Fake News Occurrence

---
name: fake-news-occurrence
description: Assesses whether multilingual LLM responses contain misinformation or fake content when queried on specific topics. The evaluation relies on an automated GPT-based judge to determine the presence of fake information in model generations. Use when the user has predictions and gold and needs to compute fake news occurrence.
metadata:
  skill_kind: metric
  source_arxiv: 2406.13748
  bibtex_key: lu2024learnunlearn
  confidence: low
---

# fake-news-occurrence

> Learn and Unlearn: Addressing Misinformation in Multilingual LLMs — Lu et al. (2024) (arXiv:2406.13748, 2024)

## What this evaluates

Assesses whether multilingual LLM responses contain misinformation or fake content when queried on specific topics. The evaluation relies on an automated GPT-based judge to determine the presence of fake information in model generations.

## Datasets

- **Fake News Dataset (from Figure 2)** — total ?; splits: test (-1)

## Metrics

- `fake news occurrence` **(primary)** — range: binary
  - Binary judgment indicating whether fake information exists in a model's generated response. Primarily evaluated using GPT as an automated judge, with human evaluation used as a validation subset.

## Input / output format

**Input**: Model-generated responses to questions about fake news, paired with the original question and model language.

**Output**: GPT-based binary judgment indicating whether fake information exists in the response.

## Scoring recipe

```python
def evaluate_fake_news(model_responses, pairs):
    gpt_judgments = {p: gpt_check_for_fake_info(r) for p, r in zip(pairs, model_responses)}
    sample = random.sample(list(gpt_judgments.keys()), 10)
    human_judgments = {p: human_check_for_fake_info(gpt_judgments[p]) for p in sample}
    discrepancies = sum(1 for p in sample if gpt_judgments[p] != human_judgments[p])
    if discrepancies == 0:  # or no statistical difference
        return gpt_judgments
    return human_judgments
```

## Common pitfalls

- GPT-based evaluation may introduce systematic bias if not validated against human judgment.
- Human validation relies on translation tools, which may affect accuracy for non-English languages.
- Small validation sample size (10 per question-model language pair) may not capture all edge cases.

## Evidence (verbatim from paper)

> To verify that the evaluation by GPT is not the source of our results, for each question-model language pair in the trained model’s responses on fake news from [Figure 2], we randomly selected 10 data points for human evaluation. Human evaluators reviewed model generations and check if fake information exists, with help of translation tools and without knowing GPT’s judgment. Number of discrepancies between the human evaluations and GPT’s evaluations is counted. As in [Figure 6], there was no statistical difference between the human and GPT judgments in any language, we concluded that GPT provides a reliable evaluation for our purpose.

## Citation

```bibtex
@misc{lu2024learnunlearn,
  title={Learn and Unlearn: Addressing Misinformation in Multilingual LLMs},
  author={Lu et al. (2024)},
  year={2024},
  note={arXiv:2406.13748}
}
```

- arXiv: 2406.13748