What a whirlwind week in the AI and legal world! From the Sam Altman and OpenAI saga (is he in, is he out? Spoiler: he's back in!) to groundbreaking developments in EU AI Act and copyright cases, it's been a rollercoaster. This week's newsletter is packed with the latest, juiciest updates and some exclusive snippets from the IAPP's AI Governance Course. Plus, a lawyer meme to lighten the mood!

On the docket today:

• AI Mistake Costs Attorney His Job

• EU AI Act: Germany, France and Italy Reach Agreement (Voluntary COC Wins)

• Sarah Silverman’s Copyright Case Isn’t Going Very Well

• BONUS: Juicy Snippets from IAPP's AI Governance Course

• Hilarious Lawyer Meme

  New here? → Subscribe

AI Mistake Costs Lawyer His Job 

Zachariah Crabill, a 29-year-old lawyer, faced a significant career challenge at Baker Law Group due to his reliance on AI in legal work. Here's a quick rundown of his experience:

• Turning to AI Under Pressure: Faced with intense workplace stress and deadlines, Crabill used ChatGPT to assist in drafting a motion, incorporating Colorado case law.

• Unverified AI Outputs: The motion, created with the help of ChatGPT, unfortunately included several fictitious lawsuit citations, which Crabill overlooked before filing.

• The Aftermath of an AI Error: After Crabill realized and reported the mistake to the judge, he was reported to a statewide office, leading to his termination. He speculates, though, that the use of ChatGPT might not have been the only factor in his firing.

Despite this setback, Crabill continues to believe in the potential of AI to boost legal productivity and has since established his own company offering AI-based legal services.

That’s one way to turn a negative into a positive.

Let's take this as a wake-up call to always double-check our work, especially when using AI tools like ChatGPT. While they can be incredibly helpful, they're not foolproof. It's a reminder for all of us: Verify, verify, verify! Stay sharp and diligent to prevent these types of errors from repeating. Keep a close eye on your work, everyone! 👍📝

EU AI Act: Germany, France and Italy Reach Agreement (Voluntary COC Wins)

In the evolving landscape of AI regulation within the European Union, the recent developments and perspectives combine to form a nuanced picture:

• On 18 November, France, Germany, and Italy (FGI) aligned on a new AI regulatory approach, advocating for a voluntary yet binding code of conduct for EU AI providers. This proposal aims to address disagreements that threatened to derail EU legislation.

  • The focus is on 'foundation models' of AI, with a requirement for developers to create 'model cards' detailing their models' capabilities and limitations.

  • Initially, the code of conduct will be without sanctions, but there is potential for penalties in the future based on compliance monitoring.

• Contrasting with this approach, recent events like the ousting of OpenAI CEO Sam Altman have sparked skepticism about the efficacy of self-regulation in the AI industry:

  • European Parliament lawmaker Brando Benifei and Dutch Minister for Digitalisation Alexandra van Huffelen stress the need for "sound, transparent, and enforceable" regulations, indicating doubts about relying solely on voluntary agreements within AI companies.

  • AI expert Gary Marcus warns against trusting companies for self-regulation, highlighting internal conflicts of interest and advocating for a robust EU AI Act.

My thoughts: Call me a skeptic but I don’t trust companies to self-regulate effectively. It may be a good move in the short-term, especially to keep the EU AI Act negotiations moving forward, but it’s definitely not a good long-term solution. There eventually needs to be a transition to hard regulation. 

What are your thoughts on this?

Sarah Silverman’s Copyright Case Isn’t Going Very Well

Sarah Silverman's high-profile copyright lawsuit against OpenAI and Meta is undergoing significant developments, attracting attention in the legal and tech communities. Here are some updates:

• Silverman's Copyright Allegations: Sarah Silverman and other authors claim their copyrighted works were used in training AI models by OpenAI and Meta without permission. The case is currently under the scrutiny of U.S. District Judge Vince Chhabria, known for his expertise in copyright law.

• Reflection of Orrick’s Ruling: The case parallels a recent decision by Judge William Orrick, who significantly reduced a similar lawsuit against generative AI companies. Orrick's ruling dismissed most of the claims, leaving only a few for further consideration.

• Chhabria's Critical View: Judge Chhabria has shown skepticism, especially regarding the claims that Llama, Meta’s AI tool, infringes copyrights through its text generation. He questions the legal basis for considering AI outputs or the AI model as copyright infringements.

• Uncertain Future of Core Claims: Some of the lawsuit's claims are likely to be dismissed by Chhabria, but the central accusation against Meta for using the authors' books in AI training remains unaddressed.

• Potential Legal Precedents: The proceedings in Silverman's case, alongside Orrick’s recent decision, signal a trend in judicial examination of copyright law in relation to AI-generated content. These cases could establish significant precedents in determining copyright infringement within the AI industry.

