Genie AI, a leading artificial intelligence (AI) and machine learning (ML) technology company, today provides an update on the progress made while developing the UK鈥檚 first AI Legal Assistant (image provided below).听
Genie AI are building proprietary Large Language Models (LLM) which lean on OpenAI鈥檚 GPT-4 technology. GPT-4 passed the multiple-choice portion of the Uniform Bar Exam (UBE), and both the written components, exceeding all prior LLMs by scoring in the top 10% of all humans who take the legal exam. For reference, GPT-3 scored in the bottom 10%.听
Exponential improvements in the ability of foundation models such as this, came from increasing the scale of the underlying model, rather than fine-tuning for law. In other words, GPT-4 outperforms smaller AI models that have been specifically trained on legal datasets.听
London-based Genie AI however, will outperform GPT-4 at these legal tasks by the end of the year (As measured by accuracy under UK law):听
- Feature 1 – Real-time document-level workflows. Users can type in any query at the document level, e.g. 鈥淪uggest standard clauses relevant to [UK-based Series A SaaS Companies with a Subscription model], if missing from this document鈥�, and watch as the AI Legal Assistant completes the work in real-time.
- Feature 2 – Red-flag review. With every AI-assisted document, clauses will automatically be flagged as red, yellow, or green based on their riskiness and favourability for you or the counterparty (and explain its working), then provide in-line suggestions which a user can accept or decline based on an explanation.
- Feature 3 – Legal explanations. The AI Legal Assistant simplifies, summarises or explains the whole document or clause. Legal expertise from senior legal counsel lies within the prompts used to generate these UK-specific legal explanations.
Further, Genie AI will go beyond the simple generative chat interface popularised by ChatGPT, Bard and Bing. Instead, Genie鈥檚 AI Legal Assistant will be built native within the Genie editor so it can comment, review, reformat, highlight, annotate, suggest and export, just like a human can, making Genie AI the UK鈥檚 first ‘multi-modal’ AI Legal Assistant.听
More from News
- Experts Comment: Is Britain’s Reliance On Foreign Technology Becoming A National Security Risk?
- Full Fibre Adoption Continues To Accelerate Across The UK
- OpenAI, Anthropic And Google Are Discreetly Building Their Own AI Standards Body 鈥� What Would That Actually Decide?
- How Have Digital Nomads Created A New Market For Businesses?
- England鈥檚 New Tourist Tax Has No National Cap 鈥� What Does That Mean For Hospitality Businesses?
- The EU Cyber Resilience Act Starts Today: Can Businesses Really Report A Cyberattack In 24 Hours?
- Can Binge-Watching Be Addictive By Design? Inside The State Lawsuit Against Netflix
- Could Australia鈥檚 Proposed Algorithm Rule Break The Current Echo Chamber That Is Social Media Today?
To outperform GPT-4, Genie鈥檚 AI Legal Assistant:听
- Uses legal and commercial datasets spanning 100,000s of business contracts and legal precedents
- Relies on 100s of prompts engineered by senior lawyers, and tailored to legal use cases 鈼� Queries several other models (in addition to GPT-4)
Genie鈥檚 1500 open-source legal templates are now used by 1000 law firms and companies with in-house legal teams, 10,000 organisations overall, and 200 more join every day.听
The proprietary data generated by the comments, suggestions and edits made by these users is then anonymised, structured and fed into Genie AI鈥檚 own LLM to fine-tune.听
Alex Papadopoulos, ML Research Scientist at Genie AI, said: “We are proud of the progress we’ve made in developing our AI legal assistant. Proprietary usage data combined with 100k+ megacap company agreements form the foundation of our LLM. With our continuous customer research, we will develop the world’s best AI Legal Assistant for the UK and beyond.”听
Genie AI is also testing its AI Legal Assistant in other complementary common law jurisdictions (including the United States, Canada, India, Australia and New Zealand).听
