Trust Oriented Explainable AI for Fake News Detection

arXiv:2603.11778 · cs.CL · Submitted 2026-03-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Trust Oriented Explainable AI for Fake News Detection".

Jane: Here's the gist of the paper. This article examines the application of Explainable Artificial Intelligence (XAI) in NLP based fake news detection and compares selected interpretability methods to enhance model transparency and trustworthiness The results show that XAI enhances model transparency and interpretability while maintaining high detection accuracy Model Comparison of Interpretability Techniques The work compares SHAP, LIME, and Integrated Gradients across two neural network architectures to determine their explanatory value Each method provides distinct explanatory value: SHAP offers detailed local attributions, LIME provides simple and intuitive explanations, and Integrated Gradients performs efficiently with convolutional models The study also highlights limitations such as computational cost and sensitivity to parameterization Overall, the findings demonstrate that integrating XAI with NLP is an effective approach to improving the reliability and trustworthiness of fake news detection systems Neural Network Architectures in NLP Natural Language Processing (NLP) uses various neural architectures including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and attention based models such as transformers Recurrent Neural Networks (RNNs) introduced feedback loops that allow information from previous time steps to influence future computations, enabling the preservation of context in longer sequences Long Short Term Memory (LSTM) architecture utilizes a cell state and three distinct gates to manage information through sigmoid and tanh activation functions The Transformer architecture is characterized by a symmetric stack of N identical layers, designed to optimize both contextual representation and sequential generation Implementation Details of XAI Methods The implementation of XAI methods involved integrating the explanation layer with the existing input data representation, where sequences of token identifiers were used For SHAP, a processing procedure was established where a text masker generates multiple versions of a sentence with hidden words and then SHAP asks the model to evaluate these versions LIME creates multiple slightly modified versions of the sentence close to the original and trains a simple surrogate model describing the local behavior of the classifier Integrated Gradients is implemented as a gradient method operating directly on the embedding representation, transitioning from an ”empty” text to the actual input Performance Metrics and Results The classification quality was assessed using several metrics, including completeness, sufficiency, AOPC (emphArea Over the Perturbation Curve), and Flip@k For LSTM, SHAP delivers the best results with the strongest removal effect and highest AOPC For CNNs, Integrated Gradients is the best method, combining the highest completeness and AOPC with the lowest Flip@k The results also show that explanation quality depends strongly on model architecture, with SHAP working best for LSTMs while Integrated Gradients outperforms others for CNNs Limitations and Future Directions The study identifies limitations associated with XAI, including computational overhead, sensitivity to parameterization, and the risk of misinterpretation by end users A primary risk is the illusion of causality, where explanations often reflect correlations learned from training data rather than true domain level causes Explanations are also susceptible to manipulation, as models can be intentionally modified to generate misleading explanations while preserving biased or undesirable decision rules The findings offer a foundation for future work involving transformer based architectures, rule based explanations, end user studies, model calibration, and linguistically grounded perturbation protocols Data Preparation and Model Training The experiments were conducted using the publicly available ISOT Fake News Dataset Data preprocessing involved building a dictionary limited to the 20,000 most frequently occurring words and standardizing sequence length to a maximum of 750 tokens Two neural network architectures were evaluated: a sequential LSTM based model and a convolutional alternative (CNN) Both models were trained using the binary crossentropy loss function, the Adam optimizer, and the accuracy metric The data partitioning described earlier was used; during training, a separate portion of the data was used as the validation set while final quality assessment was performed solely on the test set Analysis of Explanation Visualization A native text graph was used for the visualization of the SHAP method, where each token is assigned a color and intensity corresponding to its contribution to the model’s decision The observed saliency maps well reflect the inductive properties of the compared architectures, with CNN attributions clustering into compact, local patches In contrast, in the LSTM model, the inputs are more dispersed in time, resulting in highlights covering wider areas of text Despite quantitative differences, all XAI methods add value by increasing transparency, revealing artifacts, supporting error analysis, improving stakeholder communication, and informing model design The results also show that explanation quality depends strongly on model architecture: SHAP works best for LSTMs while Integrated Gradients outperforms others for CNNs No XAI method is universally optimal; effectiveness is architecture dependent Ultimately, all XAI methods add value by increasing transparency, revealing artifacts, supporting error analysis, improving stakeholder communication, and informing model design Final Assessment of Interpretation The adopted fidelity assessment procedure is also a limitation because all metrics were based on input perturbations implemented by replacing tokens with PAD While this solution ensures technical consistency, it can lead to unnatural examples and distort the behavior of models, especially LSTMs The obtained metric values should therefore be interpreted in terms of relative comparisons rather than absolute quality measures Finally, it is important to emphasize the limitations of interpretation because the explanations obtained in this study are local in nature and cannot be generalized to the global performance of the model The observed differences in the attribution distribution between LSTM and CNN are largely due to the nature of the architectures, not solely to the performance of the explanation methods themselves References

