Artificial intelligence (AI) systems, which are increasingly used in various industries, can unintentionally absorb biases present in the data they are trained on. This can lead to errors and unintended consequences, and responsibility for these issues may not rest solely with the companies that develop the AI. Experts suggest that organizations using AI should thoroughly understand their systems and anticipate potential problems. While large language models (LLMs) are often marketed as efficient tools that can enhance decision-making and reduce costs compared to hiring human workers, they can also carry hidden risks and liabilities. According to Gartner, global spending on AI is expected to exceed $2.5 trillion by the end of 2026. Several AI companies, including OpenAI, Replika, Clearview, and Character.ai, have faced fines for issues arising from biased training data in their models. In one notable case, a Canadian court ruled that Air Canada was fully responsible for damages after its AI chatbot provided incorrect information to a passenger about a refund, despite the chatbot being a third-party product. In January 2025, the National Transportation Safety Board (NTSB) began an investigation into Waymo's robotaxis after they were observed ignoring a school bus stop sign in Austin, Texas, even though the bus driver had clearly signaled that they were stopped. Similar incidents were reported in Atlanta, prompting the National Highway Traffic Safety Administration (NHTSA) to look into Waymo's autonomous driving algorithms. As a result, Waymo recalled over 3,000 vehicles and implemented fixes. However, the NTSB launched another investigation after a new violation occurred following one of the fixes. Missy Cummings, director of the Mason Autonomy and Robotics Center and former safety advisor to the NHTSA, noted that a specific change in Waymo’s engineering model was responsible for the series of traffic violations. Waymo has since transitioned to the Waymo Foundational Model, an end-to-end (E2E) learning model that mimics real driving behaviors based on Gemini models. Unlike previous algorithms, this model does not prioritize traffic rules as strictly. A system trained on real-world driving video data is only as effective as the data it observes, meaning it may learn and replicate reckless behaviors, such as speeding past school bus stop signs. Cummings explained that E2E models analyze short video clips to understand spatial relationships and actions within a scene, but they require vast amounts of data to be reliable. Unfortunately, there is currently limited public research on how much data is needed to ensure safety in these systems. These challenges are not unique to autonomous vehicles. Similar issues have been reported in other sectors, such as medical technology, where biased algorithms can impact patient care. In 2019, a study by Ziad Obermeyer, a professor at the University of California, Berkeley, found that Optum’s Impact Pro risk assessment algorithm systematically classified Black patients as healthier than white patients, even when their medical conditions were similar. The algorithm used health costs as a proxy for health needs, which led to disparities in care access. Obermeyer noted that Black patients had lower annual medical expenses due to unequal access to healthcare, not because of lower health needs. If the algorithm had been accurate, more Black patients would have been included in care programs. New regulations under Section 1557 of the Affordable Care Act, effective in May 2025, aim to clarify responsibility in cases of discrimination. The rule prevents healthcare providers from shifting blame to AI tools used in patient care. Deepika Shrivastava, director of operations at The Doctors Company, emphasized that doctors should treat AI systems like any other clinical tool, carefully documenting their use in case of legal scrutiny. In the realm of human resources, a 2023 class-action lawsuit against Workday alleged that its AI-based candidate screening platform disproportionately rejected applicants based on age, disability, and ethnicity. The court ruled that AI providers like Workday are considered extensions of the employers who use their platforms, meaning that responsibility for AI-driven decisions extends beyond the technology itself. Tabitha Weinstein, former HR executive from the Maryland State Public Safety, advised HR managers to thoroughly understand the AI platforms they use, ensure that providers have conducted bias audits, and maintain human oversight to override AI decisions when necessary.