Workers Sound Alarm on AI Risks, Urge Caution in Adoption

Concerns about artificial intelligence (AI) are mounting among those directly involved in its development and assessment. Employees tasked with training AI systems have voiced serious warnings about the potential dangers associated with these technologies. An article published by The Guardian highlights insights from AI workers who advise caution regarding the use of AI, particularly in sensitive areas like healthcare.

The interviewees, who have firsthand experience in training AI models, shared alarming observations about the biases embedded within these systems. They reported facing challenges such as inadequate training, vague instructions, and unrealistic deadlines, all of which undermine the quality and reliability of AI outputs. Many of these workers are now cautioning friends and family about the technology’s risks, with some even restricting their children’s use of AI.

This skepticism is not entirely new; concerns about misinformation and biases in AI have been raised by various experts. Notably, a campaign group known as Pause AI has compiled an “AI Probability of Doom” list, which assesses the likelihood of severe negative outcomes resulting from AI advancements. This list includes voices from established AI researchers who have extensively documented the potential pitfalls of artificial intelligence.

Even influential figures in the AI industry are advocating for a cautious approach. In a podcast aired in June 2025, Sam Altman, CEO of OpenAI, remarked, “People have a very high degree of trust in ChatGPT, which is interesting because AI hallucinates. It should be the tech that you don’t trust that much.” While Altman does not discourage the use of ChatGPT outright, his comments reflect growing concerns about over-reliance on such technologies.

Many AI workers, including those who participated in The Guardian interview, have experienced the pressure of performing tasks that often prioritize speed over thoroughness. They describe a work environment where they are expected to improve AI models while grappling with unclear guidelines and insufficient training. One worker encapsulated this sentiment, stating, “We’re expected to help make the model better, yet we’re often given vague or incomplete instructions, minimal training, and unrealistic time limits to complete tasks.”

The process of developing AI, particularly in the context of large language models like GPT, involves two main stages: language modeling and fine-tuning. During the initial phase, AI systems are trained on extensive datasets, including text from the internet and books, to identify patterns in language. The fine-tuning phase introduces human testers who evaluate and rank the AI’s responses to enhance its safety and usability.

Although companies like OpenAI employ senior research engineers for specialized tasks, much of the routine evaluation work is outsourced to third-party contractors across the globe. This outsourcing raises questions about the quality and consistency of the evaluation process. Continuous testing remains essential even after an AI model’s release, with “red-teaming” being a critical part of this effort. In this context, workers actively probe AI systems for errors, biases, and unsafe behaviors, contributing to ongoing improvements.

Despite these measures, errors in AI outputs can have serious consequences. Recent investigations, including one by The Guardian, revealed that Google AI’s health-related responses sometimes provided misleading information regarding liver function tests. Such inaccuracies could mislead individuals with serious health conditions into believing they are healthy. Following this revelation, Google updated its AI system and removed the overview concerning liver function tests.

As the conversation around AI continues to evolve, the perspectives of workers in the trenches offer essential insights into the technology’s risks and limitations. Their experiences highlight the need for enhanced training, clear guidelines, and a cautious approach to AI adoption, particularly in critical areas affecting public health and safety.