Stanford Report Shows Bigger AI Models Cause More Ethical Harm
The 2022 AI Index reveals a troubling trend where more capable AI systems produce greater ethical harms, prompting a surge in industry and academic research focused on measuring these negative impacts.
The 2022 AI Index Report from the Stanford Institute for Human-Centered Artificial Intelligence highlights a critical challenge in the tech world: as AI systems become more capable, they also generate more harm. Researchers historically focus on publishing technical achievements, but they are much slower to measure and report the negative consequences of these advanced models.
Co-author Helen Ngo points out that as companies deploy AI in the real world, they must understand how these systems perpetuate harm and start measuring shortcomings alongside capabilities. A counterintuitive finding shows that larger, more powerful AI models are actually more likely to produce outputs that misalign with human values.
Fortunately, both industry and academic interest in AI ethics is growing rapidly. The report examines specific issues like toxic language model outputs, difficulties in ensuring AI truthfulness, and gender bias in machine translation, highlighting that while progress exists, significant work remains to properly quantify AI systems along ethical dimensions.