200 | 👎🙅‍♀️ AI deepfake nude ruins a teacher's career

In today’s Brainyacts:

1. The “accuracy” of legal AI research tools

2. OpenAI’s GPT-5 is coming

3. ChatGPT Free gets much better and other AI model news

4. Deepfakes ruin careers and other AI-related content

👋 to new subscribers!

To read previous editions, click here.

Lead Memo

🎯 ⛓️‍💥 Study Assesses Accuracy and Reliability of Leading AI Legal Research Tools

A new study by researchers at Stanford University and Yale University provides the first comprehensive evaluation of the reliability of leading AI-powered legal research tools. The study focused on assessing the tendency of these tools to "hallucinate", or generate false or misleading information when responding to legal research queries.

Key Findings:

• The AI research tools made by LexisNexis (Lexis+ AI), Thomson Reuters (Westlaw AI-Assisted Research), and Practical Law (Ask Practical Law AI) hallucinated between 17% and 33% of the time when responding to a variety of legal research questions. This is an improvement over general-purpose AI chatbots like GPT-4, which have been found to provide incorrect responses 58% and 82% of the time on legal queries.

• There were substantial differences between the tools. Lexis+ AI was the most accurate, Westlaw AI-Assisted Research and Ask Practical Law AI lagged behind.

• Hallucinations persisted across different categories of questions, including general legal research, jurisdiction-specific queries, factual recall, and even questions with a false premise that the tools should have flagged as incorrect.

Expanded Definition of "Hallucination": Importantly, the study proposes an expanded definition of what constitutes a "hallucination" for AI legal research tools. A response is considered hallucinated if it is either:

1. Incorrect - containing any false statements of law or fact, or

2. Misgrounded - citing sources that do not actually support the propositions claimed.

Responses can be factually correct but still "hallucinate" if they cite inapplicable or irrelevant authorities. The study authors argue this expanded definition is crucial for the legal domain, where the accuracy and applicability of cited authorities are essential.

Implications for Lawyers: The study provides critical information for lawyers seeking to responsibly adopt and supervise the use of AI research tools.

Key takeaways include:

• Leading AI research tools can provide efficiency gains, but they are not fully reliable and lawyers must still carefully verify the accuracy of AI-generated work products, especially cited authorities.

• Differing performance between tools means lawyers must understand the specific strengths and limitations of the AI system they use. Reliance on vendor assurances about accuracy is not sufficient.

• Lawyers must ensure they meet their ethical duties of competence and supervision when using AI research tools, which requires being able to spot and remedy AI errors.

The study emphasizes the need for continued testing and transparent evaluation of AI systems by third parties to empower lawyers to effectively adopt this powerful new technology while upholding their professional responsibilities. Leading AI companies are urged to enable this by facilitating greater testing of their systems.

Ok, this all sounds terrific and I applaud the researchers for undertaking this effort. It is sorely needed if for no other reason than to inform customers that some claims of accuracy are, well, inaccurate.

But here is where I disagree with this research. The study frames hallucinations primarily as a bug or flaw to be mitigated, there's a compelling case to be made that they can also be a feature that enables creative insights and novel lines of inquiry (as I have written and spoken about many times).

In many ways, what the study labels as "hallucinations" could be reframed as a form of creative reasoning or exploratory ideation. When an AI tool generates a response that goes beyond a strict recitation of existing legal authorities, it is engaging in a kind of speculative or generative process that has the potential to surface new connections, arguments, or theories. This is not unlike the kind of creative legal thinking that human lawyers engage in when they are brainstorming novel legal strategies or exploring uncharted areas of law.

From this perspective, the value of AI legal research tools may lie not just in their ability to accurately summarize existing law (btw don’t we already have tools that do this), but also in their capacity to help lawyers generate new ideas and approaches. A certain degree of "hallucination" or creative extrapolation may be desirable, as long as lawyers understand that the output is provisional and requires further investigation and validation.

Moreover, it's worth questioning the implicit assumption that AI legal research tools should aim for 100% accuracy. We don't hold human legal researchers to this standard. Junior associates, and even experienced lawyers, routinely make mistakes or mischaracterize the law in their research memos and briefs. The process of legal research and writing is often iterative, with initial drafts containing errors or overstatements that are caught and corrected through feedback and editing.

