#26 Society and Technology Bulletin

AI-powered chatbots have rapidly assumed roles such as personal assistants customer service representatives and even psychological counselors within just a few years. This advancement stems from large language models trained on trillions of words of data collected from the internet. Riding the momentum of this success technology leaders like Elon Musk and NVIDIA CEO Jensen Huang argue that humanoid robots trained with similar methods will soon replace surgeons factory workers and domestic helpers. But is this truly possible?

Robotics pioneers are taking a more cautious stance toward this optimism. Professor Ken Goldberg of industrial engineering at UC Berkeley directly addresses these claims in two new papers published in Science Robotics. To counter the prevailing “humanoid hype” Goldberg introduces the concept of a “100 000-year data gap.”

According to Goldberg’s calculations it would take a human 100 000 years to read all the text currently available on the internet. Large language models are trained precisely on this massive volume of data. Yet no comparable volume of data exists for robots to master objects in the physical world. Mastering physical objects requires processing them at a level that is nearly impossible to achieve through data alone. Moreover motor skills are far more complex than language acquisition. Simple tasks such as securely grasping a glass placing a lightbulb in its socket tightening a screw with appropriate pressure or stirring soup without spilling it present formidable challenges.

All these everyday human movements remain unsolved massive problems for robots. This is where Moravec’s Paradox comes into play. According to this paradox what is simple for humans is hardest for robots. A robot can master chess yet safely lifting a glass remains a distant goal.

Although various approaches have been tried to overcome the data deficit for training robots no definitive solution yet exists. One such approach is simulation. Training robots thousands of times in virtual environments to walk jump or tumble has yielded promising results for certain movements. However when it comes to complex manual tasks simulations fail to translate effectively into the real world. Another method is teleoperation. Humans remotely controlling robots to generate data is also a viable solution. Yet this approach progresses extremely slowly. Eight hours of human effort produce only eight hours of robot data. At this rate closing the 100 000-year data gap would require an equally vast amount of time.

Thus Goldberg dismisses the notion that “if large language models succeeded with this approach we can apply the same to robots” as an overly simplistic analogy.

Goldberg highlights another critical shift in the robotics field: a paradigm split. Today the robotics community is divided into two camps. The first prioritizes the “data is sufficient” approach. This view which is especially common among younger researchers argues that engineering calculations and physical modeling are unnecessary and that sufficient data alone will enable robots to learn anything. This camp is strongly supported by investors and the media because it appears more immediately practical. The opposing camp holds that “engineering is indispensable.” Represented by more established researchers this perspective argues that robots cannot become functional without physics mathematics and engineering models. To them data alone lacks meaning without engineering support.

Goldberg aligns with the second view: “Through engineering we can make robots sufficiently functional to bring them to market. As robots are used they gather data and improve over time. This is a realistic roadmap.” Yet this roadmap appears ambitious in its timeline for delivering breakthrough results in the near term.

According to Goldberg blue-collar jobs particularly those requiring craftsmanship and manual dexterity will remain safe for a long time. Humanoid robots replacing construction workers electricians or plumbers is not feasible in the near future. In contrast some routine white-collar tasks are at risk. Standardized procedures such as form filling patient intake processes and customer service interactions previously viewed as rule-based are now increasingly susceptible to replacement or augmentation. Similarly image-intensive fields like radiology also appear ripe for AI assistance.

Yet Goldberg emphasizes that even in these areas a uniquely human element is missing: “A computer can never say to you ‘I understand you.’” Machines incapable of empathy have serious limitations in customer experience and patient care. While a social life adapted to these limitations may be possible it would appear profoundly dull.

Goldberg’s core message is clear. The development of humanoid robots is advancing rapidly but not nearly as quickly as popular narratives suggest. We are unlikely to see surgical robots or household assistants within five years. Rather than fostering pessimism Goldberg stresses that the process may take decades and therefore expectations must remain realistic. Otherwise a bubble could burst triggering widespread disillusionment and backlash against technology.

