Limits of AI: Why AI Struggles with Tacit Skills and Innate Human Abilities

Despite the promise and significant investment, upon getting stuck, high-profile AI projects raise the issue of the limits of AI. The challenge of achieving broad, reliable, and human-like intelligence in artificial intelligence (AI) has been a matter of both promise and frustration. While AI applications have shown impressive results in areas such as data-driven predictions, language processing, and visual recognition, the limitations are also evident. The cases of autonomous vehicles and IBM’s Watson in healthcare highlight a recurring issue: progress with automating codified, structured tasks often stalls when AI faces tasks requiring human intuition, Tacit Knowledge, and the kind of common-sense reasoning that’s second nature for humans. Besides, the impediments to automating innate abilities have stalled Honda’s ASIMO–Humanoid for elderly care. This essay explores why many advanced AI projects encounter hurdles when aiming to automate human-like capabilities and why achieving true general intelligence remains a distant goal.

Introduction to AI’s Initial Successes and Partial Progress

AI’s initial successes largely rely on advancements in machine learning and deep learning, where algorithms are trained to recognize patterns and make decisions based on vast data. This type of intelligence, often termed Narrow AI, has shown promise in fields where problems can be broken down into structured data-driven tasks. The progression has been rapid: AI algorithms now excel at tasks like image classification, speech recognition, and language processing. This proficiency led to rapid investments across various industries, with companies envisioning AI-driven transformations in fields from healthcare to transportation.

However, these applications are often limited to well-defined, Codified Knowledge. Codified knowledge consists of explicit information that is easy to structure and encode. Tasks like image labeling, predictive analytics, and even basic customer service automation leverage codified data, allowing AI to replace or supplement human labor efficiently. As a result, AI has seen success in fields such as retail for customer personalization and finance for fraud detection.

AI’s Limitations in Automating Tacit and Innate Human Abilities

Despite these successes, problems arise when AI attempts to handle unstructured, nuanced tasks that rely on tacit knowledge—the implicit, experiential knowledge humans acquire through life experience—and innate abilities like common-sense reasoning and contextual awareness. For example, the $80 billion invested in autonomous vehicles has yet to yield a fully reliable, universally deployable self-driving car. The setbacks in autonomous driving arise largely from real-world complexities: unexpected environmental changes, unpredictable human behavior, subtle signals and a need for instantaneous decision-making that current AI cannot yet emulate.

Similarly, IBM’s Watson experienced setbacks in automating medical prescription and diagnostic processes. Watson was designed to analyze vast amounts of medical data and assist in decision-making. However, it struggled to replicate the intuitive decision-making of physicians who draw on years of clinical experience and a kind of common-sense reasoning that’s difficult to codify in algorithms. Watson’s limitations in this area illustrate the challenge AI faces when encountering tasks that require subjective judgment, empathy, and situational awareness. Although progress is being made to overcome the limits of AI, innate abilities could be sharpened to increase the hurdle.

The Role of Tacit Knowledge in AI’s Shortcomings Underpin AI Limits

Tacit knowledge represents the implicit understanding that individuals accumulate over time, often specific to context and difficult to articulate. Tasks like recognizing sarcasm, understanding social cues, or performing a medical diagnosis often involve layers of contextual knowledge. In autonomous vehicles, for example, a human driver can instinctively anticipate the movements of a pedestrian based on subtle visual cues, while an AI might not. This “common sense” relies on a vast network of tacit understanding about the physical world, social norms, and expected behaviors.

To understand why AI struggles with tacit knowledge, it’s essential to examine the structure of current machine learning and deep learning systems. These systems are built on statistical methods, which means they learn from vast datasets to predict outcomes based on patterns. While these systems can learn patterns, they cannot develop true comprehension. For instance, an AI might recognize that a pedestrian is crossing the street but may not fully interpret unusual pedestrian behavior that deviates from its training data.

Premature Saturation and the Diminishing Returns of AI

Premature saturation occurs when initial rapid advancements in a technology slow down as more complex challenges emerge, leading to diminishing returns. In AI, this is evident when the problem shifts from data-driven decision-making to tasks requiring contextual judgment and situational awareness. In fields like healthcare, natural language processing, and robotics, AI’s performance improves quickly at first but eventually reaches a plateau, where further advancements require disproportionately larger investments with smaller gains. The phenomenon was starkly illustrated by IBM Watson’s experience in healthcare. While Watson excelled in processing large datasets, it struggled to interpret complex, ambiguous medical cases that human doctors handle routinely.

This saturation effect is partly due to the brittle nature of AI models. They excel in environments similar to their training data but often fail when applied to novel situations or edge cases. Autonomous vehicles, for instance, can perform well on familiar roads but falter in rare scenarios like handling unusual road signs or unpredictable pedestrian actions. This brittleness highlights AI’s limitations in achieving generalization, a key component of human intelligence that allows us to transfer knowledge and skills across domains.

