Title Page
DEVELOPMENT OF AN AI-BASED SYSTEM FOR DETECTING FAKE NEWS ON NIGERIAN SOCIAL MEDIA
A Research Project Submitted to the Department of __________________________
Faculty of __________________________
__________________________ University
In Partial Fulfilment of the Requirements for the Award of the Degree of __________________________
By
__________________________
Matriculation Number: __________________________
Date: __________________________
Certification
This is to certify that this research project titled “Development of an AI-Based System for Detecting Fake News on Nigerian Social Media” was carried out by __________________________ under the supervision of __________________________ and has been approved as meeting the requirements for the award of __________________________.
Supervisor: __________________________
Signature: __________________________
Date: __________________________
Head of Department: __________________________
Signature: __________________________
Date: __________________________
Declaration
I declare that this research project is my original academic work, except where the ideas, findings and materials of other researchers have been appropriately acknowledged. The modelled results presented in Chapter Four are simulated research data developed for the purpose of demonstrating how the proposed system could be evaluated. They are not presented as results obtained from a real-world deployment or survey.
AI-Based System for Detecting Fake News
Dedication
I dedicate this research project to God Almighty, my family, teachers and everyone who continues to support my academic and professional development.
Acknowledgements
I sincerely appreciate my supervisor and lecturers for their guidance and academic contributions throughout this research. I also appreciate my family, colleagues and friends for their encouragement and support.
I am particularly grateful to everyone whose previous research and technical work contributed to my understanding of artificial intelligence, machine learning, natural language processing and misinformation detection.
Abstract
The rapid growth of social media has changed the way Nigerians receive, discuss and share information. Although social media provides fast access to news and public information, it also creates an environment in which false, misleading and manipulated information can spread rapidly. Nigeria’s growing digital population makes the problem increasingly important. DataReportal reported 109 million internet users in Nigeria in October 2025 and 47.8 million social media user identities at the same period. These figures illustrate the scale of the digital environment in which information can circulate.
This study focused on the development of an artificial intelligence-based system for detecting potentially fake news on Nigerian social media. The proposed system uses natural language processing and machine learning techniques to analyse the textual characteristics of social media news content and classify it as likely genuine or likely fake. The study adopted a design and development research approach. A proposed dataset structure containing labelled Nigerian social-media-style news statements was used to design the classification pipeline. Text preprocessing, tokenisation, TF-IDF feature extraction and supervised machine learning were incorporated into the system. Logistic Regression was selected as the principal baseline classifier because it is suitable for binary text classification and can work with sparse feature representations.
Because real-time data collection and deployment were not conducted for this academic project, the performance results in Chapter Four are explicitly simulated/modelled rather than claimed as empirical results from a live system. The modelled evaluation indicates that the proposed architecture can provide a practical foundation for automated preliminary screening of suspicious content. The research concludes that AI should support, rather than completely replace, human fact-checking because determining whether a claim is factually true may require external evidence, source verification and contextual understanding. The study recommends continuous retraining with Nigerian datasets, inclusion of local languages and Nigerian Pidgin, explainable predictions, human review and integration with trusted fact-checking sources.
Keywords: Artificial Intelligence, Fake News Detection, Social Media, Machine Learning, Natural Language Processing, Nigeria, Text Classification.
Table of Contents
Title Page
Certification
Declaration
Dedication
Acknowledgements
Abstract
Table of Contents
List of Tables
List of Abbreviations
Chapter One: Introduction
Chapter Two: Literature Review
Chapter Three: Methodology and System Design
Chapter Four: System Implementation, Modelled Results and Discussion
Chapter Five: Summary, Conclusion and Recommendations
References
List of Tables
Table 3.1 Proposed Functional Requirements
Table 3.2 Proposed Dataset Structure
Table 4.1 Modelled Dataset Distribution
Table 4.2 Modelled Classification Results
Table 4.3 Modelled Confusion Matrix
Table 4.4 Modelled Performance Metrics
Table 4.5 Comparison of the Proposed Classifiers
List of Abbreviations
AI – Artificial Intelligence
API – Application Programming Interface
ML – Machine Learning
NLP – Natural Language Processing
TF-IDF – Term Frequency-Inverse Document Frequency
UI – User Interface
URL – Uniform Resource Locator
TP – True Positive
TN – True Negative
FP – False Positive
FN – False Negative
Chapter One: Introduction To AI-Based System for Detecting Fake News
1.1 Background of the Study
Social media has become an important part of modern communication. Platforms such as Facebook, WhatsApp, X, Instagram, TikTok and other digital services allow users to create, receive and share information almost instantly. In Nigeria, the growth of internet connectivity has increased the importance of these platforms as channels for news, political communication, business information, entertainment and public discussion.
