Thomas Wolf

Two years ago to this day, we got some of the first results back on the prototype that would eventually become aiR for Review. Given just a simple set of instructions that we had crafted from the review protocols that was repeated in parallel document by document (yes, we were building Agentic AI Workflows in legal technology *two years ago*), we were calculating point measures of recall as high as 85%, and with tweaking saw final point measures of recall as high as 100%. We learned a lot in the first year of testing. One thing we observed was that acceptable precision was typically harder to achieve than acceptable recall. We put significant design, engineering, and applied science effort into improving the experience writing and improving prompts. As a result, we have a great interactive experience that allows a user to easily modify their instructions and test their performance. We also upgraded the underlying foundation model. Initial testing started with GPT-4 in 2023. After Turbo's release in Fall 2023, we switched to that. When GPT-4o was released in May of 2024, we validated and deployed it before the GA release of aiR for Review in September 2024. For each change to our model including foundation model upgrades, we gave our users plenty of notice about the upgrade, extensively tested and measured qualitatively and quantitatively, and documented and shared the differences. This week, we announced a number of changes and improvements to aiR for Review that we are either releasing or will be coming soon. First, after extensive validation, we're upgrading the underlying foundation model from the May release of GPT-4o to the November release. Despite being in the same family, the November release of GPT-4o is a substantially different model than the May release. It is more attentive to detail, easier to steer, and gets less lost in longer and complicated documents. We've observed precision improvements of as high as 10% when compared side by side with the May release. But because November is more attentive to detail, it's less likely to "reach" for a questionable document when its instructions are ambiguous. If you're a member of the Relativity community, there are more details here: https://lnkd.in/ePkRdnRB Second, we're introducing a new feature for automatically drafting prompts directly in just a few minutes from case materials such as the complaint or RFP. Remarkably, this feature has outperformed experienced aiR users in terms of its initial precision and recall. Finally, we're introducing built-in validation for aiR for Review. This validation process follows industry best practices in random sampling and providing precision and recall estimates, giving our users more confidence in the prompt that they've built, allowing them to reliably estimate the fraction of the responsive documents that they will find before they start a large-scale review.

91

3 Comments

🚀 Practical Guidance for Evaluating AI Misuse Safeguards—Exactly What We Need 🛡️ AI governance often feels like navigating a maze of abstract principles and vague directives. That’s why the UK AI Safety Institute’s new guide on evaluating misuse safeguards for frontier AI systems is such a breath of fresh air. 🌬️ It’s practical, structured, and gives developers tools they can actually use. 🔑 What Makes This Guide So Valuable? It lays out a clear 5-step plan for assessing how well AI systems are protected against misuse: 1️⃣ Define Safeguard Requirements: • Start with clear goals: What risks are you mitigating? Who are the threat actors? What’s considered an unacceptable outcome? 2️⃣ Create a Safeguards Plan: • Identify and document which safeguards you’re using—from access controls to model refusal training—and how they address specific threats. 3️⃣ Collect & Document Evidence: • Use red-teaming, bug bounties, and coverage evaluations to gather robust data on how effective your safeguards really are. 4️⃣ Plan for Post-Deployment Assessment: • AI systems evolve, and so do threats. The guide emphasizes ongoing monitoring and rapid response to new vulnerabilities. 5️⃣ Decide if Safeguards are Sufficient: • Bring it all together. Is the evidence strong enough? Is your post-deployment plan ready for real-world challenges? 📋 Why This Matters So Much ⚡ Theory is easy. Implementation isn’t. This guide doesn’t just tell you to “be responsible”—it shows you how to be responsible with actionable steps and templates. 🛠️ Practicality is everything. AI misuse isn’t a distant threat—it’s happening now. Developers and regulators need tools like this to make informed, timely decisions. 🌍 We need more resources like this—clear, focused, and designed to help teams actually do the work, not just check compliance boxes. 💬 Your turn: How are you approaching AI misuse safeguards? What tools or processes have made the biggest difference for you? Let’s share solutions. ⬇️ #AIGovernance #AISafety #ResponsibleAI #AICompliance #FrontierAI #MisusePrevention #AIEthics __________________________________ Did you like this post? Connect or Follow 🎯 Jakub Szarmach, AIGP, CIPM Want to see all my posts? Ring that 🔔

