Editorial · AI Safety
What Nobody Is Saying About India's Use of Home Videos to Train AI Robots
The use of home videos to train AI robots in India has sparked a growing concern among workers who feel their jobs are being threatened by the very technology they are helping to create. Every morning, hundreds of workers in textile factories strap on small recording devices to track and record their every move, from adjusting machine levers to calibrating moving parts. This exercise, known as egocentric data collection, is meant to teach machines how people perform physical tasks, but it has also exposed a sharp imbalance of power between workers and factory management.
Workers are often not told what is being recorded, where the footage is going, or how it may eventually be used. They are simply asked to wear the devices throughout their shift, with no control over how their data may be used to automate parts of their job or replace them altogether. This lack of transparency has led to feelings of unease and mistrust among workers, who are already struggling to make ends meet in a highly competitive industry. The fact that they are being asked to generate behavioural data, including years of tacit skill and muscle memory, without any say in how it will be used, is a clear indication of the power dynamics at play.
The demand for egocentric data is on the rise, with robotics labs needing between 100 million to 1 billion hours of pre-training data over the next two to three years. This has created a lucrative market for companies that collect and sell such data, often without considering the impact on the workers who are generating it. The end goal of collecting such data is to build robots that can operate in the real world with human-like adaptability and precision, but it is unclear whether the benefits of this technology will trickle down to the workers who are making it possible.
The use of home videos to train AI robots also raises important questions about the future of work and the impact of automation on employment. As machines become more advanced and capable of performing complex tasks, there is a real risk that workers will be replaced by robots, leading to widespread unemployment and social unrest. This is particularly concerning in countries like India, where the job market is already highly competitive and workers are struggling to make ends meet. The fact that workers are being asked to generate data that may ultimately be used to replace them is a stark reminder of the need for greater transparency and accountability in the development and deployment of AI technology.
As we look to the future, it is clear that the use of home videos to train AI robots is a trend that is here to stay. However, it is imperative that we consider the impact of this technology on workers and take steps to ensure that they are protected and empowered. This may involve providing workers with greater control over their data, ensuring that they are fairly compensated for their contributions, and implementing measures to mitigate the negative effects of automation on employment. Only by taking a more nuanced and equitable approach to the development and deployment of AI technology can we ensure that its benefits are shared by all, rather than just a privileged few.
Editorial perspective - synthesised analysis, not factual reporting.
Terms in this editorial
- egocentric data collection
- A method where workers wear recording devices to capture their actions, used to train AI robots. This technique aims to teach machines how people perform physical tasks but raises concerns about worker privacy and job security.
If you liked this
More editorials.
AI Alignment and Ethical Computing: A Call for Interdisciplinary Dialogue
Recent advancements in artificial intelligence (AI) have sparked significant discussions about how these technologies can be aligned with human values while maintaining ethical standards. The MIT Ethics of Computing Research Symposium, held on June 5, 2026, brought together experts from various fields to explore the social and ethical implications of AI. This event emphasized the importance of collaboration between computer scientists, philosophers, policymakers, and educators to ensure that AI technologies are developed and deployed responsibly. The symposium featured a keynote address by Jon Kleinberg, the Tisch University Professor of Computer Science and Information Science at Cornell University. Kleinberg discussed the challenges of algorithm-human handoffs, highlighting the need for clear communication and mutual understanding between humans and AI systems. He stressed that while AI can enhance decision-making processes, it must always remain under human oversight to prevent unintended consequences. One of the key panels focused on AI alignment, a topic that raises fundamental questions about governance and accountability. Dylan Hadfield-Menell, an associate professor of Electrical Engineering and Computer Science (EECS) at MIT, moderated a discussion where experts debated how to instill "human values" into AI systems. Iason Gabriel, a philosopher and research scientist at Google DeepMind, argued that AI should be designed to interpret human moral values rather than mimic them perfectly. He likened this approach to a judge's role, where the system adheres to rules while also considering context and fairness. Bailey Flanigan, an assistant professor of political science with a joint appointment in the MIT Schwarzman College of Computing, emphasized the need for broader societal input in determining who gets to govern AI systems. She highlighted the importance of establishing clear guidelines and regulatory frameworks to ensure that AI technologies benefit all segments of society equitably. The symposium also showcased student research during an afternoon poster session. These projects demonstrated the diverse ways in which ethical considerations are being integrated into AI development. For instance, one team presented a framework for responsible computer vision deployment, focusing on issues like bias mitigation and transparency. Another project explored the use of AI in air pollution forecasting, aiming to create more accurate and accessible tools for public health officials. Looking ahead, the symposium serves as a reminder that ethical computing is not just an afterthought but a core component of technological progress. As AI becomes more pervasive, it will be essential to foster interdisciplinary dialogue and collaboration. This approach ensures that technical advancements are paired with thoughtful reflection on their societal impact. The MIT Schwarzman College of Computing's Social and Ethical Responsibilities of Computing (SERC) initiative is leading the charge in this effort. By supporting cutting-edge research and creating platforms for open discussion, SERC is helping to shape a future where AI technologies are developed with accountability and empathy. As computing and AI continue to evolve, initiatives like these will play a crucial role in guiding the field toward ethical and inclusive innovation. In conclusion, the symposium underscored the importance of viewing AI not just as a tool for efficiency but as a technology that requires careful stewardship. By embracing an interdisciplinary approach, we can ensure that AI systems are aligned with human values and contribute to a more equitable and just society. The challenges ahead are complex, but through collaboration and thoughtful dialogue, they are surmountable.
