What linguistics predicts about where AI fails

In 2015 a machine learning researcher described the situation to Christopher Manning: natural language processing was a rabbit in the headlights of the deep learning machine, waiting to be flattened.

He was broadly right about what would happen and wrong about what it would mean. Hand-engineered linguistic features did stop helping. Grammar-based pipelines were replaced. Parsers and part-of-speech taggers went from being the system to being a preprocessing step nobody runs.

And yet every article in this series has been an instance of a linguistic category doing explanatory work: distributional semantics explaining why embeddings confuse opposites, pragmatics explaining why models over-explain, compositionality explaining why performance collapses on recombination, morphology explaining why arithmetic fails.

Linguistics lost the engineering argument and kept the diagnostic one. The features stopped helping because models learn the structure on their own. The categories survived because they remain the best available description of what a language system has to get right, and therefore of the specific ways it can fail.

Why linguistic features stopped working

This deserves a fair account, because the usual telling is triumphalist in one direction or resentful in the other.

Researchers did try to inject linguistic knowledge into neural models: parse trees as inputs, morphological features, explicit syntactic structure. Reviews of that literature find no substantial performance improvement over models without it.

The reason is not that the linguistics was wrong. It is that models learn the structure implicitly during ordinary training, so supplying it explicitly is redundant. Probing work established this: syntactic structure is recoverable from the internal representations of models nobody taught syntax to. Hierarchical relationships, dependency structure, morphological features, all present in the representations of systems trained only to predict text.

Once that was demonstrated, the case for hand-supplying the same information collapsed. You cannot improve a system by giving it something it already has.

Worth noting a detail that gets dropped: earlier work found that models learned syntax-sensitive dependencies well and did better when structure was modelled explicitly. The advantage narrowed to nothing as models and data grew. This is the bitter lesson in its precise form: not that structure is unhelpful, but that the benefit of supplying it shrinks faster than the cost of scaling.

The failure was annotation, not representation

The formal tradition deserves a more accurate obituary than it usually gets.

Formal semantics from Montague onward, through categorial grammars and unification-based frameworks, produced representations that were logically sound. You could perform real inference over them, check consistency, and give formal guarantees about what a sentence meant. Nothing in current systems does this.

The problem was never the representation. It was that every grammar rule had to be written by a linguist.

That is a scaling constraint rather than a conceptual one, and it is why the tradition lost to statistical methods and then to neural ones. Approaches requiring human labour per rule cannot compete with approaches requiring only compute per parameter, regardless of which produces the better description.

The distinction matters because it predicts something. If the constraint was annotation cost, then a technology that removes annotation cost changes the calculation. Systems that can produce structured representations without hand-written rules are now possible, and whether that revives any of the formal tradition is an open question rather than a settled one.

What linguistics is for now

The role changed rather than ending, and the new role is arguably more useful.

Linguistics is the instrument, not the input. The productive contact between the fields is now probing: using linguistic categories to determine what a model has represented. This works because linguistics supplies a hypothesis space. Asking whether a model represents grammatical number, or argument structure, or scope, requires someone to have identified those as things a language system must handle, and that identification is a century of work.

Without it you can only observe that a model failed on some inputs. With it you can say the model failed on long-distance agreement, which is a category with known properties, known human behaviour to compare against, and a known set of related cases to test next.

That is the difference between a bug report and a diagnosis.

The levels, and what each one predicts

The traditional division of linguistic analysis turns out to map cleanly onto distinct classes of AI failure, which is a decent argument that the division was carving something real.

Phonology and orthography, the level of form. How sounds and written symbols are structured. Predicts: failures on tasks defined over characters, since subword tokenization discards that level. Counting letters, reversing strings, rhyme, anything where the unit of analysis is smaller than a token.

Morphology, the level of word structure. How words are built from meaningful parts. Predicts: failures on agglutinative and morphologically rich languages, where a single word carries information English spreads across several, and where tokenization fragments a morpheme into pieces that are not morphemes. This is a large part of why non-English performance degrades.