62

1 Comment

PATENTING CORE AI IN THE UK - BACK TO THE PAST Following the Court of Appeal decision in Emotional Perception AI, which reversed the previous High Court decision that Artificial Neural Networks per se are 𝘯𝘰𝘵 excluded from patentability as a "program for a computer", the UKIPO has again updated its examination guidelines (see link in comments) for AI to take account of that. No surprise here, we are back to the status quo prior to the EPAI High Court decision. At least until this case reaches the Supreme Court, which we expect to happen later this year. Will we go back to the future then? 𝗖𝗼𝗿𝗲 𝗔𝗜 - 𝗘𝗣𝗢 𝗮𝗻𝗱 𝗨𝗞𝗜𝗣𝗢 The UKIPO guidelines have a helpful discussion of EPO decisions that could be considered to relate to Core AI but a key point is missed. While in all these decisions the application was refused because the applicant failed to demonstrate a technical effect across the whole claim scope, this does not in itself tell us whether Core AI is patentable or not. Rather, the cases were refused because the technical effect was not plausibly present over the whole claim scope. For example, merely shrinking the size of a neural network will save compute but if the covered embodiments include shrinking a network with commensurate degradation in performance in line with any compute savings, there is no technical effect. The key-point is here that these EPO decisions do not close the door on a technical effect being present where computational efficiency in a way that is specific to the way computer work. They only underline that this needs to be achieved across the whole claim scope. To what extent the same might apply under the "a better computer" heading at the UKIPO remains to be seen. 𝗧𝗵𝗲 𝘄𝗶𝗱𝗲𝗿 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 - 𝗔𝗜 𝗶𝗻 𝘁𝗵𝗲 𝗨𝗞 With the current government's focus on making the UK the most AI friendly country in the world, it is perhaps ironic that our patent system seems to be taking a step in the opposite direction. However, it is not clear that this will have a negative impact on the AI eco-system in the UK. For example, there is no indication that the similarly restrictive patenting regime in the US is hampering AI innovation and growth there. What is more, recent developments have shown the commercial relevance of open source approaches. Deepseek's open release of its R1 reasoning model, which competes with the state of the art at a fraction of the cost and compute resources, suggests that the patent system may not be that central to the growth of AI innovation. If the clearly continuing trend started by Meta's semi-open release of Lllama models continues, the UKIPO's (and EPO's) stance of focussing patentability on the application of AI may yet turn out to be compatible with how the wider ecosystem develops.