The Future of Bilingual Voice Agents: A Call for Ethical and Inclusive Design
The rise of bilingual voice agents presents a critical opportunity to bridge linguistic gaps in customer service and enterprise interactions. However, as we saw in recent benchmarks, the technology struggles to handle code-switched speech accurately. This gap is not just technical-it raises ethical concerns about inclusivity and accessibility for multilingual users. If voice agents are to serve diverse populations effectively, developers must prioritize designs that account for the complexity of human language use, particularly in bilingual contexts. Current models demonstrate significant word error rates when faced with code-switched speech, highlighting a fundamental flaw in their training data and architectures. The benchmark results reveal that even state-of-the-art ASR systems struggle to maintain accuracy across different language pairs. This failure extends beyond mere transcription errors-it undermines the ability of voice agents to understand and respond appropriately to user needs. To address this challenge, developers must adopt ethical design principles. First, they should diversify training data to include more code-switched examples from real-world interactions. Second, models must be fine-tuned to recognize and preserve linguistic nuances that are critical for maintaining the integrity of user intent. Finally, companies must establish clear guidelines for error handling, ensuring that users receive accurate and respectful responses even when systems falter. Looking ahead, the integration of multilingual capabilities into voice agents will require collaboration between linguists, data scientists, and ethicists. By prioritizing inclusive design, the AI community can create tools that not only improve accuracy but also reflect the diversity of human communication. The stakes are high-failure to address these issues risks excluding millions of bilingual speakers from the benefits of modern technology. In conclusion, the future of voice agents lies in their ability to serve all users with equal precision and respect. Ethical design must be at the forefront of this evolution, ensuring that no one is left behind in the AI-driven world of tomorrow.
The Unseen Risks of Anthropic's Claude Mythos AI Model
The recent development of Anthropic's Claude Mythos AI model has sparked significant concern among cybersecurity experts and financial institutions. This advanced AI model not only detects security vulnerabilities but also exploits them with alarming efficiency, potentially destabilizing critical systems and eroding public trust in financial institutions. Anthropic claims that Mythos is a general-purpose AI capable of identifying and exploiting software flaws faster than human capabilities. While this may seem like a positive advancement for cybersecurity, the reality is far more dangerous. The model's ability to uncover vulnerabilities in major operating systems and web browsers highlights a pressing issue: it illuminates risks that were already present but undetected by existing security scanners. This capability could enable malicious actors to exploit these vulnerabilities before they can be patched, leading to significant disruptions in financial systems. In response to the potential threat posed by Mythos, Anthropic has launched Project Glasswing, a coalition of major tech companies and financial institutions. The initiative aims to use Mythos in preview mode to identify and fix vulnerabilities before hackers can exploit them. However, this approach raises ethical concerns about who should control such powerful AI tools and whether their release could be misused for malicious purposes. The broader implications of Anthropic's Claude Mythos are profound. If widely released, similar models could significantly alter the cybersecurity landscape, making it harder for organizations to protect against increasingly sophisticated threats. The financial sector, in particular, is vulnerable to such risks, as undetected vulnerabilities could lead to operational outages and reputational damage. Looking ahead, the development of AI models like Mythos underscores the urgent need for stricter regulations and independent verification processes. While Anthropic's intentions may be noble, the potential consequences of unrestricted access to such powerful tools are too great to ignore. The financial industry must remain vigilant and prioritize proactive measures to mitigate these risks before they escalate into broader systemic threats. In conclusion, Anthropic's Claude Mythos AI model represents a double-edged sword for cybersecurity. While its capabilities could theoretically enhance security by identifying vulnerabilities, the potential for misuse is equally significant. As the technology continues to evolve, it is crucial for stakeholders to approach its deployment with caution and prioritize ethical considerations over technological advancement.
