So how does ER work? In its purest form, advocates of ER tell students to ignore the grammar and not worry about unknown words. They promise that by aiming for fluency, reading will lead them to proficiency. A central tenet of ER is that learners internalize language structures implicitly through extensive, enjoyable reading, which leads to natural fluency over time. ER aligns with Stephen Krashen’s Input Hypothesis, which posits that language acquisition occurs through exposure to comprehensible input that includes some items slightly beyond one’s current level, thereby allowing learners to intuitively grasp language rules and usage without explicit instruction (Krashen, 1982; Day, 2019). ER also allows learners to develop cultural and contextual fluency by engaging with authentic texts that reflect the subtleties of human language, including idioms, slang, and cultural references. These elements are critical for achieving advanced proficiency, as they involve more than just understanding vocabulary and syntax; they require an awareness of social norms, humor, and context-specific usage. 

This seems to be exactly the approach LLMs have taken, which raises a provocative question: has AI been doing ER? The similarity of LLMs and ER is uncanny, but let’s take a closer look, starting by considering what we mean by “AI.”

Artificial Intelligence is an imprecise and unhelpful term. “Artificial” is usually understood as “not real” even though the actual meaning is “man-made.” Likewise, “intelligence” is difficult to define, often leading to a lazy and circular idea of “what we humans can do but machines cannot.” As the technology progresses, the goal posts are moved to protect our vanity. We must remember that, when we are developing technology, our aim is not to match human ability, but to supersede it. We do not use bicycles because they move at the same speed as we can run, but because they can move faster, with less effort, carrying more. Our intelligence is limited by our biology, but the technology we develop is not, and at some point, AI will become more intelligent than us, and we will struggle to know what that means. This may already have happened. 

To understand what this might mean in practice, it helps to look at what exactly underlies the technology. More precise terms than AI are Machine Learning and Deep Learning. Machine learning somewhat mimics human learning by giving a computer a task and some training data, and then test data to check whether the task can be accomplished. Deep learning is so called because the algorithms which turn inputs into outputs include hidden layers. Essentially, we have a black box and do not know exactly how the results come out.   

People frequently cite mistakes AI makes as evidence of its lack of intelligence, but we should not count on AI mistakes as evidence of incapacity. Users of ChatGPT become frustrated when answers are wrong and there is no explanation why. Interestingly, humans also make mistakes and don’t know why they made them. The inability of LLMs to process numbers correctly highlights that here the AI is focusing on mastering patterns in language, not on replicating all human abilities. And anyone who has experimented with AI chatbots will be chronically aware that they usually treat rules as mere guidelines, and often ignore them completely.

Beyond these limitations, large language models sometimes show surprising emergent abilities. Although they have been described as “stochastic parrots” (Bender et al., 2021), LLMs sometimes demonstrate abilities that extend beyond simple pattern recognition. They can generate creative responses, solve novel problems, and produce outputs in complex, emergent ways that have never been explicitly programmed. This reflects an unexpected sophistication in artificial neural networks, reminiscent in some ways of emergent phenomena observed in human cognition. While the inner workings differ from the brain, the analogy emphasizes learning through extensive exposure rather than prescriptive instruction.

Both LLMs and ER learners depend on contextual input to develop proficiency. Just as ER emphasizes immersion in a wide variety of texts to internalize vocabulary and structures naturally, LLMs improve through exposure to enormous amounts of language data. The key similarity is learning by doing rather than learning by memorizing rules. In both cases, repeated engagement with meaningful content allows patterns to be internalized organically. This suggests that ER’s approach of focusing on abundant, comprehensible input remains highly relevant in modern language learning contexts. But is AI’s language training truly following ER principles, or are the similarities merely metaphorical?

While it is tempting to see the success of LLMs as simply a victory of data input over rule following, we should not overlook the exponential increase in computer power and memory size on the slow road to computer language proficiency. Computers are sometimes thought of as electronic brains, but they are very different to our own neural networks. While electronic computers ultimately derive all their functions from one-dimensional strings of ones and zeros, the 86 billion neurons in the brain each has a thousand synapses making connections in three dimensions. And the brain does not neatly divide architecture, processing and memory into ROM (read only memory), RAM (Random Access Memory) and permanent storage. We cannot therefore equate a brain to a computer with an 86-gigabyte hard drive, or say it is close to the 175 billion parameters that allowed GPT-3 to produce impressive language. However, we can see that computers have recently entered the ballpark of the human brain’s computing power, and neural networks and parallel processing imitate the way our brains process information (see Frank et al., 2008).