1: Athira A.B., S.D.

Tom: Madhu Kumar, and Anu Mary Chacko.

Jane: A systematic survey on explainable ai applied to fake news detection

2: Scott M Lundberg and Su-In Lee.

Tom: A unified approach to interpreting model predictions

3: Malatthi Sivasundaram, V.

Title and authors: Tom: Okay, so we know the title is "Trust Oriented Explainable AI for Fake News Detection." It sets the stage for how we build models that people can actually rely on when they’re dealing with information like this.

Jane: But what does that actually mean in plain language? I don't want to hear about weights or layers; I want to know what this means for someone just trying to figure out if a piece of online news is real or fake.

Lu: It means we are moving beyond just having a model that spits out a "fake" or "real" answer. We’re asking the model, and then showing us *why* it made that call, so we can actually understand the reasoning behind it, which is a big step toward building systems people trust.

Meng: So you’re saying instead of just trusting the output score, we trust the explanation for that score? Because if I'm deploying this on a live feed, I need to know why it flagged something so I can debug when it messes up.

Lalam: Exactly. The paper suggests that by using tools like SHAP or LIME, we get different kinds of explanations. For instance, SHAP gives you detailed local attributions for each word in a sentence, while LIME gives you a simpler explanation of what parts of the text were most important for the decision.

The paper's summary: Tom: So, to summarize this paper about "Trust Oriented Explainable AI for Fake News Detection," they implemented classification models and then used these interpretability techniques to see how well they explained their decisions.

Jane: Can you unpack that for me? What are the key findings regarding the performance when you compare SHAP, LIME, and Integrated Gradients? Because I’m picturing a lot of different results depending on what tool you choose.

Lu: The paper shows that each method offers a different kind of explanation. SHAP gives you detailed local attributions for every word, LIME provides simple and intuitive explanations that are easy to grasp, and Integrated Gradients works pretty efficiently with convolutional neural networks.

Meng: That efficiency is key for production, right? Because when we deal with massive amounts of text data, the computational cost of generating an explanation needs to be low.

Lalam: Right. The study also points out that there are some limitations to these methods, like the computational cost or how sensitive they can be to how you set up the parameters for them. But overall, they conclude that integrating XAI with NLP is a really effective way to boost how reliable and trustworthy fake news detection systems are.

The paper's improvements: Tom: So, beyond just comparing the methods, what did the authors suggest we should do next? What’s the actual improvement they propose for this field?

Jane: I’m hoping they suggest a way to make these explanations even better. Because as I said earlier, right now feature attribution might not be enough to explain why a piece of text is deceptive.

Lu: They point out that current methods like SHAP and LIME primarily offer feature attribution rather than a causal narrative, which often fails to explain the actual linguistic nuances or the logical fallacies that make content deceptive.

Meng: So they’re moving beyond just showing us *which* words were important to showing us *how* those words create the deception? That sounds like it could help us catch more complex manipulation.

Lalam: Yes, a pivotal shift in the state of the art is the movement from feature level attribution to Generative Justification. They suggest leveraging capabilities from Large Language Models to introduce frameworks that provide this kind of justification for why something is flagged.

Conclusion: Tom: So, wrapping up on "Trust Oriented Explainable AI for Fake News Detection," the paper confirms that XAI methods really do enhance model transparency and interpretability while keeping the detection accuracy high across different architectures.

Jane: It seems like the main implication is that we don't just want accurate models; we need models whose reasoning we can actually inspect, and this paper shows us how to get started with tools like SHAP or Integrated Gradients.

Lu: It does confirm what they found: the effectiveness of XAI depends heavily on the model architecture you’re using. SHAP works best when you have a recurrent network like an LSTM, while Integrated Gradients seems better for convolutional networks like CNNs.

