AI-Generated Content Surpasses Expectations in Emotional Resonance

The Surprising Weakness of AI in Visual Perception

Despite dizzying advancements in the artificial intelligence (AI) world, it has been revealed that the latest technology models are far behind expectations in basic visual perception. A new study shows that multimodal artificial intelligence systems like GPT-4o and Google's Gemini can barely surpass a 50% success rate in tests involving recognizing and interpreting simple objects. This situation indicates that there is still a long way to go for AI to achieve human-like seeing and comprehension skills.

The Striking Results of the Research

Researchers showed the models images of objects frequently encountered in daily life and asked them to describe, count, or understand the relationship between them. The results revealed that, contrary to their impressive performance in complex text generation, the models face serious limitations in interpreting the visual world. The models struggled to understand the context of objects, notice obvious errors in images, or grasp the spatial relationship between multiple objects.

The Chasm Between 'Seeing' and 'Understanding'

These findings expose a deep chasm between artificial intelligence's ability to 'see' and its ability to 'understand'. According to Wikipedia's definition, artificial intelligence is idealized as "a system that exhibits high cognitive functions specific to human intelligence," yet current systems, while capable of processing visual data, fail to place it within a conceptual framework like humans do. This is considered one of the biggest obstacles to the reliable use of AI in the real world.

Implications for Education and Ethics

This development also reignites debates about the use of artificial intelligence in the field of education. As emphasized in the Ministry of National Education Artificial Intelligence Applications Ethics Statement, artificial intelligence should be used "solely to support pedagogical goals, enhance teaching quality, and develop students' higher-order thinking skills." However, this weakness of AI models in basic perception strengthens the view that they should be limited to supportive roles within teacher-artificial intelligence collaboration models, not as independent teaching tools.

Companies' Approach and Future Roadmap

Companies like Google state that they take user feedback seriously with the goal of making assistants like Gemini "the most useful and personal artificial intelligence assistant." However, this latest research indicates a need not just for more data and more powerful computing, but for a qualitative leap in basic perception and comprehension architectures. For AI to be truly considered 'intelligent,' it needs to comprehend visual scenes holistically and contextually, rather than breaking them down into parts.

Conclusion and Future Expectations

While artificial intelligence is advancing at an incredible pace in areas like writing, planning, and brainstorming, it is failing in understanding the visual world, one of humanity's most basic skills. This situation serves as an important warning for technology developers. Future research is expected to focus on models not just recognizing data patterns, but grasping the meaning, intention, and causality behind visual inputs. AI's ability to fully realize its potential appears to depend on closing this 'seeing-understanding' gap.