Colorado first to pass artificial intelligence law | Arsh.ai posted on the topic | LinkedIn (2024)

Colorado first to pass artificial intelligence law to protect consumers - The Colorado Sun: Colorado first to pass artificial intelligence law to protect consumersThe Colorado Sun http://dlvr.it/T74PQN #ai #artificialintelligence

1

Congratulations, Trace Jackson, on the publication of your article on AI tools in legal practice in The Florida Bar’s WCS News & 440 Report! Your forward-thinking insights are helping to shape the future of our profession. We are proud of your accomplishment and excited to see your continued impact. Read the article: https://lnkd.in/e5fw_-JdLearn more about the IP & technology law: https://lnkd.in/ejpmGMvF #artificialintelligence #ai #aiandthelaw #legalinnovation #floridabar #futureoflaw

15

2 Comments

How Should Businesses Implement Artificial Intelligence Tools, Legally - The National Law Review: How Should Businesses Implement Artificial Intelligence Tools, LegallyThe National Law Review http://dlvr.it/T88HSp #ai #artificialintelligence

RT @SSRN: Generative #AI in EU Law: This paper delves into the legal & regulatory implications of Generative AI & #LLMs in the European Union context.Read More: https://t.co/6mRZxXbZxuSubscribe: https://t.co/8kJY3k7Zc2

38

Leading legal AI research tools still hallucinate. Like, a lot. 🤖⚖️Stanford and Yale researchers tested Lexis+ AI, Westlaw AI-Assisted Research, Thomson Reuters Ask Practical Law AI, & GPT-4 on 200+ legal queries.The legal-tuned tools hallucinated on 17-33% of queries vs 43% for GPT-4. Hallucinations ranged from fabricating case holdings to misinterpreting legal actors & authority.Vertical AI is promising but hallucination-free legal AI remains elusive for now. Read more here: https://lnkd.in/eFh-5QFP

560

82 Comments

Eerie timing 😀 In the second talk mesh conference As Michael Geist shares how Canada is trying to emulate the European model of penalizing hallucinations in AI legislative bots that return wrong citations. Wrong decisions with zero transparency on how those decisions were made are more harmful so it is essential for regulators to balance between getting the investment for innovation and also being transparent. Social media in search results is marked high risk for legal AI systems in Canada something EU hasn’t yet. What about US ? - Tends to be self regulated - Looking to the industry to lead- identify self harm signals- Leading investors and commercializations of AI even though Canada may done the research - focus on the litigation has been on copywriting - what goes into the systems as inputs to the LLM - without permissions from the creators and original writers - companies are trying to follow suit by saying that they should opt out of being found with the SEO - output of the LLM - may feel like copyright infringement aren’t holding ground. For eg : New York Times articles which were copied by many other websites and was made available in various sources other than the original were used in the output in its style and exact responses because of the numerous occurrences. -

1

AI models still hallucinate. The following study was conducted on legal #AI #research models. We should not expect the situation with healthcare AI models to be much better. We are still in the early stages of AI development.#ArtificialIntelligence #Hallucination #Legal

2

In AI, an hallucination is a wrong or mis-leading response. AI is still in it's infancy, so this is too be expected for the foreseeable future. Anyone that uses AI to generate content should be aware of this. As an experiment, I opened this document in Adobe Acrobat, which now has an AI Assistant (in beta). I asked it to summarize this document, to which it's response was:"The document titled "Hallucination-Free?Assessing the Reliability of Leading AI Legal Research Tools" provides a comprehensive evaluation of AI-driven legal research tools, specifically focusing on LexisNexis's Lexis+ AI, Thomson Reuters's Ask Practical Law AI, and Westlaw's AI-Assisted Research.1 The evaluation reveals that while these tools claim to mitigate hallucination risks, they still exhibit a significant rate of hallucinations, ranging from 17% to 33%.The document discusses the limitations and implications of AI tools in the legal practice, emphasizing the need for lawyers to verify propositions manually and the importance of rigorous benchmarking and public evaluations.2It also highlights the challenges and variations in performance among different systems, providing examples of hallucinations and explaining their causes.3The document concludes by emphasizing the need for legal professionals to supervise and verify AI outputs and provides a typology for differentiating between hallucinations and accurate legal responses.4"The numbers in the summary were links to access the section of the document it summarized.

1

In response to the rapid rise of artificial intelligence in the law, the NJSBA has published a wide-ranging report outlining AI’s impact on the legal profession, with practical guidance on how attorneys can benefit from AI while navigating the technology safely and ethically. The 36-page report, created by the NJSBA Task Force on Artificial Intelligence and the Law, addresses the fundamental considerations attorneys must make when implementing AI in their legal practice. Read the full report here, https://ow.ly/t3QZ50S8ru7 #NJStateBar #NJSBA #AI #ArtificialIntelligence #ArtificialIntelligenceLaw

3

Fine-tuning an LLM for legal uses just seems like exactly the wrong approach to me... In my mind, the better approach would be an agentic workflow like this (example for legal research, but could be for other areas):1. Have an agent to pull out specific references in cases that match what you're asking from cases themselves (RAG / Summarized Chunks of Text / etc to figure out specific text to analyze), giving the specific reference, why it thinks that reference fits the case, and any other cases referenced in that chunk2. Have another agent judge the output to determine if it really is relevant3. For each "other" case referenced, recursively have agents pull out the relevant data from those (or limit it to a certain "depth" or "breadth" of search (i.e. 50 cases deep, or 500 cases in total)4. Generate a report with references (both cases + line numbers / exact text to validate)Obviously this could (and should!) be extended, with added functionality Maybe add in some research tools to find other interpretations of the cases/references specifically, and determine if you want to include those in the analysis.Maybe add in a separate agent to suggest arguments based on the references, and create a whole new agentic flow to find evidence to support/detract from those arguments.There are a world of possibilities, but fine-tuning an LLM seems like a bad way to do it.

2