Anthropic has just announced that Claude will start leaving an invisible mark in the text it generates, in an announcement that may have unintended consequences for expert witnesses.

Supported Claude models will embed machine-readable watermarks into generated text. Anthropic says those marks will travel with the words when copied and pasted into other documents and may survive subsequent editing.

The immediate purpose is provenance. As the volume of AI-generated material increases, watermarking offers a potential means of identifying content that has passed through a particular AI system.

For expert witnesses, however, that raises a rather more difficult question: if an expert report contains an AI watermark, what exactly does that prove?

Perhaps less than might first appear.

Detection is not the same as authorship

One of the most important qualifications comes from Anthropic itself.

Anthropic cautions that detecting its watermark does not establish that Claude wrote or originated the material. Rather, it indicates that the content may have been processed by Claude.

That distinction could be particularly important in expert evidence.

An expert might write an entire report themselves and subsequently ask an AI tool to correct spelling and grammar. Another might use AI to improve the structure of material they have already written. A third might provide their analysis and ask an AI system to turn it into polished prose.

Those activities are quite different from asking an AI system to examine the underlying evidence, perform the analysis and suggest the opinion the expert should express.

Yet all involve AI-generated or AI-processed text.

Consider four possible uses:

• An expert writes their report and uses AI to check spelling and grammar.
• The expert writes the substance but asks AI to improve its structure or expression.
• The expert provides their analysis and asks AI to draft passages expressing it.
• The expert asks AI to analyse the evidence and assist in formulating the opinion.

All involve AI, but they plainly do not raise the same questions about the independence of the expert’s analysis and opinion.

That makes the presence of a watermark potentially useful evidence, but not necessarily evidence of what might initially be assumed.

Provenance of the text is not provenance of the opinion

This is perhaps the central issue for expert witnesses. A watermark may provide evidence about the provenance of the words. It might indicate that Claude, or another AI system using similar technology, has been involved in generating or processing them:

• It cannot, by itself, establish the provenance of the opinion.
• It does not tell us who examined the evidence.
• It does not tell us who identified the relevant issues.
• It does not tell us whether the expert carried out the underlying calculations or analysis.
• And, most importantly, it does not tell us whether the conclusion was independently reached by the expert or suggested by an AI system.

The same paragraph could conceivably carry an AI watermark whether Claude corrected three grammatical errors in it or generated the paragraph from scratch.

For expert evidence, those are very different propositions.

The absence of a watermark proves little too

There is also a problem in the opposite direction.

No watermark does not necessarily mean no AI.

Anthropic acknowledges that detection can become more difficult where generated material has subsequently been heavily edited, paraphrased, translated or combined with other text. A passage may also simply be too short to provide sufficient information for reliable detection.

This creates an important limitation if watermark detection ever begins to be used when examining expert reports.

A positive result may indicate AI involvement without establishing the extent or nature of that involvement whilst a negative result cannot necessarily establish that there was no AI involvement at all.

Watermark detection should therefore not be treated as a binary test of whether an expert “used AI”.

A new question in cross-examination?

The development also raises an interesting practical possibility.

“Your report contains a watermark indicating that it was processed by Claude. How did you use it?”

The answer might be entirely unremarkable.

The expert might explain that they wrote the report themselves and used Claude to check its grammar. They might explain that they used it to condense a lengthy passage they had already written. Alternatively, the answer might reveal that AI played a much greater role in analysing material or formulating the opinions contained in the report.

The takeaway is that whilst the watermark itself cannot distinguish between those situations, it may, however, provide the starting point for questions which can.

That means the increasingly important question may not simply be:

“Did you use AI?”

It may instead be:

“Precisely how did you use it?”

Watermarking is becoming a real technology

Anthropic is not alone in pursuing this approach.

Google has already developed SynthID Text, a watermarking technology for AI-generated text. Rather than inserting a visible label or conventional metadata, SynthID subtly affects the model’s choice of tokens as text is generated. Across a sufficiently long passage, those choices create a statistical signature which can subsequently be detected.