DataReportal’s Digital 2026 Nigeria report indicates that Nigeria had approximately 109 million internet users in October 2025. The same report estimated 47.8 million social media user identities in Nigeria at that time. The report also cautions that social-media user identities should not automatically be interpreted as unique individual people because of methodological limitations.
The growth of social media has created significant benefits. Information can be distributed quickly, citizens can participate in public discussions, journalists can obtain eyewitness information and organisations can communicate directly with their audiences. However, the same characteristics that make social media useful also make it suitable for the rapid distribution of misinformation and disinformation.
Fake news is commonly used to describe false or misleading information presented in a form that resembles legitimate news. Researchers have noted that fake-news detection is particularly difficult on social media because the textual content is only one part of the problem. User behaviour, source characteristics and patterns of information propagation may also provide useful signals. Shu et al. describe social-media fake-news detection as a distinctive problem involving content as well as auxiliary social information.
The Nigerian situation deserves particular attention because social media has become an important communication environment while misinformation can affect political discussions, public health, security, business decisions and interpersonal relationships. A 2025 study of Lagos social-media users reported substantial exposure to fake news and concluded that technical interventions should be combined with continuing media-literacy efforts.
Artificial intelligence provides an opportunity to develop systems capable of processing large quantities of online content. Machine-learning algorithms can learn patterns from labelled examples and subsequently classify new pieces of text. Natural language processing allows computers to transform human language into representations that can be analysed computationally.
In this research, I therefore propose an AI-based system that can act as an initial screening mechanism for potentially fake news on Nigerian social media. The purpose is not to claim that an algorithm can determine the truth of every social-media statement. Instead, the proposed system is designed to identify linguistic and textual patterns associated with previously labelled fake and genuine content and present a probability or classification that can be subjected to human verification.
1.2 Statement of the Problem
The volume and speed of information shared through social media make manual monitoring difficult. A person or fact-checking organisation cannot realistically inspect every post, headline, forwarded message or online claim that circulates across Nigerian social-media platforms.
Users may also share information before verifying it. Research involving Nigerian social-media users has identified factors such as rapid news sharing, self-promotion, self-expression and altruism as relevant to the spread of false news, while social-media knowledge can influence users’ behaviour.
Traditional manual fact-checking remains important, but it can be slow when large quantities of content are involved. There is therefore a need for automated tools that can identify suspicious content for further investigation.
A second problem is that many generic fake-news detection systems are developed using datasets from other countries and contexts. A model trained primarily on foreign political language, cultural references and news styles may not perform equally well on Nigerian content.
A third problem concerns the meaning of “fake”. A machine-learning model may identify statistical patterns associated with fake content without actually establishing whether a claim is factually correct. A robust system must therefore be designed as a decision-support tool rather than an unquestionable authority.
1.3 Aim of the Study
The main aim of this study is to develop an AI-based system for detecting potentially fake news on Nigerian social media.
1.4 Objectives of the Study
The specific objectives are to:
1. examine the characteristics of fake-news content circulated through Nigerian social media;
2. design a dataset structure suitable for Nigerian fake-news classification;
3. develop a text preprocessing pipeline for social-media content;
4. extract relevant textual features using natural language processing techniques;
5. develop a machine-learning classification model for identifying potentially fake content;
6. design a user interface through which users can submit news text for analysis;
7. evaluate the proposed system using standard classification metrics; and
8. identify the limitations and possible improvements of an AI-based fake-news detection system in Nigeria.
