Can AI ever surpass human intelligence?

When reading about AI, comparisons often emerge suggesting that certain models reach specific levels of intelligence, sometimes measured using tests designed for humans. We typically express human intelligence using IQ scores - the higher the IQ, the more intelligent a person is. Comparisons of models like Grok, Copilot, Ernie, ChatGPT, and Google Gemini often include IQ-like assessments across various domains.

With the rapid pace of technological advancement, many assume AI models will one day surpass human intelligence, potentially reaching four-digit IQ levels. Given the exponential growth in computational power over the last century, is this likely to happen soon?

This article explores when, if ever, AI might reach or exceed human intelligence, and whether advancements in compute power lead to linear increases in AI capabilities. I consulted the latest models to gather insights and aimed to provide you with the tools to consider these questions yourself. I think it’s both fascinating and essential to understand the capabilities and limitations of language models to maximize their potential.

Let’s dive in to explore how AI could one day, perhaps, surpass human intelligence.

The Architecture of Artificial Intelligence

The idea of AI models reaching or exceeding human “IQ levels” is compelling but also complex, as AI and human intelligence are fundamentally different. While human IQ reflects broad problem-solving abilities, reasoning, and adaptability across varied domains, AI models are specialized tools that excel at pattern recognition within specific constraints. They are not directly comparable to human IQ.

Improvements in current AI models are largely achieved by increasing scale - adding parameters, data, and sophisticated algorithms. However, practical limits exist. As models grow, they face challenges like diminishing returns, increased computational demands, and efficiency trade-offs. At a certain point, adding more parameters may yield only marginal gains, suggesting a possible “glass ceiling.”

Self-Improvement

Language models operate within fixed parameters and cannot redesign or improve themselves. While they can simulate adaptive behaviors and generate complex language patterns, they don’t independently evolve their architecture or create new learning mechanisms. This lack of self-modification limits their potential to “break through” IQ-like thresholds, as they rely entirely on human-directed updates and architecture.

AI models excel in narrow, specialized tasks but lack general intelligence - the adaptable, cross-domain reasoning that characterises human IQ. Even with advanced architecture, an AI may remain constrained by its specialization, limiting its capability for “higher IQ-like” flexibility and creativity.

Creativity and Novel Structures

For an AI to transcend its limits, it would need to develop new structures or languages of thought, similar to how human geniuses invent new theories or fields of knowledge. Current models lack true creativity - they generate patterns based on past data, not through inventing original methods of reasoning. This means that while language models can deepen within a given domain, they would struggle to create entirely new paradigms on their own.

For language models to “surpass” a certain IQ level, they might require architectural shifts allowing for autonomous learning, cross-domain reasoning, or self-modification - capabilities beyond current transformer-based models. Approaches that incorporate symbolic reasoning, causality, or neuro-inspired architectures might facilitate more complex, generalized problem-solving. However, these would need to enable AI to function more autonomously, resembling human-like thought rather than simply scaling up pattern recognition.

Emergent Abilities and Beyond-IQ Capabilities

Language models sometimes display emergent abilities - skills that arise unexpectedly at certain scales, like rudimentary problem-solving. However, these abilities are often limited and unpredictable, suggesting that scaling alone may not produce the directed, high-level intelligence associated with human IQ.

While language models can theoretically grow more powerful and sophisticated, they’re limited by architectural constraints, lack self-directed improvement, and cannot reinvent their frameworks. Without breakthroughs enabling self-modification, cross-domain reasoning, and true general intelligence, language models may approach a practical “IQ limit,” where gains in complexity no longer translate to broader intelligence.

The Real IQ of AI

Assigning an “IQ” to language models is misleading because IQ measures human-like general intelligence across diverse cognitive abilities.

Here’s why IQ doesn’t meaningfully apply to language models:

• Human IQ reflects a wide range of abilities, whereas language models are narrow systems optimized for language tasks. They don’t possess generalized reasoning abilities or flexibility beyond language processing.
• Language models excel at recognising patterns in data but lack core cognitive functions like self-reflection, independent learning, and deep reasoning. They don’t “understand” the information they process or form concepts beyond statistical patterns in language data.

Lack of Consciousness and Intent

Human IQ is often tied to intentionality, awareness, and goal-oriented problem-solving. AI models lack goals, self-awareness, and intentionality; they predict responses based on training data without understanding or purpose.

IQ also implies the ability to learn and improve independently, but language models operate within fixed parameters determined by human training. They don’t improve over time, nor can they modify their own architecture or learn from new experiences in real time as humans do.

Although language models can demonstrate emergent abilities, such as solving logic puzzles, these abilities reflect training on vast datasets, not evidence of a generalized intelligence that aligns with IQ.

While it’s tempting to use IQ as shorthand for model sophistication, it’s a misleading analogy. Instead, language models are best understood by evaluating their capabilities within specific tasks and limitations, not by applying human-centric measures.

Evolution of Intelligence

Human evolution is a slow process, particularly in terms of biological changes. We don’t see significant increases in human intelligence within a few generations. However, certain aspects of intelligence, like problem-solving and knowledge access, have expanded through cultural evolution, technology, and improved education. By contrast, AI can “improve” quickly, but this improvement is fundamentally different from biological or cognitive evolution in humans.

Biological evolution occurs over thousands of years, and any genetic changes linked to intelligence take time to manifest. While humans are indeed evolving, intelligence changes due to biology alone are subtle and gradual.

Over generations, better access to education, technology, and nutrition has contributed to increased cognitive abilities. For example, the “Flynn effect” shows IQ scores rising over generations, likely due to schooling, problem-solving experiences, and societal complexity rather than a genetic increase.

Can Artificial Intelligence Evolve?

AI models improve rapidly with faster computers, larger datasets, and advanced architectures. However, this improvement isn’t autonomous - it requires human intervention, redesign, and retraining. This is more akin to “engineered evolution” than natural evolution.

Current AIs don’t evolve independently, self-modify, or pass on “traits” as organisms do. AI models rely on fixed architectures and only improve when researchers create new versions.

Self-Learning Models in the Future

Some AIs engage in limited self-learning (e.g., reinforcement learning systems like AlphaZero), but they function within strict parameters and aren’t capable of open-ended, cross-domain learning like humans.

For AI to “evolve” like biological systems, we’d need systems that autonomously alter their architectures, test new configurations, and pass on successful “traits.” This concept, often referred to as “AI self-improvement” or “recursive self-improvement,” remains theoretical.

Human vs. AI Intelligence Growth

Humans pass knowledge and skills through education and culture, with each individual building on accumulated societal knowledge. Cognitive improvements are cumulative but limited by biology.

AI improves quickly by “upgrading” with new data and architecture, but these changes don’t represent true evolutionary growth. AIs don’t adapt autonomously; instead, they rely on intentional design, resulting in narrow rather than general improvements.

Conclusion

While human intelligence shows incremental improvements due to cultural evolution, biological evolution is slow. AI, in contrast, can rapidly improve in specific areas thanks to technology, but lacks autonomy and self-directed growth needed for true “evolution.” Future advances may enable self-improving systems, but achieving something akin to biological evolution remains speculative and would require breakthroughs in autonomous learning.

So when will this shift to general artificial intelligence truly occur, and should we fear it? I hope this article has given you some insight into these questions and perhaps new ideas about the future with, rather than overshadowed by, AI and robots.