In partnership with
Most of the AI risk this newsletter routinely covers concerns what a tool produces. This week the question is what it destroys. JUSTICE has warned that AI transcription tools used across family justice generate text to fill gaps in poor audio and then delete the original recording, often within thirty days. Elsewhere Harvey raised $550 million and bought a business which monitors AI agents, OpenAI released a new frontier model, and a UK AI governance platform picked up a Magic Circle client.
AI in Practice
The transcript survives, the recording does not
JUSTICE published a warning on 9 September about AI transcription tools in the family justice system, written by its deputy legal director Ellen Lefley. Two problems are identified. The first is familiar: some transcription tools fill gaps in poor quality audio by generating plausible text rather than marking the passage as inaudible. That behaviour is a design feature rather than a fault, because it improves the apparent quality of the output. The second is new to this writer, and it has had far less attention. Many of these tools delete the original recording automatically, in many cases within thirty days, so by the time anyone questions the transcript there is nothing left to check it against.
The evidence for the first problem is not speculative. The Ada Lovelace Institute's February research on AI transcription in social work found instances in which AI-generated summaries recorded that a person had expressed suicidal thoughts when the person had not. Lefley's point is that a court's ability to find facts rests on being able to test a record, and where the audio has gone the transcript stops being a record of what was said and becomes simply a document which asserts it.
The obvious application is care proceedings, where a social worker's note of a conversation may be the only account of it and may be squarely in dispute. This author would put the point rather more broadly. A firm running an AI notetaker across client meetings, conferences with counsel and telephone attendances is generating exactly the same class of document, and most of those tools carry a default retention period which nobody at the firm has looked at. When a client says they were never advised of something, the attendance note is what the firm reaches for, and a note produced by a tool designed to smooth over an inaudible passage rather than flag it, with the audio then since purged, is a weaker document than the contemporaneous manuscript note it replaced.
Two comparisons are worth drawing. The police and the Crown Prosecution Service have already agreed joint operating procedures for AI transcription in criminal work which stress retaining the original audio, so the retention question has been answered once already in a neighbouring part of the system. HM Courts and Tribunals Service has been piloting its own in-house tool, Justice Transcribe, across Crown court and family proceedings since the spring, on the reasoning that transcripts currently cost victims hundreds or thousands of pounds and that AI might make them affordable. Both propositions can hold at once: the access to justice case for cheap transcription is real, and so is the evidential case for keeping the source material, and they only pull against each other where nobody settles the retention question in advance.
Takeaways
Act: Find out what retention period your firm's AI notetaker or transcription tool applies to the original audio, and whether it can be changed. If the answer is thirty days and nobody at the firm chose that number, a vendor has chosen your evidential position for you.
Watch: What HMCTS publishes from the Justice Transcribe pilot, and in particular whether it says anything about retaining source audio, because that will become the reference point everyone else is measured against.
Risk: An attendance note or care record which cannot be verified, produced by a tool built to fill gaps rather than mark them, in a matter where that note is the only account of the conversation.
Read: JUSTICE, Ada Lovelace Institute, and the reporting at Legal Futures
On your radar
Harvey has raised $550m and bought a company which watches what AI agents do: Harvey announced on 9 September a $550 million round at a $15.5 billion valuation, co-led by Diffusion and Lightspeed, alongside the acquisition of Guardrails AI, a San Francisco business whose open source tools detect when an agent departs from its intended behaviour and manage that risk while it is running. Its co-founders Shreya Rajpal and Zayd Simjee join Harvey's product and engineering teams, and it is Harvey's fourth acquisition this year. Harvey says 80% of Am Law 100 firms use the platform, though it gives no equivalent UK figure. Why it matters for UK lawyers: the valuation is not that interesting but the purchase is, because a vendor buying agent monitoring three weeks after the SRA put supervision at the centre of its warning notice echoes the sentiment this newsletter expressed in recent weeks. If your firm is on Harvey, ask your account manager whether the Guardrails technology will show anything to the firm's own risk team or only to Harvey's engineers. (Artificial Lawyer, Legal IT Insider)