1.5 Research Questions
1. What characteristics can be used to distinguish potentially fake news from genuine news content?
2. How can Nigerian social-media news content be prepared for machine-learning classification?
3. What machine-learning approach is suitable for an initial fake-news classification system?
4. How can an AI-based system provide users with an understandable indication of potentially fake content?
5. How effective is the proposed model under the simulated evaluation framework?
6. What limitations may affect the performance of an AI-based fake-news detection system in Nigeria?
1.6 Research Hypotheses
H0₁: An AI-based text-classification system does not provide a useful basis for distinguishing potentially fake news from genuine news.
H1₁: An AI-based text-classification system provides a useful basis for distinguishing potentially fake news from genuine news.
H0₂: There is no meaningful difference in the classification performance of the proposed machine-learning algorithms.
H1₂: There is a meaningful difference in the classification performance of the proposed machine-learning algorithms.
1.7 Significance of the Study
This study may be useful to social-media users because an automated screening tool can encourage them to pause and verify suspicious information before sharing it.
Journalists and fact-checkers may also benefit from automated prioritisation. Instead of manually examining every item, they could use a detection system to identify content requiring closer investigation.
Researchers and students in computer science, information technology, mass communication and related disciplines may use the study as a foundation for further research into Nigerian misinformation detection.
Technology developers may also find the system architecture useful when developing information-verification tools adapted to Nigerian language and social-media conditions.
1.8 Scope of the Study
The study focuses on the development of an AI-based system for detecting potentially fake news using textual information associated with Nigerian social media. The proposed system concentrates primarily on text classification.
The research does not attempt to build a complete commercial social-media monitoring platform. It also does not claim to determine the objective truth of every online claim. Image manipulation, deepfake video detection, audio verification and comprehensive social-network analysis are outside the main scope of the prototype.
1.9 Operational Definition of Terms
Artificial Intelligence: Computer-based techniques that enable systems to perform tasks normally associated with human intelligence, including learning and classification.
Fake News: False or misleading information presented or circulated in a form resembling news or factual information.
Machine Learning: A branch of AI in which algorithms learn patterns from data to make predictions or decisions.
Natural Language Processing: Computational methods used to process and analyse human language.
Social Media: Digital platforms through which users create, publish, communicate and share content.
Classification: The process of assigning an item to a predefined category.
TF-IDF: A numerical representation that assigns weights to terms based partly on their frequency within a document and their distribution across documents.
Fact-Checking: The process of investigating a claim against credible evidence to determine its accuracy.
Chapter Two: Literature Review Of AI-Based System for Detecting Fake News
2.1 Concept of Fake News
Fake news is a complex concept because false information can take several forms. It may be deliberately fabricated, unintentionally inaccurate, misleadingly framed or presented without sufficient context. Researchers therefore distinguish between concepts such as misinformation, disinformation and rumours.
Surveys of the field show that automatic detection can consider the content of a claim, writing style, source characteristics, propagation behaviour and external knowledge. Zhou and Zafarani organise fake-news detection research around knowledge, style, propagation and source-related perspectives.
2.2 Misinformation and Disinformation
Misinformation generally refers to false or inaccurate information shared without necessarily having an intention to deceive, while disinformation normally involves deliberately misleading information. The distinction is important because an automated system may be able to analyse the characteristics of a message but may not be able to establish the intention of the person who created or shared it.
For this reason, the proposed system uses the practical term “potentially fake” rather than claiming that every classified item represents intentional deception.
2.3 Social Media and Fake News in Nigeria
Nigeria has a large and expanding digital information environment. The available 2026 country report from DataReportal estimates 47.8 million social-media user identities and 109 million internet users based on data available around October 2025.
The Nigerian context also has specific linguistic and cultural characteristics. English is widely used online, but Nigerian Pidgin and indigenous languages are important communication forms. Social-media users also employ abbreviations, slang, code-switching, screenshots, forwarded messages and informal spelling. These characteristics can make automated text classification more difficult than classification of carefully edited conventional news articles.
Recent Nigerian research has continued to document the practical problem. A 2026 study examining Nigerian internet users investigated how people experience, interpret and respond to online disinformation.
2.4 Artificial Intelligence and Fake-News Detection
AI-based fake-news detection normally involves several stages. First, relevant data are collected and labelled. Second, the text is cleaned and transformed into machine-readable features. Third, a classification algorithm is trained. Finally, the model is tested on previously unseen examples.
