💬 NLP

Emergence of a phonological bias in ChatGPT

This paper demonstrates that ChatGPT exhibits a human-like phonological bias by preferentially using consonants over vowels to identify words across different languages, suggesting that such cognitive traits can emerge in large language models despite their distinct training mechanisms compared to human language acquisition.

Original authors: Juan Manuel Toro

Published 2026-02-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Juan Manuel Toro

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Secret Sound Preference of ChatGPT

Imagine you are teaching a robot to read a book. You wouldn't expect the robot to suddenly develop a "favorite" way of hearing words, right? You'd think it would just process every letter equally, like a super-organized librarian sorting books by the alphabet.

But according to this new research by Juan Manuel Toro, ChatGPT has a secret habit. Just like human babies and adults, the AI has developed a strong preference for consonants (like b, t, k, s) over vowels (like a, e, i, o, u) when it tries to figure out what a word is.

Here is the simple breakdown of what happened, using some everyday analogies.

1. The "Skeleton vs. The Flesh" Analogy

Think of a word like a human body.

Consonants are the skeleton. They give the word its shape and structure.
Vowels are the flesh and skin. They fill in the gaps and give the word its specific "flavor" or sound.

If you take a skeleton and swap out the flesh, the body still looks like the same person. But if you swap out the bones, it becomes a completely different creature.

The study found that ChatGPT acts like a detective who looks at the skeleton first. If you show the AI the word "Natural" and ask, "Is this more similar to 'Nalural' (vowel changed) or 'Nalural' (consonant changed)?" the AI almost always picks the one where the consonants stayed the same.

2. The Experiment: A Game of "Spot the Difference"

The researcher played a simple game with ChatGPT 100 times in English and 100 times in Spanish.

The Setup: He gave the AI a real word (like "Zebra").
The Trick: He offered two fake words:
- Option A: "Zebra" (Changed the vowel: Zebra → Zibra)
- Option B: "Zebra" (Changed the consonant: Zebra → Zrbra)
The Question: "Which one looks more like the original?"

The Result: ChatGPT chose the version with the changed vowel (keeping the consonants) 76% of the time in English and 74% of the time in Spanish.

It didn't matter if the word was short or long, or if it was a noun or a verb. The AI consistently ignored the vowels and focused on the consonants.

3. Why Does This Matter? (The "Magic" of Emergence)

This is the most fascinating part. Humans have this bias because our brains are wired that way. We learn that consonants are the "anchors" of words, while vowels can change based on how we feel, where we are from, or how fast we are talking.

But ChatGPT isn't a human. It's a massive math machine trained on terabytes of text. It wasn't programmed to like consonants. No one told it, "Hey, ignore the vowels!"

The Analogy: Imagine you throw a million puzzle pieces into a giant box and shake them. You don't tell the box how to solve the puzzle. Yet, somehow, the pieces start to fit together in a way that looks exactly like a human picture.

This suggests that consonant bias is a natural law of language, not just a human quirk. When you train a machine on enough human language, it naturally discovers that consonants are the most important clues for identifying words.

4. The Takeaway

This paper tells us two big things:

AI is getting eerily human-like: Even though it learns differently than a baby, it ends up with the same "instincts" about how language works.
Language has a hidden structure: The fact that both humans and AI prioritize consonants suggests that this is the most efficient way for any intelligent system to process words. It's like the "operating system" of language itself.

In short: ChatGPT isn't just mimicking us; it's discovering the same deep secrets about how we speak that we have known for centuries. It's a robot that learned to listen to the skeleton of the word, just like we do.

1. Problem Statement

The paper investigates whether Large Language Models (LLMs), specifically OpenAI's ChatGPT, exhibit phonological processing biases similar to those observed in human language acquisition and processing.

The Human Context: Humans display a well-documented "consonant bias," where they rely more heavily on consonants than vowels for lexical access (word identification). This bias is observed across ages, languages (e.g., English, Spanish, French), and modalities. The prevailing hypothesis is that consonants carry more lexical information (disambiguating words), while vowels carry more indexical information (speaker identity, intonation).
The Research Question: Do current AI models, which are trained on statistical patterns of text rather than biological cognitive development, spontaneously develop this same consonant bias, or is it an artifact of human-designed architecture?