OpenAI's new frontier model landed on 3 September, and the legal reaction is about supervision: GPT-6 Astra was released on 3 September, with better computer use and stronger performance on long multi-step tasks. Legal IT Insider gathered early reactions, including Natalie Foster of Legally Fond, who found it vastly different from its predecessor and described "a much stronger route to an outcome rather than having to constantly replenish prompts", and Harvey's head of applied research Niko Grupen, who called it a significant quality improvement on complex legal work. The cautions in the piece are useful. The agentic runs remove the tasks junior lawyers used to learn on, that greater autonomy needs an audit framework to match it, and Matthew Letts's warning that leaning on frontier models creates a supply chain dependency when a good deal of the work could run on smaller local ones. Why it matters for UK lawyers: your firm does not choose when its legal AI vendor swaps the model underneath, and a model which completes a long task in a single run changes what a supervisor is actually reviewing. Ask each vendor what its upgrade path is and whether the firm will be told before a change happens. (Legal IT Insider)
In-house agents which run without anyone pressing a button: Chamelio launched an Agent Hub on 8 September for in-house legal teams, built around agents which work continuously rather than on request. The examples given are chasing signatures which have sat for more than five days, producing weekly summaries of where deals are stuck, watching for a customer milestone which triggers an amendment, and reviewing renewals sixty days out. Chief executive Alex Zilberman put the pitch plainly, saying the promise of AI is "taking work off their plate entirely" rather than doing the same work a little faster. Why it matters for UK lawyers: an agent which acts because a condition was met, rather than because a lawyer asked, is an agent nobody is reviewing at the moment it acts, which is a different supervision problem from the one most firm AI policies were written to solve. If you sit in or advise an in-house team looking at this category, settle before the pilot which agent actions require a human approval step. (Artificial Lawyer)
A UK AI governance platform has raised $5.4m, with a Magic Circle firm among its clients: AI Score, co-founded by Alex Harland and Benita Tibb (who spent more than five years at White & Case), raised a $5.4 million seed round led by Fuel Ventures on 4 September, following a $1 million pre-seed in November 2025. The platform is sold to organisations wanting to control which AI tools are used, measure how they perform and track compliance with frameworks including the EU AI Act. Clifford Chance's global technology co-chair Jonathan Kewley is a senior adviser, and the company says its clients include a Magic Circle firm and a FTSE 250 company. Why it matters for UK lawyers: AI governance is quietly becoming something firms buy rather than something they write, and a bought tool is only worth the questions it can answer. If your firm looks at this category, test any product against the question the SRA's warning notice actually poses, which is whether you can reconstruct how a particular document was produced. (The Global Legal Post)
Ad Break
In order to help cover the running costs of this newsletter, please check out the advert below. In line with my promises from the start, adverts will always be declared.
Your voice. Every platform. No writing required.
You ghost your own socials by Wednesday. SureThing learns your voice and ships native posts to LinkedIn, X, Instagram, and TikTok, without you writing a thing.
For Review
Scribe and prejudice? Exploring the use of AI transcription tools in social work (Ada Lovelace Institute)
The research behind this week's lead, published in February and worth reading properly rather than through the reporting. Go to the material on how the tools handle inaudible passages and what they retain, because those findings transfer directly to the AI notetaker sitting in your own client meetings.
Read or listen: Ada Lovelace Institute
AI in the family justice system (Nuffield Family Justice Observatory)
A briefing from March mapping where AI is already in use across family proceedings and where the risks sit. If you have to brief a family team, or a local authority client, on what is coming, this is the single most useful document to hand them.
Read or listen: Nuffield Family Justice Observatory
Merrill – The NewMod For Litigation Work (Artificial Lawyer)
A founder interview with Rick Merrill, who has launched a California firm doing high volume litigation defence for large corporates on fixed fees per case or per tranche, built on its own technology stack rather than off-the-shelf legal AI, and aiming to undercut regional competitors by at least a third. It is a US business and the piece is an interview rather than an assessment, so read the claims accordingly. It earns its place here because fixed recoverable costs already push UK volume litigation towards the model Merrill says AI makes profitable, which makes it a reasonable preview of an argument defendant insurance practices will hear before long.
Read or listen: Artificial Lawyer
Practice Prompt
Try the below prompt to audit what your firm's AI notetaking and transcription arrangements would actually produce if a client, an opponent or a judge questioned the resulting note, which is the gap this week's JUSTICE warning exposes. It works from what you can find out about the tool and the practice around it, and what comes back is a list of questions to put and the points where the firm's evidential position is weaker than it assumes. Ensure you fill in context and constraints and other aspects marked with {}. Remember to adhere to the Golden Rules and do not upload confidential or privileged information to public tools.
