News You Can Use

In a preregistered Minnesota trial involving 91 upper-level law students, using Gemini 2.5 Pro to synthesise an unfamiliar legal doctrine increased the quality of the initial work by 59.8%. When AI was removed for the next task, the group showed no immediate comprehension loss. Its advantage on a later application memo disappeared once the quality of the earlier synthesis was controlled for: AI had improved the material they were reasoning from, rather than their underlying reasoning. When everyone later used AI to revise their own memo, weaker work generally improved while some stronger work deteriorated, although regression to the mean makes the size of that decline unsafe to use. In a separate preregistered study of 52 professional developers, AI users scored 4.15 points lower on a 27-point comprehension test and received no speed benefit. Those who generated code and asked one follow-up question scored 86%; those who stopped after generating it scored 39%. A third study found no measurable distortion when AI was used as an information tool, but wholesale delegation made writing more homogeneous and could alter its argument.

Whether AI helps or harms depends heavily on how much of the thinking is delegated. It can improve the artefact that carries work into the next stage, while also preventing the user from acquiring the knowledge they would have picked up by doing it themselves. This explains why genuine practitioner enthusiasm and disappointing organisation-wide results can coexist. Deployment should therefore be designed around behaviours, not usage totals: break work into defined components, require users to interrogate and explain outputs, and avoid mandatory AI revision at the end of work that was already good. Asking whether somebody used AI tells us very little. Asking whether they could defend the resulting work in a demanding conversation, without going back to the model, is a much better test.

Harvey and EngramLab created Calderwood & Harkness, a synthetic law firm containing 46 clients, 266 matters and 9,288 files across 15 practice areas. Harvey then tested GPT-5.6 Sol and Claude Opus 4.8 on 250 tasks against that shared institutional record. Both models generally found the core information but struggled to know when they had found everything, particularly where a task required several separate facts from across the corpus. Harvey’s published baseline said the models satisfied around half the grading criteria, although the rubric was revised four days later, so the qualitative failure is more durable than the percentage. Separately, DeepJudge published the draft Agent Handoff Protocol, with Harvey planning beta support and Thomson Reuters intending to support it in CoCounsel. AHP moves a user and an approved package of task context between complete agent products: an objective, selected messages, files and a thread identifier. System prompts, hidden reasoning, credentials and execution privileges stay behind.

Legal work does not begin and end inside one prompt or one product. A matter accumulates facts, decisions, permissions and work product over months, then crosses between research, drafting, review and specialist tools. Re-uploading documents and re-explaining the task at every boundary loses information and creates inconsistent copies. Giving every product permanent access to everything creates a different problem. AHP is an early proposal rather than an adopted standard, but it identifies the right unit of portability: selected matter context under the user’s control. Harvey’s benchmark shows why that context also needs structure. Search alone cannot tell an agent whether its answer is complete. Firms will need permissioned indexes, summaries and memory that persist across tasks, alongside auditable rules governing what moves between systems. The context layer is becoming both the source of useful performance and the next source of switching costs.

Ireland’s High Court issued Practice Direction HC 142 on the responsible use of generative AI in court documents, effective from 1 September. It applies to pleadings, submissions, affidavits, witness statements and expert reports in new and existing civil proceedings. Responsibility remains with the party, lawyer, witness or expert regardless of AI use, and witnesses must declare that AI was not used to generate substantive content in statements or affidavits. Formatting and spellchecking are permitted. Meanwhile, Anthropic has begun applying invisible text watermarking and C2PA provenance metadata to Claude outputs globally as part of its response to the EU AI Act’s transparency rules. Anthropic has not disclosed its technical implementation. An informed reconstruction by Kai Magnus and Daniel Miessler suggests durable text watermarking is likely to rely partly on statistical patterns in model word choice. Such a signal can survive copying but weakens as the text is rewritten or paraphrased.

Detection cannot settle the question a court declaration asks. A human-written document passed through Claude for proofreading may carry a mark, while heavily rewritten AI-generated text may not retain enough of one to detect. The mark can indicate that Claude processed the text at some stage; it cannot by itself establish who authored the substance or how much judgement the human exercised. That distinction reaches beyond court filings into privilege reviews, disclosure exercises and internal investigations. Organisations should record AI use when it happens rather than trying to reconstruct authorship from the finished document: what went into the system, what came back, what changed, who checked it and who approved it. Technical provenance can support that record, but responsibility still rests with the person signing their name.