Summary
Highlights
AI researcher Sasha Luccioni addresses common fears about AI's future existential threats, arguing that immediately pressing dangers like climate change contributions, data exploitation, and discrimination are often overlooked. She emphasizes that AI operates within society, impacting real people and the planet, and advocates for transparency, disclosure, and tools to better understand AI's consequences.
AI models require significant energy, contributing to climate change. Luccioni's research on the Bloom large language model revealed that its training consumed as much energy as 30 homes in a year, emitting 25 tons of CO2. She notes that larger AI models like GPT-3 have even greater carbon footprints, yet tech companies rarely disclose these figures. She introduces CodeCarbon, a tool to measure AI's energy consumption and carbon emissions, helping developers make sustainable choices like deploying models on renewable energy.
Luccioni highlights how AI models are trained on artists' and authors' work without consent. She mentions 'Have I Been Trained?', a tool created by Spawning.ai that allows individuals to check if their work is in massive AI training datasets. This tool has provided crucial evidence for artists like Karla Ortiz in copyright infringement lawsuits. She also discusses the collaboration between Spawning.ai and Hugging Face to create opt-in/opt-out mechanisms for dataset creation, asserting that human-created artwork should not be an 'all-you-can-eat buffet' for AI training.
AI models often perpetuate stereotypes and biases, leading to real-world harm. Luccioni references Dr. Joy Buolamwini's work, which exposed facial recognition systems' poorer performance on women of color. Such biased systems in law enforcement can lead to wrongful accusations and imprisonment. To combat this, Luccioni created the Stable Bias Explorer, a tool that reveals biases in image generation models by examining how professions are depicted, often showing significant overrepresentation of whiteness and masculinity compared to real-world statistics.
With AI becoming integral to society, Luccioni argues for its accessibility to ensure that everyone understands how it works and when it fails. She emphasizes that there are no simple solutions to complex issues like bias, copyright, or climate change. However, by creating tools to measure AI's impact, we can establish guardrails for society and the planet. This information empowers companies to choose sustainable and ethical models, helps legislators develop effective regulations, and enables users to select trustworthy AI. She concludes by reiterating that focusing on current, tangible impacts is crucial for building a better AI future, as we are collectively shaping its direction.