Meng: For deployment, that's helpful because it tells us not to assume one explanation method will work universally; we need to match the tool to the model we’re running.

Lalam: To finish up, the paper emphasizes that while these explanations help with error analysis and stakeholder communication, they aren't perfect. They mention a limitation where explanations are often local in nature and can’t be generalized to the overall performance of the model.

Tom: So, so we’ve seen how SHAP excels with LSTMs, Integrated Gradients is strong for CNNs, and even though there are limitations regarding causality and interpretation, using XAI is definitely a value-add for building better detection systems.

Jane: It's clear that this work on "Trust Oriented Explainable AI for Fake News Detection" gives us concrete tools to make the black box a little less scary by providing specific, measurable explanations.

Lu: It’s a foundation; it shows that we need to keep looking at hybrid architectures and moving toward generative justification to really capture those deeper logical fallacies.

Meng: For production, it’s good knowledge that we have different trade-offs when choosing between those explanation tools based on whether our underlying model is sequential or convolutional.

Lalam: The core message is that XAI adds value by helping us see the artifacts in the model's decision-making process, even if it doesn't solve every problem with causality or manipulation risks.

Warsaw University of Technology

cs.CL

Submitted: 2026-03-12

Updated: 2026-03-12

Comments: 9 pages, 4 figures, 2 tables

DOI: 10.1109/CPEE71962.2026.11768167

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 75/100

The gist: This article examines the application of Explainable Artificial Intelligence (XAI) in NLP based fake news detection and compares selected interpretability methods to enhance model transparency and…

Key concepts

SHAP
SHAP is a technique that assigns an importance value to each feature (in this case, text tokens) based on how much it contributes to the final prediction. It provides detailed local attributions, showing exactly which words in a sentence influenced the model's decision.
Integrated Gradients
This method calculates the contribution of each input element by summing up gradients along a path from an 'empty' input to the actual text embedding. It is particularly efficient for convolutional models and helps visualize how specific parts of an image or sequence drive the output.
Model Architecture Dependency
The quality and effectiveness of XAI explanations are not universal; they are highly dependent on the underlying neural network structure. For instance, SHAP performed best with LSTMs because it handles sequential context well, whereas Integrated Gradients was superior for CNNs due to their spatial pattern recognition.
Local Explanation vs. Global Performance
The explanations generated by XAI methods are inherently local, meaning they only explain a single prediction. This limits their ability to represent the model's overall global behavior. Therefore, explanations should be used for error analysis and trust-building rather than as a complete picture of the model's entire decision-making process.

Terminology

Summary

This article examines the application of Explainable Artificial Intelligence (XAI) in NLP based fake news detection and compares selected interpretability methods to enhance model transparency and trustworthiness The results show that XAI enhances model transparency and interpretability while maintaining high detection accuracy

Model Comparison of Interpretability Techniques

The work compares SHAP, LIME, and Integrated Gradients across two neural network architectures to determine their explanatory value Each method provides distinct explanatory value: SHAP offers detailed local attributions, LIME provides simple and intuitive explanations, and Integrated Gradients performs efficiently with convolutional models The study also highlights limitations such as computational cost and sensitivity to parameterization Overall, the findings demonstrate that integrating XAI with NLP is an effective approach to improving the reliability and trustworthiness of fake news detection systems

Neural Network Architectures in NLP

Natural Language Processing (NLP) uses various neural architectures including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and attention based models such as transformers Recurrent Neural Networks (RNNs) introduced feedback loops that allow information from previous time steps to influence future computations, enabling the preservation of context in longer sequences Long Short Term Memory (LSTM) architecture utilizes a cell state and three distinct gates to manage information through sigmoid and tanh activation functions The Transformer architecture is characterized by a symmetric stack of N identical layers, designed to optimize both contextual representation and sequential generation

Implementation Details of XAI Methods

The implementation of XAI methods involved integrating the explanation layer with the existing input data representation, where sequences of token identifiers were used For SHAP, a processing procedure was established where a text masker generates multiple versions of a sentence with hidden words and then SHAP asks the model to evaluate these versions LIME creates multiple slightly modified versions of the sentence close to the original and trains a simple surrogate model describing the local behavior of the classifier Integrated Gradients is implemented as a gradient method operating directly on the embedding representation, transitioning from an ”empty” text to the actual input

Performance Metrics and Results

The classification quality was assessed using several metrics, including completeness, sufficiency, AOPC (emphArea Over the Perturbation Curve), and Flip@k For LSTM, SHAP delivers the best results with the strongest removal effect and highest AOPC For CNNs, Integrated Gradients is the best method, combining the highest completeness and AOPC with the lowest Flip@k The results also show that explanation quality depends strongly on model architecture, with SHAP working best for LSTMs while Integrated Gradients outperforms others for CNNs

Limitations and Future Directions

The study identifies limitations associated with XAI, including computational overhead, sensitivity to parameterization, and the risk of misinterpretation by end users A primary risk is the illusion of causality, where explanations often reflect correlations learned from training data rather than true domain level causes Explanations are also susceptible to manipulation, as models can be intentionally modified to generate misleading explanations while preserving biased or undesirable decision rules The findings offer a foundation for future work involving transformer based architectures, rule based explanations, end user studies, model calibration, and linguistically grounded perturbation protocols

Data Preparation and Model Training

The experiments were conducted using the publicly available ISOT Fake News Dataset Data preprocessing involved building a dictionary limited to the 20,000 most frequently occurring words and standardizing sequence length to a maximum of 750 tokens Two neural network architectures were evaluated: a sequential LSTM based model and a convolutional alternative (CNN) Both models were trained using the binary crossentropy loss function, the Adam optimizer, and the accuracy metric The data partitioning described earlier was used; during training, a separate portion of the data was used as the validation set while final quality assessment was performed solely on the test set

Analysis of Explanation Visualization

A native text graph was used for the visualization of the SHAP method, where each token is assigned a color and intensity corresponding to its contribution to the model’s decision The observed saliency maps well reflect the inductive properties of the compared architectures, with CNN attributions clustering into compact, local patches In contrast, in the LSTM model, the inputs are more dispersed in time, resulting in highlights covering wider areas of text

Conclusion on XAI Value

Despite quantitative differences, all XAI methods add value by increasing transparency, revealing artifacts, supporting error analysis, improving stakeholder communication, and informing model design The results also show that explanation quality depends strongly on model architecture: SHAP works best for LSTMs while Integrated Gradients outperforms others for CNNs No XAI method is universally optimal; effectiveness is architecture dependent Ultimately, all XAI methods add value by increasing transparency, revealing artifacts, supporting error analysis, improving stakeholder communication, and informing model design

Final Assessment of Interpretation

The adopted fidelity assessment procedure is also a limitation because all metrics were based on input perturbations implemented by replacing tokens with PAD While this solution ensures technical consistency, it can lead to unnatural examples and distort the behavior of models, especially LSTMs The obtained metric values should therefore be interpreted in terms of relative comparisons rather than absolute quality measures Finally, it is important to emphasize the limitations of interpretation because the explanations obtained in this study are local in nature and cannot be generalized to the global performance of the model The observed differences in the attribution distribution between LSTM and CNN are largely due to the nature of the architectures, not solely to the performance of the explanation methods themselves

References

[1] Athira A.B., S.D. Madhu Kumar, and Anu Mary Chacko. p. 8

Where the paper says it

REFERENCES [1] Athira A. Madhu Kumar, and Anu Mary Chacko.

A systematic survey on explainable ai applied to fake news detection

[2] Scott M Lundberg and Su-In Lee. p. 8

Where the paper says it

Engineering Applications of Artificial Intelligence, 122:106087, 2023. [2] Scott M Lundberg and Su-In Lee.

A unified approach to interpreting model predictions

[3] Malatthi Sivasundaram, V. Prakasham, R. p. 8

Where the paper says it

[3] Malatthi Sivasundaram, V. Prakasham, R.

Rathika, P. p. 8
Where the paper says it

Prakasham, R. Rathika, P.

Subhashini, P. p. 8
Where the paper says it

Rathika, P. Subhashini, P.

Karthi, and A. p. 8
Where the paper says it

Subhashini, P. Karthi, and A.

Priyadharshini. p. 8
Where the paper says it

Karthi, and A. Priyadharshini.

