😵Shocking Effects of AI ‘Unlearning’

Making AI models 'forget' specific data can make them less useful, study finds

Key Points 

• AI models like GPT-4o are trained on scraped data from public websites without compensating original data owners, leading to legal disputes.

• Unlearning techniques, which attempt to remove specific information, are now a key focus to address these copyright issues. However, they disrupt the patterns learned by models, which are essential for generating accurate responses. 

• As a result, the models' overall predictive capabilities are affected, causing them to lose their ability to answer basic questions accurately and degrading their general performance.

🤖News - According to a new study by researchers from the University of Washington, Princeton, the University of Chicago, USC, and Google, techniques to make AI "unlearn" specific information—like sensitive data or copyrighted material—have some serious downsides. 

These methods can make advanced models like OpenAI’s GPT-4o or Meta’s Llama 3.1 405B much less effective at answering basic questions, sometimes to the point where they're not practical to use anymore.

🤔Why is that so? Generative AI models aren't truly intelligent; they function as statistical systems that predict data like words, images, and speech based on patterns learned from extensive examples such as movies and essays. For instance, given an email fragment like "Looking forward...", a model might complete it with "... to hearing back," based purely on learned patterns. There's no intentional thought behind it.

Because of this, unlearning techniques, which attempt to remove specific information, disrupt these patterns. Instead of simply erasing the data, unlearning uses algorithms to steer models away from certain information. As a result, the models' overall predictive capabilities are affected, causing them to lose their ability to answer basic questions accurately and degrading their general performance.

👨🏻‍💻Here's why this matters - It's no secret that many AI models, including prominent ones like GPT-4o, are trained using data scraped from public websites without notifying or paying the original data owners, often citing fair use. This has led to lawsuits from copyright holders such as authors and publishers. As a result, unlearning techniques have become a hot topic, with Google and several academic institutions launching a competition to develop new ways to address these copyright issues by effectively removing specific data from AI models.

🕵🏻‍♂️Has anyone found a solution yet? Not really—which is why the authors of the study say there’s still a lot of research needed. Maybe a technical breakthrough will make unlearning viable in the future, but for now, vendors will need to find other ways to stop their models from saying things they shouldn’t.

Report: AI adds to employee workload and stress

Key Points 

• 77% of employees reported that AI tools increased their workload, leading to burnout for 71% and difficulty meeting productivity demands for 65%.

• About one-third of surveyed individuals are considering quitting within the next six months due to feeling overworked.

☕News - A survey by the Upwork Research Institute and Workplace Intelligence found that while companies that adopted AI expected it to boost productivity and make jobs easier, the reality is quite different.

🧐How so? Out of 2,500 employees, freelancers, and C-suite executives surveyed from various countries:

• 77% said AI tools actually increased their workload.

• 39% said they spent more time reviewing or moderating AI-generated content.

• 23% had to learn how to use these new tools.

• 71% felt burned out.

• 65% struggled to meet their employers' productivity demands.

• And, about a third are thinking of quitting in the next six months because they feel overworked.

👥What insiders think - Kelly Monahan from the Upwork Research Institute highlighted that for AI to truly enhance productivity and employee well-being, a fundamental shift in how we organize work is needed.

Lucas Botsen, the CEO of an HR services company, added that while AI can be efficient in theory, it still needs constant monitoring and validation to meet specific operational needs.