Artificial intelligence’s limited ‘intelligence’

Artificial Intelligence (AI) is everywhere, and valid anxieties about it persist. People are worried about job displacement, data privacy, the proliferation of deepfakes, and the broader impact of automation on human social interaction.

But how intelligent is AI really?

While experts debate the exact definition of intelligence, a common middle-ground definition is: “the ability to efficiently acquire knowledge, understand information, learn from experience, reason about situations, solve problems, and adapt to novel environments.”

At first glance, modern AI appears to check most of these boxes. The reality is much more nuanced.

1. Recognition is not understanding

Large Language Models (LLMs) like ChatGPT and Gemini do not "know" things the way humans do. They are trained on massive datasets comprising hundreds of billions of words, code repositories, and images. Through this training, they map mathematical relationships between words to mimic natural human language.

When we ask LLMs a question (aka prompting), it’s not fetching a fact from a traditional database; it’s guessing the next most likely word based on the statistical patterns it has learned. Because they rely on these learned patterns rather than genuine comprehension, they struggle with novelty. When asked about concepts outside their training data, they frequently hallucinate, that is, generating highly confident responses that are factually incorrect. The problem-solving capability of AI is a function of familiarity, not true reasoning. Likewise, they are sensitive to how questions are phrased, leading to the birth of the field of prompt engineering, in which we carefully structure queries to guide the AI to the correct output, aiming to maximise the AI’s performance.

2. The Efficiency Gap

At the core of true intelligence is efficiency: the ability to form useful abstractions from limited experience and apply them to navigate uncertainty.

Humans are incredibly efficient learners. A child can learn to distinguish cats from dogs after seeing only a few examples, then generalise that concept to new environments, lighting conditions, and even unforeseen breeds. In contrast, AI models need tens of thousands of images of cats and dogs, as well as massive computational power, to achieve comparable results. True intelligence is about maximising output while minimising cognitive and physical resources. AI operates on the opposite principle: brute-force data consumption. This difference reflects a fundamental distinction: human intelligence is highly sample-efficient and grounded in rich prior experience across multiple modalities, while current AI systems are data-intensive and specialised, even as they become increasingly versatile.

3. The Absence of Sentience and Meaning

Human intelligence is closely tied to perception, emotion, and lived experience. We don’t merely process information; we interpret it through meaning, context, and feeling. AI, however, operates entirely without emotion or consciousness.

When humans witness a crisis, a cascade of emotional and ethical signals prompts us to feel empathy and offer aid. AI can analyse a photo of a painful conflict zone and generate a sympathetic response, but the pixels and words carry no intrinsic meaning to the machine. It is a simulation of empathy, not its existence. While robotics can replicate human facial expressions, the underlying algorithm experiences neither joy nor sorrow.

This distinction matters because human decision-making is often shaped by values, ethics, and emotional understanding, factors that are not inherent to AI systems.

4. The "Garbage In, Garbage Out" Dilemma

AI models reflect the flaws of the data they consume. They are trained on large-scale datasets that inevitably contain inaccuracies, biases, outdated information, and conflicting viewpoints. This low-quality data directly causes undesirable AI behaviours, including overconfidence in incorrect answers and systemic bias.

While AI tech giants employ armies of human reviewers to filter and fine-tune these models, the sheer volume of training data makes it impossible to eliminate these flaws entirely. As a result, AI systems can reproduce or amplify errors present in their training data, and their outputs should not be treated as inherently authoritative.

Why, then, does AI appear so capable?

AI's apparent brilliance relies on two pillars: scale and external tools.

Because AI has ingested a significant portion of digitised human knowledge, it rarely encounters a problem it hasn't already seen in some form. It excels at synthesising and rephrasing existing information, which can easily be mistaken for original thought.

Furthermore, modern AI is frequently augmented with hard-coded software tools and external systems. For example, while an LLM can perform basic arithmetic through statistical probability, it is unreliable at complex math. To fix this, developers give AI access to traditional calculators and code interpreters. When you ask an AI to solve a complex equation, it routes the problem to a rigid, human-written program to guarantee accuracy.

The same applies to autonomous vehicles. While AI predicts driving behaviour, hard-coded safety guardrails written by human software engineers override the AI to engage the brakes if the vehicle detects an imminent collision, regardless of the AI’s prediction. Safety and control are governed by deterministic software layers that operate independently of the AI model’s probabilistic outputs.

This combination of learned pattern recognition and engineered deterministic tool use creates Agentic AI systems that can appear highly capable, even in domains where they do not independently “reason” in a human sense.

Navigating the Transition

Whether we label it "intelligent" or not, AI is one of the most disruptive technologies ever created, boasting immense potential to boost productivity. AI models are reshaping industries, increasing productivity, and changing the nature of work. This transition is not without disruption.

As a software engineer, I see my field serving as the canary in the coal mine. Some tech leaders have rushed to replace human developers with AI to cut costs. This is a short-sighted strategy because AI systems still (and always should) require human oversight, especially to ensure accuracy, safety, security, and maintainability. Since AI is trained on vast amounts of sub-optimal or flawed public code, it frequently generates insecure or inefficient software that software engineers need to identify and fix. Without human experts to conduct rigorous code reviews and quality control, organisations relying solely on AI-generated software are building on shaky foundations. Forward-thinking leaders recognise that AI is an amplifier, not a replacement; they are using it to supercharge their existing engineering teams, allowing them to take on more work and scale safely.

We cannot avoid this transition, but we can manage it. As we integrate AI into our businesses and daily lives, our focus must be on understanding its limitations, training our workforce in new skill sets, establishing robust regulatory and security guardrails, and ensuring human oversight. AI is a powerful tool, but the responsibility to guide it safely into the future remains entirely human.