Fakes of Varying Shades: How Warning Affects Human Perception and Engagement Regarding LLM Hallucinations

🌈 Abstract

The article investigates how untrained human evaluators perceive the accuracy of and engage with (like, dislike, share) large language model (LLM)-generated content with varying degrees of hallucination (genuine, minor hallucination, major hallucination). It also examines the impact of warning on human perceptions of LLM-generated genuine and hallucinated content.

🙋 Q&A

[01] Fakes of Varying Shades: How Warning Affects Human Perception and Engagement Regarding LLM Hallucinations

1. How do untrained human evaluators perceive the accuracy of LLM-generated genuine and hallucinated content?

• Participants ranked content as truthful in the order: genuine > minor hallucination > major hallucination.

2. How does the perceived accuracy differ depending on (a) the varying degree of hallucination and (b) the presence of warning?

• Warning decreased the perceived accuracy of minor and major hallucinations, but did not significantly affect the perception of genuine content.

3. How do untrained human evaluators engage with (i.e., like, dislike, share) LLM-generated genuine and hallucinated content?

• 'Likes' and 'shares' mirrored the pattern of perceived accuracy, with 'dislikes' following the reverse order.

4. How does the engagement differ depending on (a) the varying degree of hallucination and (b) the presence of warning?

• Warning increased 'dislikes' but had negligible effects on 'likes' and 'shares'.

[02] Related Work

1. What is the difference between misinformation, disinformation, and hallucination?

• Misinformation encompasses all inaccurate information, spread with or without intent. Disinformation refers to false information spread with the intent to deceive. Hallucinations can transition into either misinformation or disinformation depending on presentation and intent.

2. What are the current approaches to addressing LLM hallucination?

• Recent studies focus on creating hallucination benchmarks, evaluating hallucinated texts, and automatically detecting hallucinations. Human evaluation remains a primary method for assessing hallucinations.

3. What do existing studies reveal about human perception of LLM-generated texts?

• Prior research indicates that humans struggle to distinguish machine-generated texts from human-written ones, often performing at or below chance levels.

4. How have studies investigated the effect of warning on misinformation perception?

• Warnings can reduce the lasting impact of misinformation, but may also lead to blind skepticism, reducing trust in authentic news.

[03] Methodology

1. How were the genuine and hallucinated responses generated?

• Genuine responses were generated by directly asking GPT-3.5-Turbo to answer questions from the TruthfulQA dataset. Minor and major hallucinations were generated using prompt engineering techniques that incorporated subtle and substantial fabrications, respectively.

2. What was the experimental design of the human-subjects study?

• The study used a 2 (between-subjects factor: control vs. warning) x 3 (within-subjects factor: genuine vs. minor vs. major hallucinations) mixed-design experiment with Prolific participants.

[04] Results

1. What was the key finding regarding the impact of warning on human perception?

• Warning decreased the perceived accuracy of minor and major hallucinations, but did not significantly affect the perception of genuine content.

2. How did humans perceive the accuracy of genuine, minor, and major hallucinations?

• Participants ranked content as truthful in the order: genuine > minor hallucination > major hallucination.

3. How did user engagement (like, dislike, share) differ across the hallucination levels?

• 'Likes' and 'shares' mirrored the pattern of perceived accuracy, with 'dislikes' following the reverse order.

4. How did warning affect user engagement?

• Warning increased 'dislikes' but had negligible effects on 'likes' and 'shares'.

[05] Discussion

1. What are the key insights regarding the use of warning to enhance hallucination detection?

• Warnings show promise for improving hallucination detection without significantly affecting the perceived truthfulness of genuine content.

2. What are the implications of the finding that humans are more susceptible to minor hallucinations compared to major hallucinations?

• This suggests the need for further research to understand the specific characteristics of minor hallucinations that make them more believable, as well as the development of tailored interventions.

3. How do the findings on user engagement relate to the potential use of RLHF for improving LLM performance?

• The patterns of user engagement, such as liking and disliking, could provide valuable feedback for RLHF algorithms to improve LLM performance in generating truthful and trustworthy content.

</output_format>