How GenAI Thinks: Breaking Down the Mechanics of Artificial Intelligence
SingTao · 1 SOURCESabout 1 hour ago2 MIN
Summary
Human cognition and Generative AI (GenAI) operate through entirely different mechanisms. A research team at the Education University of Hong Kong (EdUHK) has demonstrated how GenAI processes information by analyzing 50,000 online posts to detect misinformation . The research reveals that GenAI navigates a conceptual 'space' where words, texts, and language categories exist as objects, generating responses by finding paths through this data landscape . Unlike human artists who infuse their work with personal memories and social realities, GenAI draws from training data spanning art, architecture, religion, and rituals—often exceeding human boundaries . Experts warn that without explicit role assignments and moral guidelines, GenAI may produce harmful content or 'hallucinations' that appear truthful but are factually incorrect .
Key Points
- EdUHK researchers analyzed 50,000 online posts using GenAI to detect fake information, demonstrating how understanding AI mechanics can guide it toward truth verification
- GenAI conceptualizes words as dimensional points in 'semantic space': for example, 'dog' corresponds to carnivore categories, 'licks' represents tactile actions, and 'boy' indicates humanity
- The 2017 study 'Attention Is All You Need' introduced a 'magnet effect' where related words attract each other, forming 'magnetized paths' that influence AI's word selection
- Real information typically produces stable and concentrated paths, while fake information generates abnormal deviations—though AI lacks inherent reasoning capabilities
- Flawed training data can cause 'model collapse' where errors multiply exponentially when AI-generated content is fed back into training cycles
Why It Matters
For Hong Kong's digital landscape, where misinformation spreads rapidly across social platforms, understanding GenAI's limitations becomes essential for educators, journalists, and policymakers . The research underscores that while GenAI can be a powerful analytical tool, human verification remains indispensable—AI should be used to structure and analyze information, not as an infallible source of truth .
For Hong Kong's digital landscape, where misinformation spreads rapidly across social platforms, understanding GenAI's limitations becomes essential for educators, journalists, and policymakers . The research underscores that while GenAI can be a powerful analytical tool, human verification remains indispensable—AI should be used to structure and analyze information, not as an infallible source of truth .