Artificial intelligence is already part of the day-to-day operations of almost every law firm we work with. They use it to research, draft, and review their deliverables. According to a study published by LexisNexis a few days ago, 94% of lawyers in the United Kingdom now use artificial intelligence in their work, compared with 41% two years ago, and the figures in Latin America do not appear to be substantially different. Although these tools have become indispensable to the efficiency and quality of service required, the documents delivered to clients seem to have begun to carry, without anyone noticing yet, a mark revealing that they have passed through them. Last week, a first public way of detecting that mark emerged, although for now only in files generated by the tool, not in text that is copied and pasted.

For a firm whose business depends on its reputation, this creates, in my view, a problem on two fronts: with the client, defending the authorship and value of work that the document no longer clearly reveals; and with the provider, retaining control over the knowledge with which it delivers its services, a discussion that the recent launch of Gemini Enterprise for Legal has made urgent. Let us examine this knowing that the discussion is only just beginning.

The news is something we have already been hearing: Claude’s texts carry an invisible mark, woven into the way the model chooses its words, that travels with the text when it is copied and pasted. Less attention has been paid to three details that matter more to law firms: (i) the mark would also extend to products built on these models, such as Harvey, Legora, or Robin AI, which do not train their own underlying model but operate on models from providers such as Anthropic, OpenAI, or Google and would therefore inherit the mark those models leave in the text; (ii) there will be an official detector, currently in private preview, and tools already claim to remove the mark, meaning that no one yet knows with certainty how third parties will be able to detect it; and (iii) because the other developers signed the same code of good practices, which applies to all providers operating in the European Union, even though Claude has decided to apply it worldwide, it is only a matter of time before each model leaves its own mark.

Let us begin with the client side, using a common example. A legal opinion may have been prepared by a lawyer and passed through one of these models solely to correct or translate it. If the mark works as announced, the document would be marked without allowing anyone to distinguish whether the tool produced the analysis or corrected a sentence. The client would know that artificial intelligence was used, but not what the tool did, which part of the work actually belongs to the lawyer, or the extent to which the lawyer relied on it. If that is the case, a cloud of doubt would fall equally over the entire document, regardless of the depth of the adjustments.

There are also the technical limitations of the mark itself. Extensive editing, paraphrasing, or translation may weaken it. Short texts barely give the mark room to form, and there will be cases in which the detector flags texts that never passed through a tool. The absence of the mark will not prove that artificial intelligence was not used. In the end, with the information we have so far, the detector may be enough to raise suspicion, but not to resolve it, not even in favor of the firm.

As I tried to understand how far these new rules of the game would reach, Google launched Gemini Enterprise for Legal and joined developers that had already launched their legal divisions, such as Anthropic with Claude for Legal. That was when I began to worry about something that goes beyond everything discussed above: the provider side. If the mark leaves a signal on the finished document, this platform enters directly into the way it is produced. It connects to the firm’s systems and, according to Google, turns its precedents, styles, templates, and negotiation positions into instructions that the platform executes.

Although the company assures us that the information will remain within the private environment and will not be used to train its models, that promise only covers confidentiality. The firm’s knowledge will nevertheless begin to be executed on infrastructure that does not belong to it. More than in the contract that is signed, the firm’s dependence on the provider will lie in the accumulated knowledge within the platform, in the cost of extracting it, and in the tool’s necessity for day-to-day operations.

Google has not announced a mark for Gemini like the one described by Anthropic. The two moves show where the market is heading: platforms are increasingly participating in the way firms organize and execute their knowledge, and they are beginning to leave signals on the documents in which they have intervened. For firms positioned around their experience, this is, at the very least, dangerous.

Firms have always delegated work among different departments, associates, support staff, and external providers. That pyramid structure is, in fact, what makes the business viable and attractive. Artificial intelligence is now being incorporated into that chain to research, compare documents, and prepare drafts. The client never knew which part of an opinion had been written by a second-year associate or to what extent each person had intervened, because what matters is the final result; the mark, by contrast, will reveal when the tool participated. The principle remains the same: we can delegate the work, but we retain responsibility. If the tool introduces an error, the firm will continue to be responsible for the document delivered.

Let us now consider the financial component. For the firm, the tool is a good business proposition. A fixed-fee legal opinion that previously took twenty hours can be completed in eight with the platform, and even after paying for licenses and reviews, the margin improves. But the client who sees the mark may do a different calculation: if the tool did the work, why do the fees not fall in proportion to the hours? Because the actual savings are smaller than they appear. It remains a good business, but not as good as we thought. Licenses for these platforms continue to become more expensive, and, above all, the firm must continue training associates so that the career path and the pyramid structure that supports the business remain intact.

Against this backdrop, I believe each firm’s ability to defend its pricing will depend on its positioning. In commodity and process-driven services, where the work is comparable and several firms have access to similar technologies, part of the efficiency gains will ultimately be passed on to clients through lower prices or faster delivery. In experience-based services, and especially in expertise—the so-called rocket science—the client pays for the judgment required to resolve an unusual situation and for the risk assumed by the person making the recommendation.

The tool can reduce preparation time without replacing that judgment, but the mark threatens to blur that distinction in the client’s eyes. A firm that charges for expertise will have to explain what the platform produced, what its lawyers reviewed, and where the professional judgment underpinning the fees came into play. Otherwise, knowledge that is difficult to replicate may end up looking like standardized work. This will be the challenge we have to face in the coming months.

Finally, I believe another market that will be affected is the partnership. The partnership captures the benefit of efficiency and assumes decisions regarding quality, risk, and knowledge protection. Partners will have to understand who owns what the firm builds within the platform, how templates and playbooks are transferred, and what happens to them when the relationship ends.

Firms have spent the past year deciding which tools to incorporate and how to use them. The watermark changes the underlying question and adds new nuances, because the issue is now about preserving, in front of the client, the authorship and value of professional judgment and, in front of the provider, control over the knowledge with which the firm delivers its services. This is not a discussion that the technology department can resolve; it is a decision for the partnership. Since almost all of this was announced only a few weeks ago, we still have to see how it plays out. It is worth following closely a change that is only beginning to alter the way legal services are delivered.

Antonio Gómez Montoya
Partner at Black Swan