Is an AI Model an Expert or a Sophisticated Interpolator? The Answer Matters

A doctor who's never seen your rare disease can still diagnose it from symptoms. They can identify a pattern beyond their direct experience. An interpolator would guess it statistically from similar known cases — and often get it wrong.

When you query an AI model, you get an answer. But how did the model arrive at it? Did it apply rules it extracted from training data and generalized beyond them? Or did it statistically interpolate between similar examples it saw during training?

For most questions, you don't know. The answer looks equally fluent whether the model truly understands the problem or just plausibly predicts the next words according to patterns from training.

This question isn't academic. It determines which types of AI answers you can trust — and which require verification. An expert generalizes. When a medical AI model sees a new combination of symptoms, it extracts a principle from previous cases and applies it to this specific case. An interpolator predicts. When it sees a new combination of symptoms, it finds the most probable diagnosis according to statistical similarity with seen cases.

The problem: current LLM models do both. Some types of questions are solved by generalization — extrapolation of abstract rules beyond seen data. Others by interpolation — statistical prediction between seen examples. And because generalization and interpolation look identical from outside (fluent answer to a question), the user has no way to recognize which mechanism the model used.

This article analyzes the fundamental difference between expertise (ability to generalize beyond training distribution) and interpolation (sophisticated guessing between seen examples). It argues that most LLM use relies on interpolation, not expertise — and that consequences of this distinction determine how to integrate AI tools into high-stakes decision-making.

What Is Generalization and Why It Goes Beyond Training Data

Generalization means extracting an abstract rule from a finite number of examples and applying it to unseen instances. This enables solving problems that weren't in the training data.

Generalization is the ability to extract a principle that applies more broadly than the examples from which it was derived. If a child sees ten trees and learns the concept "tree," they can identify the eleventh tree even if it looks different from all ten previous ones. That's generalization — the abstract rule "what makes a tree a tree" transcends concrete seen instances.

Mathematical example: If someone teaches you arithmetic with examples 2 + 3 = 5, 7 + 1 = 8, 10 + 4 = 14, you can derive the rule of addition and apply it to 38 + 127. You've never seen this specific example, but you understand the principle. This is generalization beyond training distribution — the area that seen examples cover.

A medical example shows the power of generalization. A doctor sees a patient with a combination of symptoms they've never seen before. But they understand pathophysiology — how symptoms arise from underlying mechanisms. They can diagnose even when the exact combination wasn't in their experience, because they work with an abstract model of disease, not a database of cases.

If an AI model truly generalizes, it can solve new types of problems beyond training data — extrapolation. If it only interpolates, it can solve only variations on seen problems — statistical prediction between known points. The difference is fundamental.

Where do LLMs actually generalize? Zero-shot reasoning on new types of tasks — GPT-3 handled several tasks it never saw during training, without a single example. Abstract analogies — transferring a principle from one domain to another ("How to apply a medical principle to business?"). Chain of Thought reasoning — decomposing a new problem into subproblems using rules the model extracted from training.

But most LLM answers aren't based on generalization. They're based on interpolation.

What Is Interpolation and Why Most AI Answers Are Based on It

Interpolation is statistical prediction between seen data points. Sophisticated, but fundamentally limited to the space training data covers.

Interpolation means predicting a value between known points using statistical patterns. If someone teaches you 2 + 2 = 4, 3 + 3 = 6, 5 + 5 = 10, you can guess 4 + 4 = 8 by interpolating between seen examples. No need to understand the rule of addition — just recognize the pattern "same number + same number = double."

LLMs are trained on next token prediction — predicting the next word in a sequence. Good prediction means finding statistical patterns in training data. If the model learns from millions of sentences like "France is a country in Europe," "Germany is a country in Europe," it learns the pattern "X is a country in Europe" and can predict "Italy is a country in Europe." This looks like geography knowledge, but it's pattern interpolation.

A concrete example illustrates the mechanism. Query "What is the capital of France?" GPT-4 answers "Paris." It looks like the model "knows" geography. But the mechanism is statistical: sentences like "Paris, the capital of France" appeared in training data thousands of times. The model doesn't see a map, doesn't understand the concept "capital," just predicts the most probable token after "the capital of France is."