Hybrid deep learning for advanced fake news detection using explainable ai and fast text

[4] Bimal Bhattarai, Ole-Christoffer Granmo, and Lei Jiao. Explainable tsetlin machine framework for fake news detection with credibility score assessment

[5] Yifeng Wang, Zhouhong Gu, Siwei Zhang, Suhang Zheng, Tao Wang, Tianyu Li, Hongwei Feng, and Yanghua Xiao. Llm-gan: Construct generative adversarial network through large language models for explainable fake news detection

[6] Sabrine Amri, Henri-Cedric Mputu Boleilanga, and Esma A¨ımeur. Exfake: Towards an explainable fake news detection based on content and social context information

[7] Antino Kim, Patricia L. Moravec, and Alan R. p. 8

Where the paper says it

[7] Antino Kim, Patricia L. Moravec, and Alan R.

Dennis. Combating fake news on social media with source ratings: The effects of user and expert reputation ratings

[8] Greta Arancia Sanna and David Lagnado. Belief updating in the face of misinformation: The role of source reliability

[9] David Martens, Galit Shmueli, Theodoros Evgeniou, Kevin Bauer, Christian Janiesch, Stefan Feuerriegel, Sebastian Gabel, Sofie Goethals, Travis Greene, Nadja Klein, Mathias Kraus, Niklas Kuhl

[10] Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory

[11] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling

[12] Kai Heinrich, Patrick Zschech, Christian Janiesch, and Markus Bonin. Process data properties matter: Introducing gated convolutional neural networks (gcnn) and key-value-predict attention networks (kvp) for next event prediction with deep learning

[13] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. p. 8

Where the paper says it

[13] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin.

Attention is all you need

[14] Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions

[15] Mukund Sundararajan and Amir Najmi. The many shapley values for model explanation

[16] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. ”why should i trust you? p. 8

Where the paper says it

[16] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. ”why should i trust you?

”: Explaining the predictions of any classifier

[17] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks

[18] Upol Ehsan and Mark O. Riedl. Explainability pitfalls: Beyond dark patterns in explainable ai

[19] Philipp Spitzer, Katelyn Morrison, Violet Turri, Michelle Feng, Adam Perer, and Niklas Kuhl. Imperfections of xai: Phenomena influencing ¨ ai-assisted decision-making

[20] Florien S. van Royen, Hilde J. p. 8

Where the paper says it

[20] Florien S. van Royen, Hilde J.

P. Weerts, Anne A. p. 8
Where the paper says it

van Royen, Hilde J. Weerts, Anne A.

H. de Hond, GeertJan Geersing, Frans H. p. 8
Where the paper says it

Weerts, Anne A. de Hond, GeertJan Geersing, Frans H.

Rutten, Karel G. p. 8
Where the paper says it

de Hond, GeertJan Geersing, Frans H. Rutten, Karel G.

M. Moons, and Maarten van Semedoen In humble defense of unexplainable black box prediction models in healthcare

[21] Kaggle.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Use SHAP for LSTM models to achieve the best results in explanation quality, as it best reflect model architecture and produces clear textlevel explanations. This allows for a more accurate understanding of how recursive architectures use context compared to other methods.

  2. Employ Integrated Gradients for CNN models because it is the best method for them, combining the highest completeness and AOPC with the lowest sufficiency and the lowest Flip@k. This makes it highly efficient while capturing local patterns effectively.

  3. Implement a comparative evaluation strategy where XAI scores are drawn within each model, avoiding direct numeric comparisons between LSTM and CNN results, as they depend on classifier properties. This prevents misinterpreting differences as inherent superiority of one architecture over the other.

Abstract

This article examines the application of Explainable Artificial Intelligence (XAI) in NLP based fake news detection and compares selected interpretability methods. The work outlines key aspects of disinformation, neural network architectures, and XAI techniques, with a focus on SHAP, LIME, and Integrated Gradients. In the experimental study, classification models were implemented and interpreted using these methods. The results show that XAI enhances model transparency and interpretability while maintaining high detection accuracy. Each method provides distinct explanatory value: SHAP offers detailed local attributions, LIME provides simple and intuitive explanations, and Integrated Gradients performs efficiently with convolutional models. The study also highlights limitations such as computational cost and sensitivity to parameterization. Overall, the findings demonstrate that integrating XAI with NLP is an effective approach to improving the reliability and trustworthiness of fake news detection systems.

Related papers