TED Talk Summary: AI is dangerous, but not for the reasons you think by Sasha Luccioni

TED Woman Talk in Oktober 2023 | Atlanta Georgia

Sasha Luccioni, an AI researcher with over a decade of experience, recently received a perplexing email from a stranger who claimed her work in AI could lead to the end of humanity. Despite the dramatic nature of the message, Sasha acknowledges that AI is a hot topic, frequently making headlines. While some news reports highlight AI’s potential for positive advancements, such as discovering new molecules for medicine or amusing viral moments, others reveal troubling issues, such as harmful chatbot suggestions or unethical AI meal planner recommendations.

AI’s impact on society is multifaceted. While it’s challenging to predict the long-term future of AI, Sasha is deeply concerned about its current negative effects. AI models contribute to climate change, use artistic works without permission, and can perpetuate discrimination. She emphasizes the importance of tracking AI’s impacts, ensuring transparency, and developing tools to improve its reliability and sustainability.

For instance, the environmental cost of training AI models is substantial. Sasha was involved in the BigScience initiative, which created Bloom, an open large language model designed with a focus on ethics, transparency, and consent. Her study found that training Bloom consumed as much energy as 30 homes over a year and emitted 25 tons of CO2 — equivalent to driving around the world five times. This contrasts sharply with models like GPT-3, which have a far greater carbon footprint. Unfortunately, tech companies often fail to measure or disclose these environmental impacts.

AI models have grown significantly in size over the past five years, which has escalated their environmental costs. Recent findings show that larger models emit 14 times more carbon than smaller, more efficient ones for the same task. As AI models become more integrated into everyday devices, the environmental burden increases. Sasha advocates for the development of tools like Code Carbon, which estimates the energy consumption and carbon emissions of AI training, to help make more sustainable choices.

Beyond sustainability, Sasha addresses other pressing issues, such as the unauthorized use of artists’ works in AI training datasets. She highlights Spawning AI’s tool, “Have I Been Trained?”, which allows individuals to search large datasets for their work. Sasha used this tool to check if her images were included in datasets and found that while some images of her were present, others were not. This discrepancy explains why AI-generated images often misrepresent individuals, as seen when AI models generate unrealistic or inappropriate images based on names.

Sasha also discusses the issue of bias in AI models. AI systems can encode harmful stereotypes and biases, as exemplified by Dr. Joy Buolamwini’s work showing that facial recognition systems perform poorly for women of color. Biased AI models can lead to serious consequences, such as wrongful accusations and imprisonment. Sasha created the Stable Bias Explorer to explore and address biases in image generation models, revealing a significant representation of whiteness and masculinity in various professions.

Despite these challenges, Sasha believes in the potential for AI to be managed responsibly. She emphasizes the need for accessible tools that allow people to understand and address AI’s impacts, from environmental effects to issues of bias and copyright. By creating and using such tools, stakeholders — including companies, legislators, and users — can make informed decisions and contribute to the development of more ethical and sustainable AI.

In response to the email claiming her work could end humanity, Sasha asserted that focusing on hypothetical future risks diverts attention from the urgent, tangible impacts of AI today. She believes that addressing these immediate concerns is crucial as we continue to develop and deploy AI technology, stressing that the path forward is not predetermined and that collective action can guide AI’s impact in a positive direction.