Teaching Machines What Matters | Laura Morselli | TEDxMirandola

Key Concepts:

• Artificial Intelligence (AI) as a decision-maker
• Bias in AI models due to training data
• Data democratization, bias mitigation, and fair representation in AI datasets
• Supercomputers and their role in AI model training (e.g., Leonardo)
• Under-representation of cultures and gender in AI outputs
• Ethical and social implications of biased AI
• The need for critical thinking when using AI tools
• Data-rich vs. data-poor regions

1. AI's Growing Influence and the Question of Representation:

• AI is no longer just a tool but a decision-maker influencing healthcare, justice, hiring, and governance.
• The speaker emphasizes the critical need to examine whose perspectives are represented in the data used to train AI models.
• AI models aren't thinking like humans; they repeat patterns found in their training data.

2. Data Sources and the Mirroring of Culture:

• AI models are trained on vast amounts of digital data from the internet: news, articles, blogs, videos (YouTube, TikTok), social media interactions, and personal messages.
• AI becomes a "mirror of the culture" that produces the most digital content, which can perpetuate existing inequalities and historical patterns.

3. The Power and Limitations of Supercomputers:

• The speaker works at a leading supercomputing facility in Europe, home to Leonardo, the 10th most powerful supercomputer in the world, located in Bologna.
• Leonardo can perform 240,000 trillion operations (2+2) per second.
• Training models like those behind ChatGPT requires an immense number of operations (10^23 to 10^25), necessitating the use of supercomputers.
• Supercomputers consume megawatts of power.

4. Data Set Decisions and Their Arbitrary Nature:

• The speaker highlights the arbitrary nature of decisions regarding which data feeds AI models.
• There's often a lack of careful consideration about what constitutes "good enough" data for training.
• The tendency to "collect everything" often results in collecting what is most popular, available, and potentially biased.

5. Bias and Under-representation in AI:

• The vast majority of internet data is in English, leading to under-representation of non-dominant languages and cultures.
• AI models may not fully understand the nuances of different cultures due to insufficient representation in training data.
• Experienced users applying critical thinking can recognize and correct biased outputs. However, organizations often assume AI outputs are neutral and correct, which is a dangerous assumption.

6. The Risk of Erasing Cultural Identities:

• If AI systems, increasingly used in education and search results, cannot grasp local knowledge, there's a risk of smaller voices being erased and cultural knowledge being flattened.

7. Gender Bias Example:

• Illustrates gender bias by showing that asking AI to create a picture of "a doctor" usually results in an image of a white male doctor, even though "doctor" is a gender-neutral term.
• Similar biases were observed when prompting the AI to depict lawyers, managers, politicians, and professors.

8. The Ethical and Social Challenge:

• The challenge is not just technical; it's ethical, social, and deeply human.
• There's a need to rethink how datasets are created, validated, and used to feed AI systems.
• The importance of recognizing data-rich vs. data-poor regions (global North vs. global South, urban vs. rural areas) and addressing disparities in data availability.

9. The Role of AI Developers and Users:

• AI developers, particularly those outside of Europe, might not prioritize social justice.
• There's a need to develop AI models that reflect desired values.
• Users should be aware of AI's limitations and apply critical thinking to its outputs, never assuming correctness or neutrality.

10. Conclusion:

• The future of AI is not just about algorithms and computing power; it's about ensuring AI represents everyone, not just the loudest voices.
• The speaker emphasizes that we still have a choice to shape a future where AI benefits all of us.

Notable Quotes:

• "AI is no longer just a tool. Right now, it is a decision maker."
• "Artificial intelligence becomes the mirror of the culture that produced the most digital content."
• "Collecting everything doesn't mean collecting fairly."

Technical Terms:

• AI (Artificial Intelligence): The simulation of human intelligence processes by computer systems.
• Supercomputer: A computer with a high level of performance compared to a general-purpose computer. Used for computationally intensive tasks like AI model training.
• ChatGPT: A large language model chatbot developed by OpenAI.
• Data Democratization: The process of making data accessible to everyone in an organization or to the public.
• Bias Mitigation: Techniques used to reduce bias in AI models and datasets.

Logical Connections:

• The video begins by introducing the pervasive influence of AI and then quickly pivots to the issue of bias in AI systems.
• It connects the source of this bias to the training data and how that data reflects existing cultural and societal inequalities.
• The speaker then illustrates these points with a discussion of supercomputers and specific examples of gender bias in AI-generated images.
• Finally, the talk concludes with a call to action for AI developers and users to be more mindful of these issues and work towards creating more equitable AI systems.