I have yet to meet a lawyer who wishes they could spend more time at night solving a legal question. Or an entrepreneur who would love to wait just one more day for a response from their legal team.
When the AI transitioned into the mainstream with the arrival of ChatGPT, many posited a rapid revolution in knowledge work. Yet, closer scrutiny revealed gaps in the still nascent technology: frozen knowledge and a tendency to make things up. Both greatly limit the kind of jobs you can ask AI to do for you.
Wouldn’t it be great if we could take the truly amazing capabilities of chatbots and equip them with recent knowledge and make them less perfect liars?
I’m thrilled to share we’ve done just that for understanding the law, starting humbly with just one piece of a hotly debated regulation: the EU’s AI Act. We’ve developed a variant of AI that has an intimate understanding of that regulation. When you ask it a question, it performs whole-document reasoning to find relevant excerpts and analyzes them in tandem to provide answers of a quality that even surprised us.
Our answers include a legal trace: word-to-word quotes of the legalese surrounded with plain language commentary that connects them to the original question. By reading the answer top-to-bottom, you can easily verify whether the AI is leading you astray by making things up. You’ll spot fishy arguments when the commentary doesn’t align with quotes from legal sources and avoid having a real bad day.
The less obvious benefit of our choice to include legal traces we noticed is a much higher quality of answers. If you ask a person for a detailed line of reasoning behind their claim, you leave them with little room for fabrication. As it happens, AI can be pinned down, too.
We believe an MVP means a minimal surface area of things done exceptionally well — not many things done partway. We think we’ve done just that for applying AI to legal questions, starting with just one regulation. We’re excited to announce HotseatAI is ready for use, and eager to hear what you think.