You are assisting a solicitor, COLP or practice manager at a law firm in England and Wales. Your task is to audit the firm's use of AI notetaking and transcription against the question of whether the resulting note could be relied on, and defended, if it were later disputed. The output is a set of findings and questions for the firm, not an assessment of any product and not a conclusion that the firm is or is not compliant.
Context to apply:
- The tool or tools in use, and what the vendor says about them, from public material only: {name each, or say "unknown, and that is itself a finding"}
- Where they are used: {client meetings, conferences with counsel, telephone attendances, internal supervision meetings, court hearings, site visits}
- Whether use is firm-licensed, individually subscribed, or embedded in another product such as a video conferencing or case management platform: {describe}
- What happens to the original audio or video, and after how long: {retention period, whether it can be changed, or say "nobody has checked"}
- What happens to the transcript, and where it is stored: {document management system, the tool's own cloud, an individual's mailbox}
- Who checks the transcript against the meeting, and when: {describe the actual practice, not the policy}
- Whether attendees are told a tool is running, and whether consent is recorded: {describe, and note any matter types where this has not happened}
- Practice areas where a disputed attendance note is a realistic prospect: {e.g. conveyancing advice on a specific risk, contentious probate, professional negligence defence, care proceedings, employment}
Produce the audit under these headings:
1. What the firm actually has
Restate, in plain English, the arrangement as described above, separating what is known from what has been assumed or left blank. Where a blank exists, say what the firm would have to do to fill it and who would know.
2. The retention gap
For each tool, work out how long the source recording survives and how long the transcript survives, and identify every period during which the transcript exists but the material needed to verify it does not. Say plainly what the firm could and could not demonstrate about a disputed passage at each point on that timeline.
3. How the tool handles what it cannot hear
Draft the questions the firm should put to each vendor about inaudible or ambiguous audio, covering at least: whether the tool marks a passage as unclear or generates text to fill it; whether confidence levels are recorded and whether they are visible to the user; whether speaker attribution is inferred; whether the tool produces a summary in addition to a transcript and, if so, what the summary is generated from; and whether any of this is exportable onto the firm's file.
4. Where the note carries weight it was not built for
Identify the matter types listed in the context where the AI-generated note is likely to be the only account of a conversation. For each, state what the note would need to show to be worth relying on, and mark whether the current arrangement produces that.
5. Consent, confidentiality and privilege
Set out the questions the firm needs to answer about who is told a tool is running, what is recorded about that, where the audio and transcript sit, whether a third party processes them, what happens on a matter closing and on a fee earner leaving, and whether a privileged conference recorded through a tool raises any issue the firm has not addressed. Mark clearly where you are uncertain about ordinary practice rather than asserting a position.
6. What to change first
Rank the findings by how much evidential exposure each carries against how easily it can be fixed, and separate changes which are a setting, changes which are a process, and changes which would need a different product. Keep this short enough to put in front of a management meeting.
7. What this audit cannot tell you
Flag every point where the output rests on an assumption about the tool, the practice or the law, and say what the firm would need to establish to replace it with a fact.
Constraints:
- {Add firm-specific constraints, for example a client who has prohibited recording, an existing licence with time left to run, a video platform the firm cannot change, or an insurer requirement on file records.}
- Apply the law and ordinary professional practice of England and Wales, and the SRA Standards and Regulations. Where you are unsure whether something is ordinary practice or a regulatory requirement, say so rather than asserting it.
- Do not invent regulatory requirements, SRA expectations, case law, product capabilities, retention periods or vendor policies. If you do not know what a tool does, write the question rather than an assumption.
- Do not conclude that the firm's arrangement is adequate, compliant or safe. The purpose is to surface the gaps and the questions.
- Write in plain English and keep the vendor questions short enough that they cannot be answered with a brochure.
- This is a planning aid, not legal or compliance advice. The firm, its managers and its COLP remain responsible for the arrangement and for everything done under it.How did we do?
Hit reply and tell me what you would like covered in future issues or any feedback. We read every email!
Thanks for reading,
Serhan, UK Legal AI Brief
Disclaimer
Guidance and news only. Not legal advice. Always use AI tools safely.
Recommended Newsletters
Below are a few newsletters that I recommend, for various reasons. Check them out!
