Experiments with truth using Machine Learning - Spectral analysis and explainable classification of synthetic, false, and genuine information

Aug 5, 2026

Misinformation remains a major societal issue, and recent generative AI systems further amplify the challenge. This work analyzes synthetic, false, and genuine text using spectral analysis, visualization, and explainability to investigate why containment efforts still underperform. Multiple embedding techniques and datasets are compared, using t-SNE, PCA, and VAEs for both spectral and non-spectral views. Classifiers are evaluated with LIME, SHAP, and Integrated Gradients, showing limited separability between misinformation and truthful content.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Introduction

Related Work

Contribution

Experiments with truth using Machine Learning: Spectral analysis and explainable classification of synthetic, false, and genuine information Vishnu S. Pendyala Department of Applied Data Science San Jose State University San Jose, CA 95192-0250 Email: [vishnu.pendyala@sjsu.edu](vishnu.pendyala@sjsu.edu) Madhulika Dutta Department of Applied Data Science San Jose State University San Jose, CA 95192-0250 [Email: madhulika.dutta@sjsu.edu](Email: madhulika.dutta@sjsu.edu) Abstract—Misinformation is still a major societal problem and the arrival of Large Language Models (LLMs) only added to it. This paper analyzes synthetic, false, and genuine information in the form of text from spectral analysis, visualization, and explainability perspectives to find the answer to why the problem is still unsolved despite multiple years of research and a plethora of solutions in the literature. Various embedding techniques on multiple datasets are used to represent information for the purpose. The diverse spectral and non-spectral methods used on these embeddings include t-distributed Stochastic Neighbor Embedding (t-SNE), Principal Component Analysis (PCA), and Variational Autoencoders (VAEs). Classification is done using multiple machine learning algorithms. Local Interpretable Model-Agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), and Integrated Gradients are used for the explanation of the classification. The analysis and the explanations generated show that misinformation is quite closely intertwined with genuine information and the machine learning algorithms are not as effective in separating the two despite the claims in the literature. I. INTRODUCTION Misinformation, fake news, and lies have been adversely impacting society in many significant ways. The advent of generative AI applications such as ChatGPT has only exasperated the problem [1] . Misinformation can be multimodal. Generative AI is capable of producing misinformation in multiple modalities. However, while the difficulties in detecting fake images are well documented in the literature [2], the same is not true for misinformation in the form of text. On the other hand, some studies have reported 100% accuracy in detecting AI-generated text using simple language models such as BOW [3] . There is also abundant literature on solving the misinformation containment problem with humangenerated text but it is a well-known fact that the problem is still largely unsolved [4] . This work expands on the previous analysis of the problem [5] . Given the current gap in the literature in sufficiently identifying the reasons why misinformation containment is still an unsolved problem, there is a need to focus on why machine learning is unable to solve the problem despite the tall claims to the contrary in the literature, which is mostly based on the currently available embedding techniques. Embeddings are essentially representations of the input data in the hidden layers of neural networks. This work is an attempt to determine what makes it so difficult to identify misinformation based on the embeddings and the limitations of the current Neural Natural Language Processing (NNLP) and Machine Learning (ML) techniques in doing so using a variety of experiments to visualize, classify, and explain. II. RELATED WORK Misinformation containment is proven in the literature tobe NP-hard [6] . Misinformation detection can be addressed using diverse approaches, including algorithms such as the Kalman Filter [7], statistical techniques and first-order logic [8] . However, it is established in the literature that machine learning is a good alternative to heuristic algorithms to solve NP-hard problems [9] . A further literature survey naturally shows a comprehensive use of machine learning and deep learning in conjunction with NLP techniques to address the problem. There are multiple surveys [10] [11] describing the literature in this regard. Large language models (LL

FAQs

What does the paper investigate about misinformation detection?

It studies how effective current embedding techniques are at separating truthful information from false information, using visualization, classification, and explanation.

Which methods are used for spectral and non-spectral analysis?

The paper applies t-SNE and PCA for visualization and uses variational autoencoders (VAEs) alongside spectral and non-spectral perspectives on embeddings.

How are model predictions explained?

It uses LIME, SHAP, and Integrated Gradients to generate explanations for the classification results.

Aug 5, 2026

Upload by Clementine