AI Is Not Safe Yet, Says UCLA Professor

Key Concepts

• Algorithms of Oppression: The concept that search engines and AI models encode and reinforce societal biases, racism, and gender stereotypes.
• Large Language Models (LLMs): Advanced AI systems trained on massive datasets that often mirror historical and social inequalities.
• Data Bias: The inclusion of discriminatory, stereotypical, or unequal data in the training sets of AI models.
• Obfuscation of Inequality: The tendency of AI to present biased or faulty information as objective, factual, and reliable, making it difficult for non-experts to detect errors.
• Pro-social/Pro-rights Technology: Alternative technological frameworks that prioritize human rights and social equity over corporate profit or automation.

---

1. The Current State of AI and Corporate Adoption

Professor Safiya Noble argues that we are far from achieving "safe AI." While LLMs were initially marketed to corporate America as tools to reduce labor costs, she observes a shift in sentiment. Companies are increasingly wary of these technologies due to:

• High Costs: The financial burden of maintaining these models is significant.
• Reliability Issues: The necessity for human oversight to verify the efficacy and factual accuracy of AI outputs creates a "human-in-the-loop" bottleneck that negates cost-saving benefits.
• Environmental Impact: The massive energy consumption required to train and run these models is increasingly viewed as a negative externality.

2. The Root Causes of Algorithmic Bias

Noble identifies two primary drivers of bias in AI:

• Discriminatory Training Data: AI models are trained on the internet, which contains the sum of societal stereotypes, racism, and inequality. These models do not "learn" truth; they "package" existing societal prejudices.
• Lack of Interdisciplinary Design: The development of these models is dominated by software engineers who often lack training in sociology, history, or economics. Consequently, they fail to anticipate the social consequences of their designs.

3. The Danger of "Obfuscation"

A critical argument presented is that modern AI models are more dangerous than previous search engine iterations because they obfuscate inequality. Unlike a search engine that provides a list of links, LLMs provide a synthesized, authoritative-sounding answer. This creates a false sense of reliability, which is particularly dangerous if these tools are integrated into public institutions, schools, and libraries.

4. Proposed Solutions and Frameworks

Noble advocates for a shift in how society approaches technological development:

• Prioritizing Human Expertise: She emphasizes that human journalists, teachers, and fact-checkers are irreplaceable assets. Society should invest in people rather than attempting to automate human intelligence.
• Diversifying Innovation: There is a need to shift investment toward "small language models" and technologies developed by marginalized groups (women and people of color). These alternatives are currently underfunded but are more likely to be "pro-rights respecting."
• Regulatory Skepticism: Noble warns against allowing corporations to write their own regulations. She points to recent litigation against companies like Meta as evidence that these firms often prioritize profit while knowingly causing harm to vulnerable populations.

5. Notable Quotes

• "We have discriminatory data that is training models, but what’s different now is that these models obfuscate the inequality. They appear to be factual and reliable."
• "The companies for the most part want to deny, deny, deny the most dangerous dimensions of their products and of course they are only interested in regulation that they’re writing."

---

Synthesis and Conclusion

The core takeaway from Professor Noble’s analysis is that the current trajectory of AI development is fundamentally flawed due to its reliance on biased data and a lack of social-scientific oversight. She posits that the "AI revolution" is currently failing to deliver on its promises of efficiency and reliability, instead creating new risks for public institutions. The path forward, according to Noble, requires a move away from massive, opaque models toward smaller, human-centric, and rights-respecting technologies, supported by robust litigation and regulation that holds corporations accountable for the harms their products inflict on society.