Interpolation works great as long as the question lies in the space training data covers. When you ask about common geography ("capital of France"), interpolation suffices — the answer lies in a densely covered part of training distribution. When you ask about an edge case outside coverage (rare disease, unusual legal precedent, counterintuitive fact), the model interpolates plausibly — but often incorrectly.

Why are most answers interpolation? Training distribution covers most common questions densely. Millions of examples for similar questions. The model learns the statistical relationship between common inputs and outputs. This suffices for many everyday queries. But once the question lies outside densely covered areas, the model interpolates outside its training region and hallucinates.

It looks like knowledge, but it's pattern matching. And pattern matching fails when the pattern wasn't in the data.

Why They Look Identical from Outside — And How to Test

A fluent, authoritative answer is compatible with both generalization and interpolation. Users need diagnostic tests to distinguish which mechanism the model used.

When a model answers a question, the output is text. Fluent, grammatically perfect, stylistically coherent. This output doesn't look different if the model generalized (applied an abstract rule) or interpolated (predicted statistically). Both mechanisms produce identically-looking text. No metadata says "this answer was interpolation."

A concrete test reveals the difference: out-of-distribution question. If you ask "How much is 38 + 127?" (arithmetic beyond small numbers often seen in training), the model either correctly applies the rule of addition (generalization) or guesses a probable number according to patterns (interpolation). The answer "165" looks equally fluent in both cases.

But if you ask "How much is 3847 + 12938?" (larger numbers outside common distribution), the model might give a wrong answer. Signal that it interpolated instead of generalizing. If the model truly understands arithmetic (generalizes the rule), number size is irrelevant. If it only interpolates patterns from seen examples, larger numbers are outside its coverage.

An adversarial example shows the diagnostic procedure. Query "What is the capital of Hungary?" → Model answers "Budapest" (correctly). This could be generalization (understands the concept of capital) or interpolation (saw the sentence "Budapest is the capital of Hungary" in data).

Diagnostic test: "What is the second largest city in Hungary?" If the model answers correctly (Debrecen), it probably has broader knowledge of Hungarian geography — generalizes the concept of Hungarian cities beyond the capital. If it hallucinates (invents a plausible Hungarian city name), it probably interpolated the first answer without true understanding.

Practical test for any expert question: If an AI model answers, ask two follow-up questions. (1) "Why?" — test understanding of mechanism. (2) "What would happen if X were different?" — test reasoning ability beyond the first answer. An expert generalizes on both. An interpolator fails on at least one.

A model that says "Paris is the capital of France" and then explains "because it's the administrative center where the government resides" probably understands the concept of capital. A model that says "because it's the most famous city in France" interpolated — used plausible but mechanically incorrect reasoning.

Which Types of Questions Require Generalization — And Which Need Only Interpolation

Not all questions require generalization. Some fall into densely covered parts of training distribution — interpolation suffices and works reliably. A taxonomy of questions determines when to trust AI output.

Three categories of questions exist based on coverage and need for generalization.

Category A — Densely Covered, Interpolation Suffices

Factual questions on mainstream knowledge. "Who wrote Hamlet?", "What is the formula for circle area?", "What is the capital of France?" Training data contains thousands of examples of these questions and answers. The model interpolates between seen instances — and this suffices because the correct answer lies in a densely covered area.

Reliability: often high. If you ask multiple different models, they'll frequently answer identically and correctly. Reason: all interpolate in the same densely covered part of distribution.

Practical use: General factual queries, common definitions, mainstream tutorials. One model suffices, verification not critical.

Category B — Sparsely Covered, Interpolation Risky

Specialized domain knowledge. "What is the differential diagnosis for this rare disease?", "What precedent exists for this unusual legal case?", "How to interpret this edge case in tax law?"

Training data contains few or no examples. The model interpolates plausibly but outside a region with sufficient coverage — hallucination probable. The answer sounds authoritative but may be statistically guessed rather than factually correct.

Reliability: often low. If you ask multiple different models, you'll frequently get different answers — each interpolates differently because coverage is sparse.

Practical use: Domain-specific questions, rare cases, specialized knowledge. Two or three independent models, compare answers. If they disagree, verify through external source.

Category C — Require Reasoning Beyond Seen Examples

Multi-step reasoning where subgoals aren't in training as a whole. "Combine concept A from medicine with principle B from statistics and apply to a new case", "Design a solution for a problem with elements from three different domains."

Requires generalization — extraction of abstract rules and their composition. Interpolation fails because the entire reasoning path isn't in data. The model must abstract a principle from one domain and apply it to another. If it only interpolates, it generates plausible but logically flawed reasoning.

Reliability: highly variable, depends on model's ability to generalize. Some models have better reasoning (Claude 3.5 has better Chain of Thought than GPT-3.5). But even the best models fail on truly novel combinations.

Practical use: Complex analytical tasks, new problems requiring synthesis from multiple domains. Trust only if you can verify reasoning steps. Expert review mandatory.

A practical example illustrates the category. "How to treat diabetes?" (Category A, reliable — interpolation in densely covered area). "How to treat combination of diabetes + this rare autoimmune disease?" (Category B, verify — interpolation in sparsely covered area). "Design a new therapeutic protocol for a patient with unique combination of symptoms" (Category C, requires generalization — expert review mandatory).

The Scaling Paradox — Larger Models Interpolate Better, Don't Generalize More

Scaling (more parameters, more data) improves interpolation dramatically but generalization only mildly. That's why larger models are more reliable on common questions but still fail on edge cases.

Intuition says: larger model = better understanding = more generalization. Reality: larger model = better coverage of training distribution = better interpolation. Generalization scales slower than interpolation.

GPT-3 (175B parameters) vs. GPT-2 (1.5B parameters) illustrates this effect. GPT-3 is significantly more accurate on common questions — has greater capacity to memorize more patterns, cover more training distribution. Can interpolate in a denser network of seen examples. But on out-of-distribution questions (reasoning beyond seen examples, counterintuitive facts) the difference is smaller.

Reason: GPT-3 interpolates in a denser network of points but still interpolates. Generalization — ability to derive abstract rules — requires a different mechanism than mere parameter scaling. More parameters mean greater memorization capacity, not necessarily better abstraction.

A concrete benchmark shows the scaling limit. GSM8K (grade school math word problems) highlights large performance gaps between model generations and training recipes. It can look like "generalization" — but on many tasks the model is still best described as a powerful interpolator with sharp failure modes outside densely covered regions.

But when tested on adversarial variants (change numbers, change context, same underlying principle), accuracy drops dramatically. The model learned to interpolate between seen math patterns ("Johnny has X apples, gives Y to a friend, how many left?"), not mathematics as an abstract system. Change "apples" to "cars" and "friend" to "sibling" — same principle, different surface pattern — and the model fails more often.

Scaling is progress. But it's not a path to AGI or to a "model that understands." It's a path to a better interpolator — covers more training distribution, makes fewer errors on common questions. But edge cases requiring generalization beyond seen data remain problematic even in the largest models.

Why does this matter? If you rely on AI in professional contexts (legal, medical, business), you typically solve edge cases — not common questions (you know those or can easily look them up). Precisely where you need generalization, scaling doesn't help enough. GPT-5 will be a better interpolator than GPT-4. But still an interpolator.

Consequences for High-Stakes AI Use

If AI primarily interpolates rather than generalizes, high-stakes decisions require mechanisms that compensate for interpolation limitations. Model diversification, external verification, human-in-the-loop.

An interpolator is a useful tool if you know it's an interpolator. The problem arises when it's treated as an expert — you trust it in contexts where interpolation fails.

Three strategies for working with AI as an interpolator, not an expert.

Strategy 1 — Model Diversification

Different models had different training data → interpolate in different regions. GPT-4 had a different corpus than Claude 3.5. Gemini has access to different sources than both predecessors. Their coverage doesn't overlap completely.