36

1 Comment

🎯IP Management Pulse #44 – New Edition Out Tomorrow! The latest IP Management Pulse newsletter is here with key insights from the IP world: 1️⃣ IP in Technological Obsolescence – the shift from traditional IP to data-driven models, including platform licenses and trade secrets. 2️⃣ IP Roadmaps - Making Strategy Actionable – 😎 Alexander Johann Bjørnå explains how to implement business strategies through IP roadmaps. 3️⃣ Copyright & AI Training Data – 😎 Dr. Barry Scannell highlights a $1.5 billion settlement over AI training data usage. 4️⃣ Systematic Trade Secret Management – 😎 🔑Maria Boicova-Wynants shares why proper management is essential for trade secrets. 5️⃣ Timing of Patent Filings – 😎 Maier Fenster explains the crucial balance of filing patents at the right time. 🌟 Deep Dives: BHS Corrugated Case Study – 😎 Maria Boicova-Wynants and Dr.-Ing. Martin Bookjans dive into how BHS turned IP into a competitive strength. Explainable AI in the Courtroom – 😎 Til Martin Bußmann-Welsch from Anita Legal explains how AI transforms legal practices. Additional topics include IP strategies in the service economy and Adobe’s Creative Cloud success. When you want to know whom to follow on LinkedIn, then look at Amelia Skelding, IP Director at Swindell&Pearson, and Wouter Pors Partner at Windt Le Grand Leeuwenburgh. 📝 Thought Leadership by 😎 Giulia Donato – Giulia discusses how IP experts can leverage publishing for visibility and trust in the market. 📬 Subscribe now and catch all this and more in the next IP Management Pulse — out this Thursday! I'll give you the link in the comments 👇 🤗 Thanks to our sponsors: Patent Cockpit; uptoIP® | ABP PATENT NETWORK GmbH; infoapps GmbH; Bösherz Goebel Patentanwälte; Connor 🙏 Thanks to our co-authors - without whom this newsletter would not be possible: Amr Eldakak; Archana Pandit; Dr. Benjamin DELSOL (PhD, LL.M); David Durand; David Kalow; David Perkins — Software Patent Expert; Donal O'Connell; Dr. Dominique Christ, Erdem KAYA; Eric Stasik; Henry K.H. Wang; Dr. Ines Duhanic, LL.M. (Stockholm); Ivan Gauci Dr. Jörn Plettig; Martin Schweiger; Mustafa Cakir, RTTP, QPIP; Magnus Hakvåg, 🔑Maria Boicova-Wynants; Philipp Hammans; Prof. Kelvin Willoughby; Ricardo Cali; Udo Gennari _____________ 🚀 Hi, are you an IP expert? We're looking for IP subject matter experts for the entire innovation chain of companies. Want to work with firms in Europe to protect their innovation results and strengthen their innovative capacity? I'll give you the Link to the Call 2025 in the comments 👇 - and just send me a direct message here on LinkedIn.

69

14 Comments

Post #2 in the Quantum IP Strategy series (post 1 here: https://shorturl.at/iPZS7): In the quantum industry, it appears that most companies follow traditional IP advice. Usually, that's an okay approximation if you haven't spent any effort thinking deeply about your IP strategy. But quantum computing isn’t like other industries, and its unique properties need to be considered: ⚛️ Long commercialization timelines – A patent lasts 20 years, but will your invention still be relevant by then? ⚛️ Fast-moving research – A technique that seems novel today might be outdated in five years—or even sooner. ⚛️ A fragmented ecosystem – Most quantum companies are either small startups spun out of academia or tech giants with deep pockets—creating an IP power imbalance. ⚛️ A pre-competitive landscape – Much of quantum computing is still in the foundational R&D phase, meaning collaboration opportunities are significant and may be highly beneficial—but IP barriers and the perception of competition could shape how (or if) that collaboration happens. ⚛️ Sustainability – Most quantum startups aren’t self-sustaining in revenue and likely won’t be for many years. That means an IP strategy needs to align with funding cycles, partnerships, and licensing opportunities, not just long-term exclusivity. And you certainly don't do this by just patenting everything. So the big question is: How should quantum startups and research teams approach IP differently than other industries? In the next post, I'll talk about some of the above points and what you SHOULD do to take some of these factors into account. If you're working in quantum, how would you think about taking the above factors into account? Let’s discuss. 👇 #QuantumComputing #Patents #IPStrategy #QuantumLaw #Innovation

29

11 Comments

A big win for the ICO yesterday on their £7.5m ClearviewAI fine! (and a big win for Privacy International's intervention!) The Upper Tribunal found that Clearview AI's processing *is* in scope of UK #GDPR and the ICO *does* have jurisdiction - the First Tier Tribunal having found that Clearview AI's assistance of overseas law enforcement having taken them outside the scope of UK GDPR. The matter will now be referred back to the FTT more generally. Case ref: The Information Commissioner v Clearview AI Incorporated [2025] UKUT 319 (AAC). Full copy from Brick Court Chambers: https://lnkd.in/gFFxN_GX https://lnkd.in/g_7SZFGS #dataprotectionlaw

