AI Is Dangerous, but Not for the Reasons You Think | Sasha Luccioni | TED

Key Concepts

AI impacts, sustainability, carbon emissions, data bias, copyright infringement, transparency, AI ethics, AI accountability, AI tools, model size, energy consumption, AI governance, societal impact.

Sustainability and Environmental Impact of AI

• Energy Consumption: Training large AI models requires significant energy. Training Bloom, a large language model, consumed as much energy as 30 homes in a year and emitted 25 tons of CO2 (equivalent to driving a car five times around the planet).
• Model Size: AI models are rapidly increasing in size ("bigger is better" trend). Large language models have grown 2,000 times in size in the last five years, leading to increased environmental costs.
• Carbon Emissions: Switching to a larger language model can emit 14 times more carbon for the same task (e.g., telling a knock-knock joke). Tech companies often don't measure or disclose these emissions.
• CodeCarbon: A tool developed to estimate the energy consumption and carbon emissions of AI training code. It helps in making informed choices about model selection and deployment on renewable energy sources.
• Mitigation Strategies: Choosing more sustainable models and deploying AI on renewable energy can drastically reduce emissions.

Copyright and Data Usage

• Data Set Training: AI models are often trained on data (art, books, images) created by artists and authors without their consent.
• "Have I Been Trained?": A tool created by Spawning.ai that allows artists to search massive datasets (e.g., LAION-5B) to see if their work has been used for training AI models.
• Case Study: Karla Ortiz: An artist who used "Have I Been Trained?" as evidence to file a class action lawsuit against AI companies for copyright infringement.
• Opt-in/Opt-out Mechanisms: Spawning.ai partnered with Hugging Face to create opt-in and opt-out mechanisms for creating datasets, ensuring artists have control over the use of their work.
• Ethical Considerations: Artwork created by humans shouldn’t be an "all-you-can-eat buffet" for training AI language models.

Bias in AI Systems

• Definition: AI models can encode patterns and beliefs that represent stereotypes, racism, and sexism.
• Facial Recognition Bias: Dr. Joy Buolamwini's research showed that facial recognition systems are significantly less accurate for women of color compared to white men.
• Real-world Consequences: Biased AI models in law enforcement can lead to false accusations and wrongful imprisonment (e.g., Porcha Woodruff case).
• Black Box Problem: AI systems are often "black boxes," making it difficult to understand why they produce certain outputs.
• Image Generation Bias: Image generation systems can perpetuate biases when used in contexts like generating forensic sketches, associating terms like "dangerous criminal" with specific demographics.
• Stable Bias Explorer: A tool created to explore the bias of image generation models through the lens of professions. It reveals significant representation of whiteness and masculinity across various professions, even when compared to real-world statistics.
• Example: Scientist Image: Image generation models often depict scientists as men in glasses and lab coats, reinforcing stereotypes.
• Impact on Professions: Models often depict lawyers and CEOs as men almost 100% of the time, despite real-world diversity.
• Application at the UN: The Stable Bias Explorer was presented at a UN event about gender bias as an example of how to make AI accessible and understandable to people from all walks of life.

AI Governance and Societal Impact

• Accessibility: It's crucial that AI remains accessible so that people understand how it works and when it doesn't.
• No Single Solution: Complex issues like bias, copyright, and climate change require multifaceted approaches.
• Creating Guardrails: Tools to measure AI's impact can help create guardrails to protect society and the planet.
• Stakeholder Roles:
  • Companies: Can use impact data to choose more sustainable and copyright-respecting models.
  • Legislators: Can use these tools to develop new regulations and governance for AI.
  • Users: Can use the information to choose AI models that are trustworthy and don't misuse data.
• AI is Not a Done Deal: We are building the road as we walk it and can collectively decide the direction we want to go in.

Conclusion

The speaker argues that focusing on future existential risks of AI is a distraction from its current, tangible impacts. The emphasis should be on measuring, disclosing, and mitigating these impacts through tools, transparency, and ethical considerations. By addressing issues like sustainability, copyright, and bias, we can collectively shape the development and deployment of AI to benefit society and the planet.