Anthropic has not publicly confirmed that Claude’s watermark uses the same technical method, so the two systems should not be assumed to work identically.

The significance is broader. Text watermarking is moving beyond academic research and into mainstream generative AI products.

If the technology becomes widespread, lawyers, courts and experts may increasingly encounter claims that a document has been identified as having passed through an AI system.

Understanding what such a finding does, and does not, establish will therefore become important.

The expert remains responsible

The Academy’s guidance on the use of Artificial Intelligence by expert witnesses already addresses the more fundamental issue.

AI can be a useful tool, but it cannot substitute for the expert’s own expertise, analysis and opinion. Experts remain responsible for the evidence they give. Where AI has played a significant role, keeping appropriate records of how it was used may also become increasingly important, particularly if the expert is subsequently required to explain that use in court.

Watermarking adds another dimension to that responsibility.

It may make some uses of AI more readily identifiable. But identifying that an AI system touched the words is not the same as establishing who produced the intellectual work behind them.

For expert evidence, that distinction is crucial.

Provenance of the text is not necessarily provenance of the opinion.

As AI provenance technology develops, the ability of an expert to explain what the AI did, what the expert did, and where the opinion actually came from may become increasingly important.

The Academy’s GDPR guidance turned 5 this year so we thought it was a good time to review it in the light of a better understanding of how GDPR has been implemented in practice. One of the challenges of developing guidance for our expert members is they cover a wide range of specialisms and as a result deal with widely differing types of personal information. The material needed by a medical or psychiatric expert to assess a personal injury is clearly very different from that used by an accountant to assess the value of a business.

In addition while some experts work within large consulting firms many are micro-businesses who don’t have the benefit of in-house egal advice but still have to negotiate the (sometimes onerous)  privacy responsibilities passed down to them by instructing solicitors.

 

Our guidance is designed to help members consider for themselves what aspects of their work are covered by GDPR in a structured way and is linked to a model privacy policy that can be used in a ‘pick and mix’ way to include only those sections that are relevant.  We recommend micro-business or individual expert read the ICO’s small business guide alongside our guidance and remind you that you are almost certainly required to register and pay the required annual fee.

 

The Academy  provides a range of practical advice and guidance on other topics such as conficts, Joint Statements, privilege and so on which is available at no charge to members in the Academy’s Knowledge Hub.

Members can also call our Technical Helpline for help and advice on ethical and practical issues from dealing with instructions to getting paid.

If you would like to join the Academy please complete our Pre-Application Questionnaire and we’ll help you identify the appropriate membership level for you. Other benefits of membership regular member meetings to discuss issues affecting experts and access to specialised PI insurance.

Ignoring AI for the moment, the use of computer models of various kinds has become routine in most if not all professions, whether it’s a simple spreadsheet or a sophisticated engineering in model. It should be no surprise then when such models  make an appearance alongside experts in disputes.

Some recent cases though point up the importance of the expertise that underlies any model and that an expert’s choice of inputs and their interpretation of the outputs is critical. Andrews v Kronospan is a case about dust, smells and other nuisance alleged to be caused by a factory. Experts in dust and wood fibre analysis met and agreed a set of parameters to be used in modelling the likely extend and impact of the factory’s activities, but both the Defendant’s experts for different reasons decide sometime later that these parameters were invalid. They came up with different parameters and different results, that favoured their client, although both provided detailed explanations and reasoning for their decision.

