LINGUIST 316: Hallucinations

Hallucination

occurs when an LLM generates information that is inaccurate, misleading, or entirely fabricates rather than fully correct

Why are hallucinations a major concern?

They reduce the reliability and trustworthiness of LLM outputs, especially in high-stakes domains such as medicine, law, education, and research.

Factuality

Refers to whether the generated info is true and corresponds to reality

Faithfulness

Refers to whether the generated output accurately reflects its source material without adding unsupported info

How can an output be faithful but not factual?

If it accurately reproduces information from a source that itself contains incorrect information

How can an output be factual but not faithful?

If it includes information that was not present in the original source

Do hallucinations occur only in text generation?

No. Hallucinations occur across multiple modalities, including:

• Text (Q&A, dialogue, summarisation)

• Images (e.g., six fingers)

• Video

• Audio transcription

Why are hallucinations common in generative AI?

Generative models predict likely outputs rather than retrieving verified facts.

Why can't LLMs simply memorise all information?

The amount of information is too large, so models compress patterns and relationships rather than storing exact facts.

What is the primary cause of hallucination?

The model generates text based on statistical prediction rather than direct access to truth.

Why does incomplete information increase hallucinations?

The model must interpolate between known patterns, increasing the chance of generating unsupported details.

How can training data contribute to hallucinations?

Training datasets may contain inaccuracies, inconsistencies, gaps, or conflicting information.

How can bias contribute to hallucinations?

Biased training data can encourage the model to generate stereotypical or unsupported information.

Why doesn't an LLM simply admit uncertainty?

Fine-tuning often rewards producing an answer rather than admitting a lack of knowledge

Intrinsic hallucination

A contradiction between the source content and the generated output

Give an example of intrinsic hallucination.

A summarisation system reports the wrong date from an article it was asked to summarise.

Extrinsic hallucinaion

Information generated by the model that cannot be verified from the source material.

Give an example of extrinsic hallucination.

A summary introduces an event or fact that was never mentioned in the original text.

Which type of hallucination is harder to detect?

Extrinsic hallucinations, because they may sound plausible despite lacking source support.

Why are multilingual LLMs especially vulnerable to hallucinations?

Translation and cross-lingual generation introduce additional opportunities for meaning distortion and unsupported inferences.

Why can translation tasks produce hallucinations?

The model may infer missing information or incorrectly transfer meaning between languages.

Why can hallucinations be understood through pragmatics?

LLMs learn how language is used in context rather than learning what is objectively true.

Why does fluency not equal truth?

A response can sound coherent, relevant, and natural while still being factually incorrect.

What are LLMs primarily trained on?

Patterns of language use and communication in context.

What are LLMs not directly trained on?

A verified model of truth or real-world knowledge.

What is pragmatic alignment?

Producing responses that sound appropriate, coherent, and contextually relevant

What is epistemic alignment?

Producing responses that are factually correct and grounded in reality.

How can hallucinations be defined using pragmatic and epistemic alignment

Hallucinations are the gap between pragmatic alignment (sounding correct) and epistemic alignment (being correct).

What are Grice's four conversational maxims?

• Quantity (give enough information)

• Quality (be truthful)

• Relation (be relevant)

• Manner (be clear)

Which Gricean maxim do hallucinations most often violate?

The Maxim of Quality (truthfulness).

How can a hallucination satisfy the other maxims?

It may still appear informative, relevant, and clear despite being false.

How can hallucination be described in Gricean terms?

A pragmatic failure of the Quality Maxim under pressure to satisfy Quantity, Relation, and Manner

What is conversational implicature?

The process of inferring meanings beyond what is explicitly stated.

How do LLMs use implicature?

They learn statistical patterns about what is typically implied in conversation.

Why can implicature lead to hallucinations?

Models may treat common inferences and statistical regularities as established facts.

Are hallucinations random mistakes?

No. They are often plausible overextensions of learned language patterns.

What is common ground in pragmatics?

The set of facts assumed to be mutually known by conversational participants.

Why do LLMs struggle with common ground?

: They lack:

• Real-time world knowledge

• A stable world model

• Reliable distinction between fact and plausibility

How can failures of common ground produce hallucinations?

The model assumes shared knowledge that does not actually exist or fabricates it to maintain coherence.

Why can subtle hallucinations be more dangerous than obvious ones?

They may go unnoticed while appearing highly credible.

How can training data misalignment create hallucinations?

The target text may contain information not present in the source data.

How can hallucinations be reduced?

By improving dataset quality and faithfulness.

What does creating a faithful dataset involve?

Ensuring outputs accurately reflect source content without unsupported additions.

Why is dataset cleaning important?

It removes noise that interferes with accurate learning.

What is lossy compression?

Compression that removes information, producing approximations rather than perfect copies.

Why are hallucinations expected according to Chiang?

A: LLMs reconstruct patterns from compressed representations rather than retrieving

LLMs reconstruct patterns from compressed representations rather than retrieving exact facts.