It’ll be interesting to see how this case continues to develop.

Juicy Snippets from IAPP’s AI Governance Course

Ready for juicy AI governance insights without the hefty fee? I'm dishing out key takeaways from the $1,000 IAPP AI Governance course at no cost, all value! 🚀📚

For those that have been following over the last couple weeks, we’re still in Module 1. These notes include info on Types of AI and Machine Learning.

Module 1: Types of AI

Overview of AI Types

Artificial Intelligence (AI) is commonly segmented into three principal classifications:

• Artificial Narrow Intelligence (ANI) - Weak AI:

  • Definition: AI that is designed to perform a single task or a limited range of tasks with high proficiency.

  • Characteristics: Operates under specific constraints and limitations, hence the term 'narrow'.

  • Application Example: A chess-playing system.

  • Impact: Enhances productivity by automating tasks and enables smarter decision-making through analysis.

• Artificial General Intelligence (AGI) - Strong AI:

  • Definition: AI that aims to replicate human intelligence across a broad spectrum of tasks.

  • Current Status: Theoretical, as AGI has not been achieved yet.

  • Expected Capabilities: Strong generalization, complex task performance, learning, and goal achievement similar to humans.

• Artificial Super Intelligence (ASI):

  • Definition: AI that surpasses human intelligence across all aspects.

  • Current Status: Hypothetical and not yet realized.

  • Projected Attributes: Self-awareness, emotional understanding, and an ability to experience reality on a human level.

• Broad Artificial Intelligence:

  • Positioning: More advanced than ANI but not as developed as AGI.

  • Functionality: Involves a suite of AI systems working in concert, capable of a wider array of tasks.

  • Real-World Example: Autonomous driving vehicles represent current broad AI technologies.

Key Takeaways

• ANI is specialized and highly proficient within its designated tasks but lacks the adaptability to go beyond them.

• AGI represents a future where machines can perform any intellectual task that a human being can, but it remains aspirational.

• ASI extends the concept of AGI to levels of capability that exceed human abilities in every domain, which remains theoretical.

• Broad AI fills the gap between ANI and AGI, handling more complex and varied tasks by leveraging multiple AI systems.

Module 1: Machine Learning 

Machine Learning Fundamentals

• Definition: Machine learning (ML) is a subset of AI focusing on data and algorithms to imitate the way humans learn, incrementally improving accuracy.

• Purpose: ML enables systems to make decisions and predictions, enhancing their performance over time autonomously.

• Training Models: ML uses various models like supervised, unsupervised, reinforcement, and semi-supervised learning.

Supervised Learning

• Process: Learns from labeled data to predict outcomes or categorize inputs.

• Function: Generates functions from input-output pairs to predict responses for new data.

• Subcategories:

  • Classification: Categorizes input into labels (e.g., identifying an animal in an image).

  • Regression: Predicts continuous outputs (e.g., forecasting stock prices).

• Examples:

  • Support Vector Machine (SVM): Primarily used for classification tasks.

  • Support Vector Regression (SVR): Used for predicting continuous values.

Unsupervised Learning

• Process: Identifies patterns in data without pre-existing labels.

• Subcategories:

  • Clustering: Groups data with similar attributes (e.g., DNA pattern analysis).

  • Association: Discovers rules that highlight relationships between variables (e.g., market basket analysis).

• Applications: Anomaly detection, market segmentation, and genetics.

Reinforcement Learning

• Concept: Learns through trial and error using rewards and penalties.

• Mechanism: Adapts behavior to maximize rewards based on feedback from the environment.

• Applications: Robotics for navigation and task efficiency, predictive text generation, and real-time ad placement optimization.

Semi-Supervised Learning

• Hybrid Approach: Combines labeled and unlabeled data, leveraging the strengths of both supervised and unsupervised learning.

• Usage: Effective in situations with limited labeled data available.

• Examples:

  • Image and Speech Analysis: Categorizes and interprets visual and audio data.

  • Web Page Ranking: Orders search results based on relevance.

  • LLMs and Generative AI: Tools like ChatGPT and Dall-e, which generate text and images, respectively.

Key Takeaways

• Machine Learning's Essence: ML is central to AI's ability to learn and adapt without explicit programming.

• Supervised Learning's Role: Vital for tasks requiring precise predictions based on past labeled data.

• Unsupervised Learning's Autonomy: Finds structure in data on its own, which can be less precise but more exploratory.

• Reinforcement Learning's Adaptability: ML with a focus on learning from the consequences of actions, akin to a reward system.

• Semi-Supervised Learning's Balance: Offers a middle ground, utilizing both labeled and unlabeled data, often seen in advanced AI applications.

Meme Of The Week:

That’s all for today!

Catch ya’ll on LinkedIn