Can We Only Ask AI What We Already Know? Can AI Take Humanity Beyond What We've Ever Conceived?

I will deeply examine the questions "Can we only ask AI about what we already know?" and "Can AI take humanity beyond what we've ever conceived?" from philosophical, technical, and practical perspectives.

This research explores the following points:

• The relationship between AI's knowledge system and creativity (Can it generate ideas beyond existing knowledge?)
• Possibilities of new problem discovery through human-AI interaction (idea support, knowledge expansion)
• AI's limitations and how to overcome them (mutual complementation with human creativity)
• AI's role in actual research and discoveries (applications in science, art, philosophy)

AI and the Quest for the Unknown: Perspectives from Philosophy, Technology, and Practice

Philosophical Perspective: The Structure of Knowledge and the Possibility of Exploring the Unknown

Since ancient times, philosophers have grappled with the question: "Can we only ask about what we already know?" In Plato's dialogue Meno, the famous "paradox of inquiry" is presented. In summary, "If you already know the answer, there's no need to ask; if you don't know the answer, you won't recognize it even when it's given to you." Plato attempted to resolve this paradox with his Theory of Recollection, positing that human souls "already know something before birth." This means that questions about the unknown are possible, and latent knowledge can be drawn out through proper dialogue.

Kant's epistemology also provides a framework for this problem. Kant argued that human cognition has structural limits. We can only know things we can experience (the phenomenal world), and we cannot have certain knowledge about the "thing-in-itself" or metaphysical questions. From this perspective, understanding or asking about things completely outside what humanity has ever conceived (entirely unknown concepts or phenomena) seems to have fundamental limits. Since AI also learns based on data collected by humans, it may not directly teach us things "outside the scope of human experience" in the Kantian sense. However, by using AI as an extension of human perception and cognition, it might be possible to expand the range of experience itself. Just as the telescope opened up the new experiential realm of astronomical observation, advanced AI could reveal insights through massive data analysis and simulation that humans alone could not achieve.

Martin Heidegger left unique philosophical reflections on the relationship between humans and technology. In his famous essay "The Question Concerning Technology," he stated that the essence of technology is not merely a tool but a kind of "revealing of being." That is, technology (tools) defines the very mode by which we understand the world and disclose truth. According to Heidegger, modern technology reveals the world as "standing-reserve" (Gestell), positioning humans as beings confronting it. From this viewpoint, AI as technology could also transform the mode of truth disclosure. If we view AI not merely as an objective tool but as an entity that actively "brings patterns to light" from data, the boundary between human (cognitive subject) and AI (tool) becomes blurred. Indeed, in recent cognitive science, Andy Clark and David Chalmers' Extended Mind thesis argues that external tools like notebooks and calculators can function as part of human cognition, and AI may similarly change the relationship between subject and object of knowledge by extending human cognition. In other words, when humans explore the unknown using AI, it's no longer a one-way relationship of "what to ask AI," but is shifting to a relationship where AI and humans collaborate to pose questions and generate new knowledge.

In summary, philosophically, exploring the unknown is not straightforward, but humanity has ventured into the unknown relying on partial knowledge, metaphor, and imagination. With the advent of AI, this way of venturing may change. In Platonic terms, AI might serve as a dialogue partner that draws out knowledge inherent in our souls; in Kantian terms, it could be a tool that technologically expands the scope of experience to access phenomena humans didn't know. From a Heideggerian perspective, the relationship between humans and AI itself reconstructs the nature of knowledge and enables new disclosure of truth, while always harboring the danger of being consumed by technology and falling into uniform perspectives. How far can human thought extend using AI—this is a question that involves updating our philosophical views of humanity and knowledge.

Technical Perspective: AI's Knowledge System, Creativity, and Its Limits

From a technical standpoint, we need to analyze the operating principles, capabilities, and limits of current AI (especially machine learning and generative AI). Data-driven learning is central to modern AI, which learns patterns from vast existing data and performs inference and generation for new inputs based on this learning. For example, large language models like OpenAI's GPT-3 learn word occurrence probabilities from hundreds of GB of text corpora and generate text by predicting the most likely continuation of given sentences. Through this approach, AI can produce remarkably human-like text, code, and images from combinations of existing knowledge. However, the flip side is that AI's knowledge fundamentally cannot exceed the scope of training data. While GPT-3 shows general capabilities like "performing translation tasks with just a few examples even when not explicitly trained on them," there are also limits and drawbacks where it generates plausible but incorrect answers about topics not in its training data, or produces biased and noisy responses.

Such limits of generative AI's knowledge system lead to the question: "Is it ultimately just imitation of given data?" Certainly, image generation AI (like GANs—Generative Adversarial Networks) produces new images by learning from vast existing images, based on recombination of learned patterns. However, this recombination sometimes produces outputs that exceed human imagination. For example, GAN-generated artworks sold at high prices at Christie's (like "Portrait of Edmond de Belamy"), or style-transfer AI creating original images fusing completely different artistic styles, suggest the emergence of creativity in AI. Creativity researcher Margaret Boden classified creativity into three types (combinational, exploratory, and transformational), and today's AI can demonstrate combinational creativity (combining existing elements) and exploratory creativity (based on rules), but transformational creativity (rewriting the rules themselves) is still considered difficult.