Rather than expecting AI tools to be infallible oracles, it may be more productive to view them as research aids that can help generate leads and ideas, but that still require human oversight and judgment. Lawyers could embrace the creative potential of AI "hallucinations" while also developing the skills to critically evaluate and fact-check AI-generated output.

Spotlight

🛠️ 📈 OpenAI Confirms GPT-5 is Coming: What to Expect

OpenAI has recently confirmed that it is actively training its next flagship model, GPT-5, as announced in a blog post revealing the formation of a new Safety and Security Committee. This announcement has sparked anticipation about the likely improvements and new features of GPT-5.

Here is my take:

Enhanced Accuracy (but still not 100%)

A key improvement expected from GPT-5 is enhanced accuracy. Traditionally, larger datasets and more parameters have driven better performance in AI models. For instance, GPT-3.5 was trained on 175 billion parameters, and GPT-4 on a trillion parameters. However, recent advancements emphasize not just the scale but also the efficiency of training. Optimization techniques like knowledge distillation, which trains smaller models to mimic larger ones, and a focus on high-quality, diverse training data are expected to play significant roles in GPT-5's development. This approach can lead to more precise and reliable outputs without solely relying on increasing data volume.

Increased Multimodality

GPT-5 is anticipated to further enhance multimodal capabilities. Previous models progressed from text-only (GPT-3.5) to text and image inputs (GPT-4) and eventually to handling text, audio, image, and video inputs (GPT-4o). GPT-5 may take this a step further by incorporating text-to-video functionalities, leveraging OpenAI’s earlier text-to-video model, Sora. This would enable users to receive video outputs directly from textual inputs, significantly expanding the model’s utility.

Autonomous Action: A Preview of AGI

While GPT-5 is unlikely to achieve full artificial general intelligence (AGI), it is expected to move closer to this goal by employing autonomous agents. These agents would allow GPT-5 to understand, plan, and execute tasks with minimal user intervention. For instance, a user could request GPT-5 to "order a burger from McDonald's," and the model would autonomously navigate the ordering process via various apps. This level of autonomy would greatly enhance the practical utility of AI, making everyday tasks more seamless and efficient for users.

When?

Nobody knows the exact release date for GPT-5, and OpenAI has not provided a timeline. However, given the recent rollouts of updates like GPT-4o and the extended access for free ChatGPT users, many speculate that a significant update could be imminent. This is driven by the need to justify the $20/month subscription for paid users, who currently receive only marginal benefits over the free version.

Additionally, OpenAI has been actively releasing specific ChatGPT versions for non-profits and higher education, indicating a broader strategy to expand its user base and applications. This increased activity suggests that OpenAI is preparing for a substantial upgrade, potentially setting the stage for the introduction of GPT-5.

AI Model Notables

► All ChatGPT Free users can now use browse, vision, data analysis, file uploads, and GPTs

► OpenAI announced two new versions of ChatGPT: one aimed at nonprofits, and the other at universities - (a higher ed colleague of mine got a quote on the ChatGPT Edu of $12 per user with a minimum of 350 users

► OpenAI terminated accounts linked to covert influence operations from Russia, China, and Israel

► OpenAI is recruiting research engineers to rebuild its robotics team, aims to eventually power robots

► Why PwC is becoming ChatGPT Enterprise’s biggest user and becoming its first enterprise reseller

► Apple plans AI-based Siri overhaul to control individual app functions

► Alexa’s Fire TV search has a new AI, but it needs some work

News You Can Use:

➭ An Indianapolis teacher with two decades of experience had to change jobs after an AI-generated deepfake nude picture of her was circulated around her school.

➭ Generative AI seed funding drops 76% as investors take wait-and-see approach.

➭ With the EU AI Act incoming this summer, the bloc lays out its plan for AI governance via its new AI Office.

➭ At 2.9 watt-hours per ChatGPT request, AI queries are estimated to require 10x the electricity of traditional Google queries.

➭ Sony will use AI to cut film costs.

➭ Who owns your voice? Scarlett Johansson OpenAI complaint raises questions.

Who is the author, Josh Kubicki?

Some of you know me. Others do not. Here is a short intro. I am a lawyer, entrepreneur, and teacher. I have transformed legal practices and built multi-million dollar businesses. Not a theorist, I am an applied researcher and former Chief Strategy Officer, recognized by Fast Company and Bloomberg Law for my unique work. Through this newsletter, I offer you pragmatic insights into leveraging AI to inform and improve your daily life in legal services.