If two models disagree, at least one probably interpolates outside its covered region. One has the correct answer in data, the other doesn't — interpolates into an area where it lacks sufficient coverage. Agreement increases confidence (both interpolate correctly or both generalized). Disagreement is a warning signal.

Practically: For medium and high stakes questions, use at least two independent models. If they agree, probably safe. If they disagree, investigate why — often you'll discover the question lies in an edge case where one model lacks coverage.

Strategy 2 — External Verification

If a model interpolates correctly, its answer should be verifiable through a primary source (document, database, expert). If it interpolates incorrectly, verification will reveal it.

Requesting citations forces the model to structure the answer around verifiable claims. Even when the model hallucinates citations (and often does), forcing citations structures output in a way that's easier to fact-check than free text without references.

Practically: For high stakes questions, demand "provide sources for each claim." The model either returns correct sources (interpolated correctly in covered region) or fake citations (interpolated incorrectly outside coverage). The latter is easier to detect than general statements without citations.

Strategy 3 — Human Expert Review

For high-stakes categories (legal, medical, financial), AI output goes through expert review. The expert doesn't test every detail but tests reasoning — whether the AI answer makes sense from an underlying principle perspective.

An interpolator can generate a plausible but mechanically wrong answer. Sounds credible but violates a principle that isn't visible in surface patterns. An expert catches this because they work with an abstract model, not a database of seen cases.

Practically: AI generates a draft answer. Expert reviews reasoning. If reasoning makes sense, proceed. If not, AI probably interpolated outside coverage — discard and seek a different approach.

Platforms like CrossChat implement Strategy 1 natively. Multi-model workflow automatically diversifies across different interpolation spaces (GPT-4 + Claude + Gemini have different data). Consensus score is a metric signaling whether models interpolate in agreement (high score) or one of them probably interpolates outside coverage (low score).

Practical Conclusion

1. Categorize questions before using AI. Densely covered mainstream knowledge → interpolation reliable, one model suffices. Sparsely covered domain knowledge → interpolation risky, verify through second model or external source. Multi-step reasoning beyond seen examples → requires generalization, trust only with expert review.

2. Test whether model generalizes or interpolates. Ask follow-up questions: "Why?" (test understanding of mechanism), "What would happen if X were different?" (test reasoning beyond first answer). An expert generalizes on both. An interpolator fails because the reasoning path isn't in its data.

3. Scaling isn't a solution for edge cases. GPT-5 will be more accurate than GPT-4 on common questions (better interpolation), but edge cases requiring generalization will remain problematic. A larger model covers more training distribution but still interpolates. Don't expect a bigger model to solve the out-of-distribution reasoning problem.

4. Treat AI as an interpolator, not an expert. Draft proposals, draft hypotheses, draft diagnoses — and verify. The model is a tool for generating candidates, not an arbiter of truth. High-stakes decisions require human expert review or multi-model cross-check. An interpolator is useful, but it's not an expert.

Sources

Zhang, C. et al. (2017). Understanding Deep Learning Requires Rethinking Generalization. ICLR 2017. — Classic study on generalization vs. memorization in neural networks.
Geirhos, R. et al. (2020). Shortcut Learning in Deep Neural Networks. Nature Machine Intelligence, 2(11). — Why models interpolate instead of generalize: learn shortcuts rather than underlying principles.
Cobbe, K. et al. (2021). Training Verifiers to Solve Math Word Problems. arXiv:2110.14168. — GSM8K benchmark as a test of mathematical reasoning and generalization.
Hendrycks, D. & Gimpel, K. (2017). A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks. ICLR 2017. — How to detect when a model is outside training distribution.
Chollet, F. (2019). On the Measure of Intelligence. arXiv:1911.01547. — Argument that abstract reasoning (generalization) is the core of intelligence, not pattern matching.

Editorial History

Concept: Claude Code + Anthropic Sonnet 4.6 Version 1: Claude Code + Anthropic Sonnet 4.6 Version 2: Codex + GPT-5.2