Syntax, the level of structure. How words combine into phrases and sentences. Predicts: failures on long-distance dependency, unusual word order, deep nesting, and structural generalisation. It also predicts what models handle well, since the probing evidence says syntax is well represented.

Semantics, the level of meaning. How structure determines interpretation. Predicts: the antonym collision, since distributional semantics captures topic-relatedness rather than meaning; scope ambiguities; and failures where two expressions mean the same thing but are formally different.

Pragmatics, the level of use. How context and speaker intention determine what is communicated. Predicts: literal interpretation of indirect requests, over-informativeness, and failure to recover implicature from unusual phrasing.

Five levels, five distinct failure signatures. A failure that looks like general unreliability usually resolves to one level once you know the levels exist, and knowing which one tells you whether to change the input, the prompt, the architecture, or your expectations.

The five failures in this series, sorted by level

This series worked through specific failures without naming the frame connecting them. Sorted by level, the pattern is visible.

Orthographic. Character counting, string reversal, and the arithmetic failures traceable to digit segmentation. The unit the task requires is smaller than the unit the model receives, and no amount of reasoning recovers it.

Morphological. The cost and quality penalty for morphologically rich languages. A single word carrying what English spreads across a phrase gets fragmented into pieces that are not morphemes, so the model sees more units carrying less structure.

Syntactic. Structural generalisation collapse, where performance falls from the high nineties to the twenties when familiar pieces occupy unfamiliar positions. The relationship that matters is between structural positions rather than adjacent tokens.

Semantic. The antonym collision, where opposites embed close together because distributional similarity captures topic-relatedness rather than meaning. Two expressions can be maximally similar distributionally and maximally different semantically.

Pragmatic. Literal interpretation of indirect requests, and systematic over-informativeness from optimising for maximum rather than required informativeness.

Reading those together, the useful observation is that none of them is a reasoning failure, and all of them get reported as one. A user encountering any of these describes the system as unreliable or not very smart, and the correct response differs in every case: change the input representation, change the language strategy, change the test set, change the retrieval architecture, change the prompt.

That is the argument for the vocabulary. Not that linguistics improves models, which it stopped doing, but that five different problems wearing the same symptom get five different fixes, and without the categories they get one.

What linguistics knows that has not been absorbed

Four things that remain underused, stated as claims rather than complaints.

Morphemes are real units and tokens are not. Subword vocabularies are learned from frequency statistics and do not align with meaningful parts of words. A morphologically aware segmentation would put boundaries where meaning changes rather than where co-occurrence peaks. This has been tried and has not clearly won, partly because it requires language-specific resources, which is the annotation problem returning.

Grammatical structure is hierarchical, not sequential. Sentences have nested constituents, and the relationships that matter are between structural positions rather than adjacent tokens. Attention can represent this and is not required to, which is one account of why structural generalisation fails while local coherence succeeds.

Meaning is compositional and reference is not distributional. Two claims that have been argued about for a century, both underlying failures documented across this series, and neither settled by the technology.

Languages differ in what they force you to encode. Some require marking evidentiality, some require marking the speaker's relationship to the listener, some make distinctions English cannot express without a paragraph. A model trained predominantly on English learns the distinctions English forces and treats the others as optional decoration. This is a deeper problem for multilingual systems than the token-count penalty and gets far less attention.

Is the model the theory?

The sharpest current dispute deserves stating, because it is about what these fields are for.

One position, argued seriously, holds that large models are better theories of language learning than anything linguists produced. They acquire language from exposure, they generalise, and they do it without innate grammatical machinery, which is the empirical claim generative linguistics made and could not test. On this view the model is the theory, and the theory won.

The response is that a model is not an explanation. A system that reproduces a phenomenon without illuminating why it occurs has predictive power and no explanatory power, and science wants both. A weather simulation predicts rain without being a theory of atmospheric physics.