Shu et al. identified fake-news detection on social media as a distinct research problem because social-media information contains both content and auxiliary signals such as social engagement.
A broader survey of machine-learning approaches also demonstrates the wide range of algorithms and feature representations that have been investigated for fake-news detection.
2.5 Natural Language Processing
NLP provides the foundation for processing the textual component of the proposed system. Raw text cannot normally be supplied directly to a conventional machine-learning classifier. It must first be transformed into numerical features.
Typical preprocessing operations include conversion to a consistent case, removal of unnecessary characters, tokenisation, handling of URLs and punctuation, and removal of selected stop words where appropriate.
However, preprocessing must be designed carefully for social-media data. Excessive cleaning can remove information that may actually help classification. For example, hashtags, repeated punctuation and certain words may contain useful signals. The proposed system therefore uses controlled preprocessing rather than indiscriminate removal.
2.6 TF-IDF Feature Extraction
TF-IDF is a common representation for text-classification tasks. It transforms a collection of raw documents into a numerical matrix in which terms receive weights based on their distribution across documents. The scikit-learn documentation describes TfidfVectorizer as a tool for converting raw documents into a TF-IDF feature matrix.
For this research, TF-IDF is appropriate because it provides a relatively simple and computationally efficient baseline for classifying Nigerian social-media text.
2.7 Logistic Regression
Logistic Regression is commonly used for binary classification. Despite its name, the machine-learning implementation treats it as a classification model rather than a conventional regression model. It can produce class probabilities and can work with sparse input representations.
Its interpretability and relatively modest computational requirements make it suitable for a prototype system. It can also provide a useful baseline against which more complex models can later be compared.
2.8 Other Machine-Learning Approaches
Other approaches that may be considered include Naive Bayes, Support Vector Machines, decision trees, random forests, gradient-boosting models and neural-network architectures.
Deep-learning and transformer-based models may capture more complex language relationships than traditional TF-IDF models. However, they usually require greater computational resources and sufficiently large, high-quality training data. They may also introduce additional challenges involving interpretability and model maintenance.
2.9 Existing Research Gap
A major gap is the need for systems designed around Nigerian social-media language and information conditions. A generic model may perform well on a benchmark dataset but fail when applied to Nigerian slang, local political references, Nigerian Pidgin, mixed-language posts or culturally specific claims.
Another gap is explainability. A system that simply returns “FAKE” without explaining why the content was flagged may encourage users to treat the prediction as an unquestionable fact. A more responsible system should provide a probability or risk level and encourage verification.
There is also a gap between automated classification and actual fact-checking. A classifier can identify patterns correlated with previously labelled examples, but factual verification may require external sources. Therefore, the proposed system is positioned as a preliminary screening tool.
2.10 Theoretical Framework
This study is primarily informed by the technological perspective of automated information classification and the information-processing perspective of misinformation detection.
The machine-learning component assumes that labelled examples contain measurable patterns that can be learned by an algorithm. Once trained, the model uses those patterns to classify previously unseen text.
The information-verification perspective recognises that classification and truth verification are not identical. A high-risk prediction should therefore lead to further investigation rather than automatic censorship or automatic deletion.
Chapter Three: Methodology and System Design On AI-Based System for Detecting Fake News
3.1 Research Design
I adopted a design and development research approach. The study involves identifying a practical problem, analysing the requirements, designing an AI-based solution, specifying the data-processing pipeline, developing a classification architecture and evaluating the proposed model.
3.2 Proposed System Architecture
The proposed system consists of the following major components:
1. User Interface: Allows a user to enter or paste a news statement, headline or social-media message.
2. Input Processing Module: Receives and validates the submitted text.
3. Text Preprocessing Module: Cleans and prepares the text for analysis.
4. Feature Extraction Module: Converts the processed text into numerical TF-IDF features.
5. Machine-Learning Model: Classifies the input as likely genuine or potentially fake.
6. Result Module: Displays the classification and confidence/risk information.
7. Verification Recommendation: Advises the user to verify important claims against reliable sources, especially where the model identifies high risk.
3.3 Proposed Functional Requirements
| Requirement | Description |
|---|---|
| Text input | The system should accept news or social-media text. |
| