How AI influences our opinion | Laura Lewandowski | TEDxKLU Hamburg

Key Concepts:

• AI's influence on information and thought
• Bias in AI models (pre-training, fine-tuning, optimization)
• Censorship and control of information by AI
• Supervised thinking and outsourcing intuition to AI
• AI as a sparring partner, not a substitute for thinking
• Socratic dialogue, ARS total principle, and emotional reasoning as frameworks for interacting with AI

1. Introduction: From AI Skeptic to Concerned Journalist

• The speaker, a journalist, initially felt proud of writing all her content.
• She was skeptical of AI but became curious and experimented with tools like ChatGPT and Midjourney.
• She experienced increased productivity and efficiency using AI.
• She felt she had "cracked the code" as her reach grew while others invested heavily in content creation.

2. The Turning Point: AI Rewriting Thoughts

• During a heated political discussion with figures like Philip Bagos (known for his friendship with Javier Mille), the speaker recorded Bagos's words to create a LinkedIn post using AI.
• The AI didn't just transcribe her thoughts but changed them, flagging her content as problematic and suggesting "more respectful phrases."
• This incident led her to question whether AI has an opinion.
• She realized AI doesn't think or feel but reflects familiar data, raising concerns about who controls that data.

3. Silicon Valley Insights: AI as a Censorship Machine

• The speaker visited Silicon Valley, expecting to find AI evangelists but instead encountered skeptics.
• Mark Andreessen, a prominent investor, stated that "AI is a censorship machine from the beginning and it's 100% intentional."
• The talk shifts focus from AI tools to the control of information.
• The speaker warns that increased AI usage influences what we search, how emails sound, the news we read, and how professionals perceive truth.
• She emphasizes that if the system is biased, so are we.

4. Three Layers of AI Influence

• The speaker identifies three key layers of influence in AI: pre-training, fine-tuning, and optimization.

  • Pre-training (Unsupervised Learning):
    • AI learns by processing vast amounts of internet data, including opinions, facts, and judgments.
    • The speaker uses the analogy of a one-year-old daughter learning language by listening.
    • She raises concerns about biased information in the training data, citing Mark Zuckerberg's admission that governments pressured tech companies to filter information during the pandemic.
  • Fine-tuning:
    • AI is taught what is good or bad through Q&A examples, similar to a parent teaching a child.
    • Real humans, sometimes in countries like Nigeria or Kenya earning $2 an hour, provide this guidance.
    • The speaker highlights that these workers may not even know for which company they work.
    • She provides an example of asking ChatGPT for the lyrics of Kanye West's "Stronger," which is denied due to copyright restrictions, while a Chinese AI (Deepseek) provides the lyrics.
    • She notes that Chinese AI companies avoid topics like Taiwan or democracy, while American companies may lean towards left-wing viewpoints.
    • A study by the University of Anglia found that many AI models avoid controversial topics and promote progressive viewpoints.
  • Optimization:
    • Large language models are continuously trained and optimized.
    • Users are often asked to upvote or downvote examples, providing data for further refinement.
    • This data is collected and analyzed by humans in a laboratory setting.

5. The Question of Transparency and Control

• The speaker emphasizes that the issue is not whether AI is left or right-leaning but about transparency.
• The rules and agenda of AI models are not visible to users.
• She notes that OpenAI, initially considered left-leaning, now aims to be more open to controversial topics and free speech.
• She mentions that big companies are working with publishing houses like Le Monde, Wall Street Journal, and Axel Springer to train models on their media content, raising concerns about the impact of controversial media outlets on AI outputs.

6. The Impact on Individual Thought and Behavior

• The speaker recalls her experience of AI trying to change her words, leading to insecurity and self-doubt.
• A study by Microsoft and Carnegie Mellon in 2025 found that the less secure people feel, the more they rely on AI and the less critically they think.
• Michael Mine, a researcher from LMU Munich, coined the term "supervised thinking" to describe the effect of outsourcing intuition to AI.
• The speaker fears that AI's eloquence and intelligence may silence individuals and prevent them from speaking their truth due to fear of online cancellation.

7. Reimagining Our Relationship with AI

• The speaker emphasizes that we should not see AI as a substitute for our own thinking but as a "sparring partner."
• We should not be afraid of AI but see it as a tool to help us achieve excellence.
• She uses the analogy of an intern, suggesting that AI can provide a 50% perfect example, but it is our job to fact-check and refine it.
• She acknowledges that adapting to AI requires unlearning certain behaviors, similar to historical fears about writing and calculators.

8. Ancient Wisdom for the AI Age

• The speaker suggests drawing on the wisdom of the ancient Greeks to navigate the AI age.

• She proposes three mental frameworks:

  • Socratic Dialogue: Challenge AI's ideas by asking probing questions and seeking counterarguments.
  • ARS Total Principle: Understand both sides of an argument to arrive at a new reality.
  • Stoic Emotional Reasoning: Don't blindly accept AI's outputs just because they sound good; critically evaluate them.

9. Conclusion: The Importance of Independent Thinking

• The speaker emphasizes that AI cannot replicate human emotions and experiences.
• She believes that the ability to think, debate, and expand knowledge independently is crucial.
• She concludes that we don't need better AIs, we need better thinkers.