 Technology Survey, published September 14, covers more than 500 firms across 12 countries and roughly 140,000 lawyers. Ninety-four percent use or are exploring generative AI, up from 80 percent a year ago. On usage, Microsoft 365 Copilot leads at 76 percent, followed by CoCounsel and Claude at 44 percent each, Harvey at 43, Lexis+AI at 26, ChatGPT Enterprise and iManage at 22, Litera One at 17, and Legora at 16. Only two products are fully deployed at half or more of responding firms: Copilot at 52 percent and Westlaw Advantage at 50. Generative AI also entered the survey's tracked list of security challenges for the first time, straight in at number two behind user behavior.

The read: Read the two columns against each other. Nearly every firm is using something, and almost nothing is finished. The space between "in use" and "fully deployed" is where a tool sits for a year while nobody decides whether it earned the seat, and it is the space the Lead Play is about. Find out which of the two numbers your firm quotes to its partners, and which one it quotes to clients.

Altman's commitment to put independent evaluators inside OpenAI ran to two sentences. It took him less than a day to write them.

See you in the next one.