On the other hand, there are examples evaluated as demonstrating "leaps of imagination" by AI. Symbolic is a move shown by DeepMind's Go AI "AlphaGo." In the second game against Lee Sedol in 2016, AlphaGo played a novel move (the famous "Move 37") unprecedented in human game records. Commentators praised it as "a very original move that humans would never play," and Lee Sedol was reportedly shocked by its unexpectedness. Indeed, AlphaGo's Move 37 was a new move without precedent in Go's approximately 2,500-year history, not existing in human joseki databases. While AlphaGo learned from human game records, it devised methods through self-play of millions of games that humans hadn't conceived. What made this possible was AI's characteristic ability to perform vast move exploration without being bound by human intuition or common sense. Though limited to a specific domain, this demonstrates that AI can create new ideas and strategies independent of existing knowledge systems.

Similarly, DeepMind's algorithm-discovery AI "AlphaTensor" discovered new fast algorithms for matrix multiplication in ways unknown to humans. For matrix multiplication requiring fewer steps than conventional methods, AlphaTensor automatically found computational procedures that humans couldn't find for decades and presented algorithms that beat existing speed records in some cases. What's interesting is that while AlphaTensor's strength was exploration unconstrained by human intuition, its weakness is that it cannot explain the reasons for its solutions to humans. That is, AI can present novel solutions by not being bound by human assumptions, but understanding and interpreting why they work still requires human analysis.

As seen above, from a technical perspective, the current limitations of AI are clear. AI extracts patterns buried in vast data and produces "newness" by combining them at high speed, but that newness depends on source data. Also, AI is weak at finding problem settings that deviate from context or common sense (problem discovery) and basically optimizes within the scope of objectives or evaluation functions given by humans. In extreme terms, it's difficult for AI alone to pose "questions humans don't know," and "questions" are set by humans while AI searches for answers. In this sense, the question "Can humans using AI only ask about what they already know?" is half yes and half no. The yes aspect is that AI is faithful to given data and instructions, so it won't spontaneously start exploring unknown problems unless humans pose questions. The no aspect is that appropriately designed AI systems (like reinforcement learning agents or evolutionary algorithms that autonomously generate and verify hypotheses) can present novel solutions as results that humans didn't anticipate. Overall, AI's creativity still requires human support and frameworks, but it's beginning to show potential for achieving leaps of combinational imagination within them.

Practical Perspective: New Insights and Limits from AI-Human Collaboration

Looking at actual applications, cases are increasing where new problem discovery or solving occurs through AI-human collaboration. From scientific research to artistic creation, AI is beginning to function beyond mere automation tools as idea partners. Below we list specific examples and consider whether there was any reaching of "things humanity hasn't thought of before" in each.

• Scientific Research (Heuristic Use): Through big data analysis, AI can discover patterns that humans might overlook and propose new hypotheses or problems. For example, by reanalyzing observational data from NASA's Kepler Space Telescope with Google's machine learning algorithm, the exoplanet "Kepler-90i" was discovered. Even subtle light variations that human astronomers would have missed in manual analysis could be efficiently extracted using AI. This is a good example of AI helping discover unknown celestial bodies. In pharmaceutical development, MIT researchers used AI to discover a completely new type of antibiotic called "Halicin." The deep learning model virtually screened compound spaces too vast to explore conventionally and identified molecules with entirely different structures from existing drugs that could kill multidrug-resistant bacteria. Thus, AI dramatically expands human exploration range and is actually leading to new discoveries.

• Technology Development (Invention/Design): In engineering, AI is also presenting designs that exceed existing knowledge. Besides the AlphaTensor example discovering mathematical algorithms, in Generative Design, AI comprehensively explores design spaces to automatically generate novel product designs that humans wouldn't conceive. For example, in aircraft part design, AI-output shapes were organic like biological skeletons, greatly deviating from what human engineers typically draw, yet better met the objectives of light weight and high strength than conventional designs.

• Mathematics/Logic (Constructing New Theories): In mathematics, joint research by DeepMind, Oxford University, and others reported AI presenting new mathematical conjectures that humans then proved. In knot theory and representation theory—separate mathematical fields—AI found patterns in large datasets suggesting "unknown relationships between certain invariants" and presented these relationships as conjectures (hypotheses) to researchers. Human mathematicians, while initially surprised, verified these new conjectures and actually succeeded in proving theorems that were previously unknown. Results approaching 40-year-old unsolved problems were also obtained. What's notable in this case is that breakthroughs emerged precisely through AI-human collaboration. AI proposed patterns with vast computational resources, but humans selected among them and elevated them to certain knowledge through proof. The researcher stated that "machine learning can be a powerful framework guiding mathematical intuition," suggesting that AI complemented and expanded human ideas to contribute to unknown theory construction.