The artificial intelligence industry is facing the largest copyright lawsuit ever filed. The investigation began with allegations that Anthropic used copyrighted books without permission to train its AI models. Initially launched based on the complaints of only three authors the case has since expanded to potentially involve seven million claimants. With damages of up to $150 000 per work the lawsuit could reach hundreds of billions of dollars in financial exposure. Industry representatives warn this could collapse not just AI startups but the entire sector.

The lawsuit stems from Anthropic’s use of pirated copies of millions of books to train its chatbot Claude. What began with three authors’ complaints soon mobilized hundreds of writers and publishers. In June U.S. Judge William Alsup ruled that using books for AI training was not inherently unlawful. However the problem lay in how the books were obtained — from pirate websites. As the case neared resolution Anthropic agreed to pay a total of $1.5 billion for approximately 500 000 books. This settlement is expected to become the largest in copyright history in the AI era and set a critical precedent.

Industry stakeholders argue that continuing the lawsuit in its current form will scare investors discourage new ventures and cause the U.S. to cede its leadership in AI to Europe and China. Yet using local crises as an excuse to suppress national tensions also carries risks for the technology world. In this case global interests favoring the tech industry may align with its position but in other cases they may oppose it.

Within this tension the concept of “coercive settlement pressure” has emerged. This idea suggests that similar copyright disputes should be resolved quietly without public outcry. If companies consistently settle to avoid massive damages without litigation it could create a lasting “culture of silence.” In such a scenario all barriers appear solvable — if only through financial means.

Criticism of the lawsuit also comes from groups representing authors and libraries. Organizations such as Authors Alliance Electronic Frontier Foundation American Library Association Association of Research Libraries and Public Knowledge argue that the court has ignored longstanding issues raised in the Google Books case. These include unclear ownership orphaned works books by deceased authors and regional licensing differences.

The Google Books case began in 2004 after Google scanned and digitized millions of books. Authors Guild and publishers sued for unauthorized copying. Google defended its actions as fair use arguing it displayed only small text snippets to improve access to information. In 2013 the court ruled in Google’s favor a decision upheld in 2015 and confirmed by the Supreme Court in 2016 after it declined to hear an appeal. This ruling became a landmark precedent in debates over copyright and access to information in the digital age. Yet the Claude case signals a new turning point.

All these developments elevate the Anthropic case beyond a simple copyright dispute. It directly confronts the legal question: Is it lawful to use copyrighted works to train AI models? Yet proceeding with the case in its current form may not resolve fundamental legal issues but instead force companies into settlements creating a lasting climate of uncertainty.

“This case is of extraordinary importance. It concerns the rights of millions of authors and simultaneously the transformative technology used by hundreds of millions of researchers students and writers.” – Joint statement from author and library organizations.

Training AI models is critically important under the current paradigm because these models rely heavily on textual sources and their diversity. In this context nearly everything written becomes potential training data for AI systems. Therefore if the appeals court rejects Anthropic’s appeal the upcoming trial in four months could mark the beginning of the end not just for this company but for the entire AI industry.

Thus the growth of AI technologies and the data-centric dominance driving chatbot proliferation confronts legal frameworks designed to protect individual property rights like copyright. Yet the critical question is: What purpose should the law serve? Absolute protection of copyright risks slowing innovation. Conversely anti-copyright approaches can be justified under the banner of “public interest” particularly when aligned with the interests of large corporations. Therefore presenting anti-copyright stances merely as a legal necessity may be overly optimistic.

When we consider the issue from the perspective of users a further dimension emerges. Today every piece of content shared on social media and digital platforms carries the potential to become training data for AI models. In such a context the most urgent need is greater user awareness: users must pay closer attention to how their data is used the boundaries of digital privacy and the protection of personal information. Ultimately beyond the societal implications of these cases we as individual users must draw lessons about our own personal usage.