Fake News Detection in Portuguese Under Large Language Model-Generated Content
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
We are daily exposed to fake news, a growing problem that spreads in various forms, including rumours, advertisements, social media posts, and political propaganda. Predominantly created by humans, in recent years, we have witnessed an increase of digital content fabricated or manipulated with the use of deep learning. Large Language Models (LLMs), for instance, represent a real threat if used to generate highly convincing fake news that could evade conventional detection systems. This study evaluates the impact of LLM-generated fake news on machine learning (ML) classifiers. The ML models are trained with Portuguese-language datasets and experiments are conducted using aligned data, where each fake news sample has its true news counterpart. We assess the performance of each ML model with synthetic fake news, which was generated using a Portuguese-based LLM, namely Sabiá-3. Our results reveal significant performance degradation of ML models when assessed under mismatch conditions, e.g., when they are trained with human-generated content, and tested with LLM-generated fake news (or vice-versa). These findings highlight the need for updated detection strategies capable of handling the linguistic and stylistic nuances introduced by LLMs. To address that, a Retrieval-Augmented Generation (RAG) framework was evaluated under the same conditions as the ML models. The framework showed to be more robust under mismatch conditions, whereas ML models provided better performance when there was no distribution shift between train and test data. These results contribute to the understanding of fake news detection in Portuguese, emphasizing the importance of adapting existing models to the evolving nature of misleading LLM-generated content.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it