• Art (Creation and Expression): In the art world too, new expressive styles and works are emerging from human-AI collaboration. There are movements viewing AI as creative partners, such as painters collaborating with AI on paintings or composers co-creating music with AI. Recent research reported that when using generative AI (GPT-type models) for story idea generation, stories became "more creative and interesting" especially for people with lower baseline creativity. On the other hand, concerns are raised that if everyone relies on AI, works may become similar to each other, potentially losing diversity of creativity. This is a practical point requiring attention—simply accepting AI suggestions can lead to mass production of existing patterns—underscoring the importance of combining it with human ingenuity.

From these cases, it's clear that achievements breaking through the walls of existing knowledge are gradually emerging using AI. Particularly for patterns humans couldn't discover because data volume is too vast, or design problems where human intuition doesn't work well because dimensions are too high, AI is becoming a tool that expands the reach of human thought. At the same time, many of these achievements are accomplished through AI-human collaboration. AI presents new possibilities, and humans evaluate and judge them to polish them into valuable knowledge. Therefore, for the question "Can AI help humanity reach beyond what we've ever conceived?", a practical answer is "Not by AI alone, but through dialogic collaboration between humans and AI, there is possibility."

However, current AI also has clear limitations. First, because understanding of context and common-sense judgment is weak, there's a risk of completely nonsensical or harmful suggestions. This is because AI responds based on statistical correlations without understanding meaning, unable to judge the quality of ideas against experience and values like humans. Second, because AI optimizes according to reward functions and evaluation metrics, it's weak at creating objectives themselves. Defining new research topics or setting unprecedented goals remains the role of humans. Third, because AI's internal processes tend to be black boxes, the validity of novel ideas AI proposes isn't immediately understandable to humans. Therefore, even when AI produces groundbreaking solutions, humans might not trust them or might overlook them.

Then, how can these limitations be overcome? Several approaches are being considered:

• Exploration with human guidance: Rather than giving AI complete autonomy, a framework where humans set interesting questions and AI freely experiments within them is effective. In the mathematics example, humans asked AI "Can you find any patterns from this dataset?", AI presented pattern candidates, and humans evaluated and scrutinized them. Thus, by running a loop of human intuition and AI computational power, areas unreachable by either alone can be entered.

• Ingenuity in evaluation functions (introducing heuristic search): Traditional AI learns according to clear goals (winning/losing, error minimization, etc.), but for unknown discoveries, goals are often unclear. Therefore, researchers are exploring reinforcement learning that includes "novelty" in rewards and evolutionary algorithms that produce diverse solutions. For example, in AlphaTensor's automatic program generation discovering novel algorithms, differences from known algorithms and improvement in computational efficiency were themselves pursued. Going forward, by incorporating mechanisms that motivate AI to "try unknown areas" autonomously (like artificial curiosity and exploration bonuses), bolder exploration of ideas can be expected.

• Integration with human knowledge: Rather than pure data-driven approaches alone, incorporating knowledge systems accumulated by humans such as logical knowledge and physical laws into AI can reduce off-target results while aiding new discoveries. So-called neuro-symbolic AI (fusion of neurons and symbols) is an approach that uses both learned patterns and explicit rules for reasoning. This allows AI to devise new combinations while maintaining consistency and validity that humans value. If AI can have meta-cognition like "this is realistically meaningful" or "this range is unexplored" against existing knowledge, it could generate unknown ideas more effectively.

• Improving explainability: Research making AI present reasons for its conclusions and proposals in human-understandable form is also important. With advances in Explainable AI (XAI), if even partial interpretation of black-box contents becomes possible, humans can more easily accept novel ideas that were previously hard to accept. For example, if there are tools visualizing the basis for AlphaGo's move evaluations or why generative models suggested designs, the barrier to humans adopting AI discoveries lowers, ultimately facilitating the incorporation of unknown knowledge.

In summary, using AI for humanity to step "beyond what we've ever conceived" may be possible, but it requires approaches that leverage the strengths of both AI and humans. Humans pose questions and judge value; AI explores vast possibility spaces and makes proposals. Philosophically, this is an extension of the cognitive subject; technically, it's achieved through algorithm and data ingenuity; practically, through designing collaboration. Currently, AI is an existence that complements human imagination, but in the future, a reversed intellectual partnership where AI itself generates new questions and humans answer them might emerge. What's important is integrating humans' curiosity and insight with AI's computational power and pattern discovery ability to expand the horizon toward the unknown. Through this, humanity's intellectual quest may transcend its previous trajectory and reach new frontiers.

References/Sources: Plato's Meno, Kant's Critique of Pure Reason, Heidegger's "The Question Concerning Technology," Stanford Encyclopedia of Philosophy, MIT News, DeepMind/Oxford University news, and other academic articles.