2. Methodology

The study employed a prompt-based behavioral experiment to test ChatGPT's word similarity judgments in two languages: English and Spanish.

Experimental Design:
- Stimuli: 100 target words were selected for each language (totaling 200 trials per language). The words varied in length (2–4 syllables), syllabic structure (CV, CVC, CCV, VC), and syntactic category (nouns, verbs, adjectives).
- Non-word Generation: For each target word, two "non-words" (pseudo-words) were generated by altering exactly one phoneme:
  1. Consonant Change (C-change): The consonant tier was altered, preserving the vowel tier (e.g., natural $\rightarrow$ nalural).
  2. Vowel Change (V-change): The vowel tier was altered, preserving the consonant tier (e.g., natural $\rightarrow$ nateral).
- Control Variables: The position of the change (onset, nucleus, coda) and the syllable affected were balanced. The order of presentation was randomized to prevent position bias.
Procedure:
- English Prompt: "If you have to choose, which of these two non-words is more similar to [target word], [non-word 1] or [non-word 2]?"
- Spanish Prompt: "Si tienes que elegir, cual de estas dos no-palabras se parece más a [target word], [non-word 1] o [non-word 2]?"
- Execution: Prompts were submitted one by one without feedback to prevent the model from learning from the interaction session. No automated scripts were used for presentation to ensure natural interaction.
Statistical Analysis: The frequency of choices favoring the C-change (preserving vowels) vs. V-change (preserving consonants) was analyzed using a Wilcoxon signed-rank test.

3. Key Results

The study found a robust consonant bias in ChatGPT across both languages.

English Results:
- ChatGPT chose the non-word preserving the consonant tier (C-tier) in 76% of the trials.
- It chose the non-word preserving the vowel tier (V-tier) in only 24% of the trials.
- Statistical Significance: $Z = 1212$ , $p < 0.005$ .
Spanish Results:
- ChatGPT chose the C-tier preservation in 74% of the trials.
- It chose the V-tier preservation in 26% of the trials.
- Statistical Significance: $Z = 1313$ , $p < 0.005$ .
Consistency: The bias was independent of word length, syllabic structure, the position of the letter change, or the syntactic category of the word.

4. Key Contributions

Emergent Behavior in AI: The paper demonstrates that a complex cognitive bias (consonant bias), previously thought to be a hallmark of human biological language processing, emerges spontaneously in LLMs trained solely on statistical text prediction.
Cross-Linguistic Generalization: The bias was observed in both English and Spanish, despite these languages having different phonological distributions (English has a higher vowel-to-consonant ratio than Spanish). This suggests the bias is not merely a reflection of specific language statistics but a general property of the model's processing.
Parallels with Human Cognition: The findings draw a direct parallel between artificial and natural intelligence, suggesting that the "division of labor" between consonants (lexical) and vowels (indexical) may be a fundamental statistical necessity for efficient word segmentation and identification, regardless of the substrate (biological brain vs. neural network).
Methodological Framework: The study provides a replicable prompt-based framework for probing phonological processing biases in generative AI without requiring access to the model's internal weights or training data.

5. Significance and Implications

Understanding LLMs: The results challenge the notion that LLMs are merely "stochastic parrots" that mimic surface patterns. Instead, they suggest that deep statistical learning on natural language corpora leads to the emergence of structural linguistic principles similar to those found in human cognition.
Nature of Language Learning: The study supports the hypothesis that the consonant bias is an emergent solution to the problem of lexical disambiguation. Since natural languages generally contain more consonants than vowels, focusing on consonants provides a more efficient signal for distinguishing word identities.
AI-Human Convergence: The convergence of processing strategies between ChatGPT and human infants/adults implies that certain cognitive architectures (whether biological or artificial) may inevitably converge on similar solutions when processing the same type of data (natural language).
Future Research: This opens avenues for investigating other human cognitive biases (e.g., syntactic hierarchies, semantic priming) in AI to determine which aspects of human intelligence are emergent properties of data scale and which require specific biological or architectural constraints.

Conclusion:
Toro's paper provides compelling evidence that ChatGPT relies more heavily on consonants than vowels to identify words, mirroring a fundamental bias in human language processing. This suggests that the consonant bias is not an innate biological trait exclusive to humans but a statistical property that emerges naturally when a system learns to process language efficiently.