The Flawed Science Behind AI-Driven Policing: A Call for Accountability
In recent years, law enforcement agencies across the United States have increasingly turned to artificial intelligence (AI) tools to aid in criminal investigations. These technologies are supposed to enhance accuracy and efficiency, but a growing body of evidence reveals that AI-driven systems, particularly facial recognition software, are prone to significant errors that disproportionately affect marginalized communities. Consider the case of Porcha Woodruff, a Black woman who was eight months pregnant when six cops arrived at her home in Detroit. Using an AI-driven software program, police identified her as a suspect in a carjacking incident based on an outdated mugshot from eight years prior. Despite having access to a more recent photo from her driver's license, authorities failed to verify the match, leading to her wrongful arrest and 11-hour detention. This scenario is far from isolated; similar incidents have occurred across the country, with innocent individuals being held in custody due to flawed AI technology. The problem extends beyond individual cases. Studies show that facial recognition systems exhibit racial biases, often misidentifying people of color at higher rates than white individuals. A Georgia State University study found significant disparities when it comes to arrests based on facial recognition software, raising serious concerns about its reliability and fairness. These issues are compounded by the fact that law enforcement agencies often lack transparency in how they use these technologies, leaving citizens without recourse or understanding. The implications of these errors are profound. Beyond the immediate trauma of wrongful arrest, individuals like Woodruff face long-term consequences on their lives and livelihoods. In her case, the stress of the arrest worsened her pregnancy complications, and she lost custody of two of her children. Such outcomes highlight the urgent need for stricter oversight and accountability in the use of AI-driven surveillance tools. To address these concerns, policymakers must take action. This includes implementing independent audits to assess the accuracy and fairness of facial recognition systems, as well as mandating double-checks by human officers before any arrests are made based on AI findings. Additionally, law enforcement agencies should be required to disclose how they use these technologies and provide clear pathways for individuals to challenge misuse. The future of AI in policing is undeniably promising, but only if we can ensure its accuracy, fairness, and ethical deployment. Until then, the risks of wrongful arrests and deeper inequities in our justice system remain too great to ignore. It's time to demand accountability-not just for these flawed systems-but for the agencies that choose to rely on them without proper safeguards.
The Hidden Cost of AI Security: Why Your Model Might Be More Vulnerable Than You Think
The AI revolution has brought unprecedented capabilities to businesses and individuals alike. From automating routine tasks to solving complex problems, AI systems have become integral to modern life. However, as these systems grow more powerful, so do the risks they pose. Recent findings from Cisco's research on AI security reveal a concerning truth: even the most advanced models are far more vulnerable than their creators claim. In a groundbreaking study, Cisco tested 15 leading AI models across various scenarios. The results were startling. While single-turn attack success rates ranged from 2.7% to 64.9%, multi-turn attacks saw an alarming increase, with rates as high as 88.3%. This disparity underscores a critical flaw in how we assess AI security. Models that pass initial tests with flying colors often crumble under sustained pressure. For instance, OpenAI's GPT-5.4 showed a nine-fold increase in vulnerability when attacked over multiple turns. Similarly, Google's Gemini 3 Pro saw its success rate jump from 18.1% to 73.4%. These numbers suggest that the current metrics used to evaluate AI models are deeply flawed. The issue lies not just in the models themselves but in how we test them. Traditional benchmarks focus on isolated interactions, ignoring the reality of persistent attacks. Cisco's research highlights this gap, showing that many models perform significantly worse when faced with multi-turn attacks. This shift in perspective is crucial for understanding the true state of AI security. Moreover, the report exposes another layer of complexity: deployment-time configurations. When Google's Grok 4.1 Fast was tested with reasoning mode enabled, its success rate dropped from 88.3% to 43.5%. This revelation points to a broader problem in how models are configured and managed once deployed. The current lack of transparency around these settings leaves organizations vulnerable to exploitation. The implications of this research are profound. Businesses relying on AI must reevaluate their security strategies. Single-turn metrics, while useful for initial assessments, provide an incomplete picture of a model's resilience. Multi-turn attacks represent a real-world scenario that cannot be ignored. The same models trusted with sensitive tasks could be manipulated to produce harmful outputs through persistent interaction. Looking ahead, the future of AI security requires a fundamental shift in how we approach testing and deployment. Organizations must demand more comprehensive reporting from vendors, including attack success rates across different strategies. Additionally, they should implement stricter thresholds for model deployment, ensuring that only those with proven resilience are put into production. The road to secure AI is fraught with challenges, but the rewards are immense. By acknowledging these vulnerabilities and taking proactive steps, we can build a future where AI enhances security rather than undermines it. The clock is ticking, and the stakes could not be higher.