As the judge explained, neither expert has “been able to persuade me that their change of approach was not at least partially influenced by their desire to see whether or not their further analysis would benefit their clients’ case more than their existing analysis. I am not for one moment suggesting that they did not fully and properly discharge their duties as independent experts but I do consider that their evidence has to be considered with some caution as a result. Moreover, I am not satisfied on my analysis of the evidence that I can safely prefer their opinions based on their changed approach to those made by their opposite experts who proceeded on the basis of the initial common approach. “

In a completely different field, fingerprint analysis, there was no criticism of the inputs to an automated fingerprint comparison system. The difficulty was that one expert accepted the computer output as unequivocally correct, whereas the other, with the benefit of 30 years working on fingerprints was able to explain why some features of a fingerprint should be ignored because they typically appear when the print is smeared or of poor quality. Ultimately the judge preferred a manual approach illuminated by scientific knowledge and expertise over one that  treated the computer model as an infallible black box.

 

The message to experts is clear, you need to understand not just how to use a computer model, but what the right inputs are and how to interpret the outputs. Otherwise as the saying goes in the IT Industry you will have Garbage In and Garbage Out!

 

The Academy of Experts has a rigorous accreditation process to ensure that Practicing Members have appropriate professional skills and qualifications, have been trained in or have experience of  the role and process of being an expert witness and can provide appropriate references or evidence of their skills.

If you’re a lawyer looking for an expert you can search the Academy’s register for an accredited expert or contact us directly for assistance.

If you would like to join the Academy but are unsure if you have the appropriate skills please complete our Pre-Application Questionnaire and we will help you identify the appropriate membership level for you.  You don’t have to be already practising as an expert or have testified in court to join.

Other benefits of membership include guidance and advice on being an expert, regular member meetings to discuss issues affecting experts and access to specialised PI insurance.

The start of a new school term often feels like a time for new beginnings at work as well. If you’re looking to expand your raneg of work or just feel its time to think again about your approach to marketing our next Member event is designed to help you do just that.

Hear from author and social media expert on why we need to have an online presence for career success and top tips for how you can create a polished profile. The session is highly interactive and there will be an exercise for people to complete.

Tim will walk you through how the world has changed with social media and how that impacts your business and you professionally. If you are looking to raise your profile, showcase your achievements, or looking for that next job role or referrals, people are looking for these things online.

• How to have a strong online presence so you are not invisible in this modern world
• How to be found when people are looking for expert witnesses
• How to get referrals either work related or job related

This is an online event for members only 12.30pm 15th September 2025. 

 

 

The London International Disputes Week (https://lidw.co.uk/) is a major event in the annual calendar. It provides a conference and a series of events hosted by barristers’ chambers, law firms and experts active in the provision of legal dispute resolution services. It is of the nature of the market, that although many practitioners are based in London, this is not a requirement, and many delegates come from around the world. Members of the judiciary are also frequently to be seen.

For members of the Academy considering how to build their networks and promote their practices, this is an intense time. Bring plenty of business cards! It provides a good opportunity to meet those with whom you may serve, or may at some time require your services. Members of the Academy were out in force in the June 2025 sessions, others will be marking their diaries for the first week of June 2026.

Members who are part of large dispute groups may find that their firm collaborates with others in putting on an event. Around 114 such member-conducted events were conducted this year, catering for every niche interest. This attracts delegates from far and wide, providing an unrivalled opportunity to maintain connections. This year, themes of arbitration and artificial intelligence featured strongly.  All events have to be put on by more than one firm, fostering a collaborative approach. Most are very well conducted, with effective chairs and brief, focused questions. Direct value comes from the conversations over a glass or sandwich before and after events, when networking is promoted. The subject of the seminar serves to attract others with shared interest. The content can also be of high value.

Expert work requires the maintenance of a large number of connections over many years, such that when opportunity arises, one is considered. An event such as LIDW has a natural place in the mix of marketing activities. However you seek to engage, this is worth considering.

 

Our photo shows a panel at LIDW 2025, discussing the impact of AI on dispute resolution. Courtesy of CMS

While we still have London Legal Walk on 17^th June and  a session AI and Expert Evidence on 3rd July to look forward to those will be our last Member events until the autumn.

