Artificial Intelligence

Our deployment of Harvey, an OpenAI-backed tool based on GPT-4, began with a sandbox. In other words, we gave access to a limited number of lawyers in a ring-fenced environment. Sandboxes are crucial for any business looking to deploy generative AI because it’s hard to predict what the technology will do until you use it. We tested, adapted, and moved ahead – all in a safe and secure environment. We only rolled out Harvey to a wider group once we could mitigate its risks, and we continue to gather and act on feedback we receive.

We also established an AI steering committee and an AI brains trust to help our experts understand AI’s current and future capabilities and how it can be harnessed across every area of our business. Alongside this, all our existing governance structures, including our risk committee, now consider generative AI in their day-to-day decisions.

Clear governance and guardrails are critical to successfully deploying AI. We have specific rules of use in place and train our people how to use AI tools effectively and safely.

People are the common thread that runs through all our work with AI. We know that generative AI is an augmentative tool. Everything Harvey produces is rigorously checked, edited and finessed by our team. It enhances the work our lawyers do and helps us produce better results for our clients. In turn, it is governed and augmented by the gold-standard critical thinking and creativity for which A&O lawyers are known.

We combine a sophisticated understanding and experience of technology with deep expertise in intellectual property, data privacy, regulation, technology transactions, litigation and change management.

We help clients to manage the risks associated with this powerful technology which fall under two broad categories.

First, AI models make errors. Crucially, even those who build and train the models can’t explain and account for them. This so-called “black box” problem creates significant risks.

• Hallucinations: These are incorrect outputs that could lead to, for example, tort liabilities, consumer harm or regulatory breaches. Hallucinations can be the result of incorrect or out-of-date data, inaccurate mathematical predictions based on weighting of sources or randomisation, or historical bias in the datasets used to train the models.
• Unpredictability: A lack of explainability also creates a lack of predictability: you can’t be certain exactly what the model will say in response to a question. This can make it extremely difficult to check that it meets standards of quality and accountability.
• Response divergence: By their very nature, AI models will give multiple answers to the same question. This could be evidentially relevant if, for example, an AI chatbot built to give financial advice delivers different responses to two individuals leading to divergent outcomes.

Second, generative AI models take human content and account for it in a mathematic response. A user may therefore be working with someone else’s information without permission, credit, knowledge, or even awareness. This raises significant IP Infringement questions: for example, can the user assert ownership over the model’s output? And is their own IP safe if they are using the model?

There are also consequential questions about data privacy and data protection, for example, where an AI model has been trained using personal data or a user inputs personal data as a prompt.

Our AI Group provides answers to these substantive legal questions on a syndicated basis. You can sign up to join a series of one-hour calls with other businesses, each in a controlled environment monitored by an antitrust lawyer present. The calls deal with specific issues and are supplemented by minutes and additional written materials such as formal memos, policy guidelines, or comparative analyses.

So far, we have covered topics including a primer on AI, ChatGPT policy, IP infringement and data risks, licensing an LLM and change management, and have welcomed attendees from industries including financial services, pharma, technology and telecoms.

For more information, please get in touch with your usual A&O contact.

ContractMatrix streamlines contract drafting, review and analysis using:

• Generative AI-assisted interrogation and drafting
• Real-time access to your gold-standard precedents and policies
• Inbuilt risk management and governance designed by A&O lawyer

It has been developed in partnership with Microsoft and Harvey, which builds custom LLMs for lawyers. ContractMatrix has been tested and perfected by over 1,000 A&O lawyers and, following launch at the end of 2023 with Financial Times coverage, is now being licenced to clients. 

Please get in touch to find out more, or request a demo.