46

💡 A Primer on Large Language Models and their Limitations 👉 a paper on the strengths, limitations, applications of and research directions for LLMs 👉 according to the authors intended to be useful to those in academia and industry who are interested in gaining an understanding of the key LLM concepts and technologies 👉 and in utilising this knowledge in both day to day tasks and in more complex scenarios where this technology can enhance current practices and processes 👉 definitely a very good paper for anyone starting into exploring the world of LLMs, well written with explanations of relevant vocabulary 👏 👉 highlighting the potential use cases and risk scenarios as well as positive and negative impacts of LLMs #artificialintelligence #innovation #law #future #riskmanagement

33

2 Comments

Garfield AI - world's first AI law firm Project Lasagne? Probably a post you would have expected on April 1st but no - this is real. ➡️ “The first regulatory approval of an AI-based law firm is a landmark moment for legal services in this country,” ⬅️ said Paul Philip, the SRA’s chief executive. This is fascinating and whilst currently fairly low level administrational work is surely a sign of what is to come Currently Garfield have built in human intervention ➡️ "All the model’s output will initially be reviewed by Young (founder) , before moving to a sampling system to check for accuracy and quality." ⬅️ but for me what is critical here is whether the appropriate AI Governance is in place. Without AI legislation in the UK we will be relying on Data Protection laws and the ICO to regulate. We believe that ultimately consumer trust will determine the success of this type of initiative and robust AI governance is the foundation https://lnkd.in/eKQRWH6U https://lnkd.in/esyWUR_p

27

2 Comments

Does your IP team have the trust of your inventors? How can you build more of it? Here’s what I’ve learned after spending years on both sides, first as an engineering PhD and then as an IP leader in industry. IP-inventor trust exists when inventors: - Prioritize IP tasks within their busy schedules - Respond to your — and outside counsels’ — emails - Knock on your door with new ideas - Feel comfortable enough to voice their concerns Conversely, when legal requests get ignored with radio silence, you know that trust is missing. To work on these relationships, I’ve gotten results from: 1. Being physically present. At HP Labs, my desk wasn’t where the majority of the legal department was, but instead in the labs space. That simple visibility went a long way and exposed me to opportunities to deepen my technical understanding. 2. Attending brown bags and seminars. Aside from teaching me about what my inventors were working on and their day to day operations, attending their events showed them that I was interested in their worlds and trying to be part of it. 3. Providing business context. Helping R&D piece together how their ideas fit into the organization’s business strategy let them know that I was a true partner in identifying patentable ideas, not just a legal checkpoint. 4. Finding balanced solutions. Don’t just enforce policies, work with your inventors to improve them. Banning phones in the lab outright was one of the policies from back in my day that we had to adjust. You won’t improve protection if your inventors don’t go for it. What are your tips for gauging and building trust with your inventors? #intellectualproperty #IPstrategy #inhouseIP

37

I often hear AI is here to “replace” parts of patent drafting. I don't think that's the solution. In my experience, rushing to let AI do everything at once produces lazy, shallow drafts. The real magic happens when you slow the process down and make the AI work with you, as your thinking assistant. It is not about speed. It is about depth, precision and consistency of thought. What's your take: Are we patent attorneys too obsessed with faster, when we should be chasing better? _____ Hi 👋 it's me, Bastian. I share my best insights on patents and AI with my 1,000+ newsletter readers! ⬆️ Click "View my newsletter" under my name ♻️ Please repost if this was useful

48

34 Comments

Really interesting announcement by Anthropic on Confidential GenAI. This effectively wipes out the competitive advantage of those few legal tech vendors whose only distinctive quality is to anonymise data before using it. Using encryption at the hardware level is so much better than taking the cumbersome and error-prone approach of anonymising text locally (in reality rather pseudonymizing it) before sending it to an LLM, and then afterwards de-anonymising it again. https://lnkd.in/d9AMCtX6

63

6 Comments