The Challenge of Crossing the Threshold to General Intelligence

The AI threshold represents the barrier between narrow AI and artificial general intelligence (AGI), where machines could perform any intellectual task that a human can. While narrow AI can surpass humans in specific tasks, AGI remains an elusive goal. As AI systems reach more complex tasks requiring emotional intelligence, empathy, or nuanced judgment, they encounter the challenge of replicating the human mind’s adaptive flexibility. For autonomous driving, this threshold includes the ability to make nuanced decisions based on incomplete information, an area where human drivers excel.

Furthermore, tasks requiring an ethical judgment or an understanding of social nuances are often out of reach for current AI systems. Consider the field of law or medical ethics, where professionals must weigh the potential outcomes and societal implications of their decisions. Although AI can assist with data gathering and preliminary analysis, the ultimate decisions in these fields still rely on human judgment—a skill that emerges from a combination of training, experience, and empathy, not codifiable knowledge.

The Future of AI: Incremental Advancements and Ethical Considerations

Despite these challenges, AI is poised to continue evolving, albeit incrementally. Innovations in Reinforcement Learning, unsupervised learning, and Explainable AI aim to address some of the limitations inherent in current models. However, experts caution against overselling AI’s capabilities. While AI may supplement human labor in fields like healthcare, finance, and education, it is unlikely to replace the nuanced judgment and empathy of human professionals fully.

Moreover, as AI becomes more integrated into society, ethical considerations become paramount. Ensuring that AI systems are transparent, explainable, and free from bias is essential for maintaining public trust. In autonomous vehicles, for example, ethical dilemmas surrounding decision-making in life-or-death situations remain unsolved. As AI systems approach human-level complexity, ethical oversight will play a crucial role in guiding their development and deployment.

Conclusion: The Path Forward for AI

While AI holds immense potential, it remains constrained by limitations when tasked with replicating the depth of human intuition, judgment, and common sense. Hence, its possibilities are fraught with pervasive uncertainties. The cases of autonomous vehicles and IBM Watson in healthcare underscore that while AI can automate structured, codified knowledge, it often falls short when encountering tasks that require tacit understanding or contextual awareness. To navigate these challenges, AI development will likely focus on augmenting, rather than replacing, human labor in complex domains. Ultimately, AI’s future may lie in a symbiotic relationship with human professionals, where machines handle structured, data-driven tasks while humans provide the empathy, ethical judgment, and situational awareness that current AI systems lack. Despite progresses, whether we will be able to overcome the limits of AI fully is yet to be seen.

Key Takeaways of Limits of AI

Here are five key takeaways from the essay on AI’s limitations in fully replicating human intelligence:

1. Narrowing AI’s Success: While AI excels in narrow tasks that rely on structured, codified knowledge (like data analysis), it struggles with areas requiring tacit knowledge and context-driven understanding, highlighting the limitations of narrow AI versus human adaptability.
2. Premature Saturation in High-Profile Projects: Large-scale AI projects like autonomous vehicles and IBM’s Watson have shown signs of premature saturation. Despite billions spent, progress slows significantly as the technology reaches a plateau, indicating limits when AI systems attempt tasks demanding human-like intuition and flexibility.
3. Challenges in Automating Tacit Skills: AI systems face barriers when dealing with tasks that involve tacit knowledge, such as common-sense reasoning or intuitive decision-making, which are critical in areas like medicine and autonomous navigation.
4. Investment in AI and Diminishing Returns: The high costs associated with pushing AI beyond its current capabilities often lead to diminishing returns, where additional investment yields less significant progress. This is especially noticeable in fields attempting to move AI from narrow, rule-based tasks to more complex, generalizable tasks.
5. The Need for Explainable AI: As AI applications become more embedded in critical areas, Explainable AI (XAI) is necessary to bridge the trust gap, especially in fields like healthcare, finance, and autonomous driving, where understanding AI decisions is vital to gaining user acceptance and regulatory approval.

Research Questions About Limits of AI

Here are some research questions based on the exploration of AI’s progress and challenges in automating human knowledge:

1. What factors contribute to the limitations of AI in automating complex tasks that rely on tacit knowledge and innate abilities compared to codified knowledge?
   • This explores the differential capabilities of AI when addressing explicit, rule-based information versus context-dependent, experiential knowledge.
2. How do high-profile AI projects, like autonomous vehicles and medical AI, illustrate the barriers in crossing from narrow AI capabilities to general intelligence?
   • A question to investigate specific project setbacks and the broader implications on AI’s scope and future.
3. To what extent do diminishing returns affect the investment in AI for high-stakes applications, and what are potential strategies to overcome these limitations?
   • This considers the economic and strategic aspects of AI R&D, especially when projects reach a plateau.
4. How does the concept of Explainable AI (XAI) influence user trust and regulatory approval in fields where AI’s decision-making is critical?
   • Examines the importance of transparency in AI to improve trust and facilitate integration in sectors like healthcare and finance.
5. What role does human tacit and experiential knowledge play in AI development, and how can AI researchers better incorporate these elements into future models?
   • An investigation into bridging the gap between AI’s current capabilities and the nuanced understanding that comes with human intuition and experience.