Over the last few years we’ve had talks about Public Enquiries, Civil procedure Rules, AI, Collaborative Contracts, Use of Statistics and many other topics of interest to experts. We’re already starting to plan our Autumn programme and would welcome ideas for events and speakers from members and the wider legal & expert community.

Is there a topic close to your heart you think experts should be more aware of – or can you bring a new slant to well worn subjects like expert independence, disclosure or any of the myriad challenges we all face in discharging our responsibilities to the courts? If you can please get in touch.

All our events provide the opportunity to learn from and be supported by other experts, something that many members find invaluable. They’re free to attend for members who can also bring a Guest.

Other benefits of membership include guidance and advice on being an expert, regular member meetings to discuss issues affecting experts and access to specialised PI insurance. If you’d like to join the Academy but are unsure if you have the appropriate skills please complete our Pre-Application Questionnaire and we’ll help you identify the appropriate membership level for you.

As you may be aware the validity of computer evidence has been in the news recently because of the public enquiry into the Post Office and its Horizon system, which formed the basis of prosecutions  of sub-postmasters.

Earlier this year the Ministry of Justice issued a call for input on the use of evidence generated by software in criminal proceedings.The MoJ wished to understand how the current presumption concerning the admissibility of computer evidence is working in practice, and whether it is fit for purpose in the modern world. As the leading professional and accrediting body for expert witnesses in the UK and world-wide, The Academy of Experts convened a working group that included experts in IT and computer systems and a range of other areas of expertise who use computer-generated evidence in criminal matters. The group debated the questions posed by MoJ and developed its view to form a response.

It is the view of The Academy that the Presumption should remain, but it is vital to recognise its limitations and to welcome challenges to it. Issues of disclosure are central to addressing this situation. The reasoning and recommendations we made are set out in full in our response to the consultation.

We are grateful to everyone who contributed to the working group;  Mark Ballamy, Richard Emery, Mark Holdsworth, William Hooper, David Tonks, Gill Hunt, PeterMcArthur, Nigel Young, Jawad Hassan Zadeh

 

The diagnosis of car engine problems has come a long way since the days when a weekly check and clean of sparkplugs was necessary. Car dashboards light up with an astonishing array of warning lights and car mechanics use computer diagnostic tools to identify what’s gone wrong, and sometimes to fix it.

Unfortunately, alongside those changes we’ve all become used to sharing our opinions on social media, something which tripped up a car expert in a recent dispute about a misfiring Citroen.

In David Adam v Moneybarn the car developed faults a year after purchase and a claim, managed by Reject My Car Ltd, was  raised for the repair and/or replacement on the grounds that the car was defective when sold. Both experts agreed that a warning light was displayed and computer diagnostics showed faults but while the claimant’s expert, Mr Smith, diagnosed a misfire because ‘the whole car was shaking  and rocking badly’ and that this was a latent defect, the defendant’s expert observed no shaking and took the view that any misfire could have multiple causes rather than being a fundamental problem.

While the judge accepted that Mr Smith had both relevant knowledge and experience of mechanical faults it turned out that he had actively supported and marketed Reject My Car on Facebook and had, amongst other more temperate posts, commented “could this be the end to being treated with contempt by large, franchised dealerships?”. As the judge commented “Some posts, taken individually, might not give rise to a legitimate attack on Mr Smith’s impartiality. However, I am concerned not just to some of the more blatant individual posts but also the cumulative effect. It leads to the impression that Mr Smith has a certain view as to the car industry, and sympathy for the cases that Reject My Car Ltd take on. “

The expert was not assisted by his client neglecting to mention that he had covered 40,000 miles between buying the car and experiencing problems, the end result being that Reject My Car’s claim was consigned to the scrap heap, presumably along with the car!

The lesson for all experts is to be as careful and balanced in what you say on social media as you would be in court.  For more information on the roles of experts see https://academyofexperts.org/practising-as-expert/

Our picture is from “Grimmer Motors” in NZ