Both are right about different things, and the disagreement is about what a theory is for. If a theory is a compressed description that predicts, models qualify. If a theory is an account of underlying mechanism stated in terms a person can inspect and reason about, they do not, and interpretability research is the attempt to extract one from the other.

There is a historical observation worth attaching. This field has been through cycles before, statistical methods displacing rule-based ones, then neural methods displacing statistical ones, and researchers who have watched all of them note that familiar problems resurface in new vocabulary. The current confidence that structure is unnecessary because models learn it should be held with that in mind.

What this means practically

For anyone building rather than arguing, the payoff is a diagnostic vocabulary.

When something fails, ask which level it belongs to. A system that mangles plurals in Finnish has a morphology problem. One that loses track of what a pronoun refers to across a long document has a syntax and discourse problem. One that answers a rhetorical question has a pragmatics problem. These need different responses and the response for one is useless for another.

Test at every level, not just the top one. Most evaluation is semantic: did it get the right answer. A test set that varies morphological complexity, syntactic depth and pragmatic indirectness will surface failures that a set varying only topic cannot.

Read the probing literature for your language. If you deploy in a language other than English, work examining what models represent in that language exists and is more informative than aggregate benchmark scores.

And treat the categories as a checklist for what could go wrong. The value of a century of description is that it enumerates the things a language system has to handle. Any of them can break, and a system nobody tested at a given level has an untested level rather than a working one.

What is unresolved

Whether structural inductive bias returns. The benefit of explicit structure shrank to nothing at scale. Whether that holds for data-limited settings, low-resource languages, or architectures other than transformers is not established, and the finding that structure helped before scale suggests it may help again where scale is unavailable.

Whether probing measures what it claims. Establishing that a syntactic property is decodable from a representation does not establish that the model uses it. A probe finding structure may be finding structure the probe imposed, and this critique has not been fully answered.

Whether the annotation constraint is gone. Formal approaches lost on the cost of hand-written rules. If systems can now produce structured representations automatically, the constraint that decided the last thirty years may no longer apply, and nobody has seriously revisited what that permits.

What is lost by not asking. The strongest form of the linguistic complaint is not that models perform badly, but that a field which stops asking why has traded understanding for capability. Whether that trade has costs that show up later is not answerable now, which is exactly what makes it worth noting.

The counter-argument

The bitter lesson has an excellent record. Every domain where hand-engineered knowledge was replaced by scale saw the same outcome, and the pattern has repeated often enough that betting against it requires more than an appeal to what a field knows. Predicting that structure will return has been wrong for a decade.

Linguistic categories may not carve reality. The five levels are a useful pedagogical division and were designed to describe human language faculties, not to describe whatever a transformer does. Mapping model failures onto them may impose a familiar structure rather than discover one.

Probing has produced modest returns. The claim that linguistics contributes through probing is fair and the practical yield has been limited: a large literature establishing that models represent linguistic properties, and relatively few improvements traceable to it.

And the diagnostic framing may be post-hoc. Any failure can be assigned to a level after the fact. The framing earns its place only if it predicts failures before they are observed, and this article has mostly demonstrated the retrospective version.

The short version

Hand-engineered linguistic features stopped improving models, and reviews find no substantial gain from injecting parse trees, morphological features or explicit structure. The reason is not that the linguistics was wrong: probing established that models learn syntactic structure implicitly during ordinary training, so supplying it is redundant. Earlier work found that modelling structure explicitly did help, and the advantage narrowed to nothing as scale grew, which is the bitter lesson in its precise form.

The formal semantics tradition deserves a more accurate obituary. Its representations were logically sound and supported real inference with formal guarantees. It lost because every grammar rule had to be written by a linguist, which is a scaling constraint rather than a conceptual failure, and that distinction matters because a technology removing annotation cost changes the calculation.

Linguistics survived as the instrument rather than the input. It supplies the hypothesis space that makes probing possible, and the difference between observing that a model failed on some inputs and diagnosing that it failed on long-distance agreement is the difference between a bug report and a diagnosis.

The five traditional levels map onto distinct failure signatures. Phonology and orthography predict character-level failures, since tokenization discards that level. Morphology predicts degradation in morphologically rich languages. Syntax predicts long-distance dependency and structural generalisation failures. Semantics predicts the antonym collision and scope ambiguity. Pragmatics predicts literal interpretation and over-informativeness.

The practical payoff is a diagnostic vocabulary. A failure that looks like general unreliability usually resolves to one level once you know the levels exist, and which level it is tells you whether to change the input, the prompt, the architecture, or your expectations. Most evaluation tests only the semantic level, and a system nobody tested at a given level has an untested level rather than a working one.

Common questions

Does NLP still need linguistics? Not as a source of features, and yes as a source of categories. Attempts to inject linguistic structure into neural models have not produced substantial performance gains, because models learn that structure implicitly during ordinary training and supplying it is redundant. What survived is linguistics as an analytical instrument: it supplies the hypothesis space for probing what models represent, and the vocabulary for diagnosing what specifically failed rather than observing that something did.

Why did linguistic features stop improving models? Because the models already had the information. Probing work established that syntactic structure, dependency relationships and morphological features are recoverable from the internal representations of models trained only to predict text. You cannot improve a system by giving it something it already possesses. Notably, explicit structure did help in earlier work, and the advantage shrank to nothing as models and data grew.

Why did formal semantics and grammar-based NLP lose? Not because the representations were wrong. Formal approaches from Montague through categorial and unification grammars produced logically sound representations supporting real inference with formal guarantees, which nothing in current systems does. They lost because every grammar rule had to be written by a linguist, so the approach scaled with human labour while competitors scaled with compute.

What are the levels of linguistic analysis? Phonology and orthography, the level of sound and written form. Morphology, how words are built from meaningful parts. Syntax, how words combine into structures. Semantics, how structure determines meaning. Pragmatics, how context and intention determine what is communicated. Each maps onto a distinct class of AI failure, which is why the division is useful for diagnosis even where it stopped being useful for engineering.

How do the levels predict AI failures? Character-level tasks fail because tokenization discards the orthographic level. Morphologically rich languages degrade because subword segmentation fragments morphemes into pieces that are not morphemes. Long-distance dependency and structural generalisation fail at the syntactic level. Antonym collision and scope ambiguity are semantic. Literal interpretation of indirect requests and over-informativeness are pragmatic. A failure looking like general unreliability usually resolves to one of these.

Are large language models a theory of language? Disputed, and the disagreement is about what a theory is for. One position holds that models are better theories of language learning than linguistics produced, since they acquire language from exposure without innate grammatical machinery. The response is that a model reproducing a phenomenon without illuminating why has predictive power and no explanatory power, as a weather simulation predicts rain without being atmospheric physics. Interpretability research is the attempt to extract the second from the first.

What does linguistics know that machine learning has not absorbed? Four things. Morphemes are meaningful units and learned subword tokens are not, so segmentation boundaries fall where co-occurrence peaks rather than where meaning changes. Grammatical structure is hierarchical rather than sequential, and attention can represent this without being required to. Meaning is compositional while distributional similarity is topical, which are different properties. And languages differ in what they oblige speakers to encode, so a model trained predominantly on English learns English's obligatory distinctions and treats others as optional.

How should I test a system across linguistic levels? Vary each level independently rather than varying topic. Include morphologically complex inputs if you deploy in a language that has them. Include deep syntactic nesting and long-distance references. Include semantically equivalent expressions in different forms. Include indirect and implied requests alongside explicit ones. Most evaluation varies subject matter while holding linguistic complexity fixed, which measures one level and leaves the others untested.