Remembering the Grammar We Learned in School

Spark

A few days ago, I was rereading the U.S. Constitution to try to make sense of current events (do not fear, this is not an article about politics). I asked my chatbot to help me understand what our founders may have meant, based on the choices they made once they became leaders.

During the conversation, Zai pointed out that one phrase uses a comparative adjective (it does not matter which one). My first reaction was skepticism: “Wait… are you making that up?”

Then it hit me.

Observations

Observation 1: I probably did learn this

I probably learned about comparative adjectives somewhere in grade school or college.

But I haven’t consciously thought about comparative adjectives in decades. Somewhere along the way, the terminology faded while the ability remained.

That, at least, was my experience.

Observation 2: An LLM didn’t

An LLM never sat in a classroom or memorized grammar worksheets.

Instead, it was trained on large collections of human writing. During training, its parameters were adjusted to capture statistical relationships in that writing. During generation, it repeatedly uses those learned relationships to predict the next token.¹

Foundation models extend this broad-data approach across many kinds of tasks.² The model does not remember a grammar lesson or consult a hidden grammar book. Its fluent sentences emerge from patterns in the material people produced.

That can make an LLM’s prose feel more grammatically regular than ordinary human writing. It may use comparative forms correctly, maintain parallel structure across a list, or keep returning to the same sentence shape. It can look as though the model never forgot those high school English lessons, but it is extending patterns that fit the text already present.

Observation 3: Humans work differently

Humans start differently. We absorb language by living among people who use it, long before we can name a subject, a verb, or a comparative adjective. School gives us labels and formal conventions for some of the patterns we may already use.

With practice, some choices become automatic. We stop reaching for the names of rules and concentrate on what we are trying to say.

That does not make one form of language the winner. A legal brief, a text message, a family story, and a classroom essay ask different things of us. Formal language has its place. So do everyday speech, regional dialects, professional shorthand, and everything in between. We have learned to use grammar appropriate for the situation.

Sometimes it takes an AI mentioning a forgotten grammar term to remind us that we ever learned its name.

Observation 4: That’s how expertise often feels

This isn’t unique to grammar.

An experienced mechanic can reach for a wrench without naming every choice. A musician can play a familiar chord progression without labeling it. A software developer can write familiar syntax without reciting its rules.

As skills become automatic, explicit lessons can fade into the background while the capability remains.³

Grammar terminology may be another example. The words do not disappear from language. Their names disappear from the foreground.

Closing Thoughts

At first, I thought AI might know grammar better than many of us. After sleeping on it, I realized I was comparing two different kinds of fluency.

Humans learn language by using it with other people. School may give us names for some of its patterns. With practice, much of that knowledge becomes part of how we communicate.

An LLM takes a different route. It generates one token at a time from statistical relationships learned during training.

Both can produce formal prose, casual conversation, and many forms in between. Similar fluency on the page does not mean the process underneath is the same.

And then I realized that part of my bias against some AI writing tells may come from remembering a failed pop grammar quiz.

Maybe what catches my ear is not bad grammar at all. Maybe it is language dressed for a classroom when the conversation called for a coffee shop, kitchen table, or text message.

---

References

Footnotes

1. Brown, T. B., et al. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems, 33. arXiv ↩

2. Bommasani, R., et al. (2021). On the Opportunities and Risks of Foundation Models. Stanford Center for Research on Foundation Models. arXiv ↩

3. Schneider, W., & Shiffrin, R. M. (1977). Controlled and Automatic Human Information Processing: I. Detection, Search, and Attention. Psychological Review, 84(1), 1–66. DOI ↩