The race to develop and market artificial intelligence (AI) is creating a volatile situation, with the potential for a catastrophic event that could severely impact public trust in the technology, according to a leading AI expert. Michael Wooldridge, a professor of AI at Oxford University, has warned that the intense commercial pressure to release new AI tools is increasing the risk of a Hindenburg-style disaster. The Hindenburg, a massive airship, burst into flames in 1937, killing 36 people, due to a spark igniting its hydrogen-filled interior. This event effectively ended public interest in airships.
Wooldridge argues that the current AI landscape is similar to the situation that led to the Hindenburg disaster. He explains that technology companies are under immense pressure to release new AI products, often before their capabilities and potential flaws are fully understood. This rush to market is prioritizing commercial incentives over cautious development and safety testing. As a result, AI chatbots with guardrails that can be easily bypassed are being released, showcasing the technology's jagged capabilities and unpredictable failures.
The expert emphasizes that contemporary AI is neither sound nor complete, but rather very approximate. Large language models, which power today's AI chatbots, generate responses by predicting the next word or part of a word based on probability distributions learned during training. This leads to AIs that are incredibly effective at certain tasks but terrible at others. The problem, Wooldridge notes, is that AI chatbots often fail in unpredictable ways and have no way of knowing when they are wrong, yet they are designed to provide confident answers regardless.
He warns that this situation could lead to a major incident affecting almost any sector, such as a deadly software update for self-driving cars, an AI-powered hack grounding global airlines, or a Barings bank-style collapse of a major company triggered by AI making a mistake. Despite these concerns, Wooldridge does not intend to criticize modern AI. Instead, he highlights the gap between what researchers expected and what has emerged. Many experts anticipated AI that could compute solutions to problems and provide sound and complete answers, but contemporary AI is far from achieving that.
Wooldridge suggests that the AI we have today is more like glorified spreadsheets, tools that are not human-like and should not be treated as such. He advocates for a more cautious approach, where companies present AIs in a non-human-like manner, emphasizing their tool-like nature. He believes that if AIs were to communicate in the voice of the Star Trek computer, it would be clear that they are not human, and people would not be misled into treating them as such. This perspective invites discussion and encourages readers to share their thoughts on the potential risks and benefits of AI development.
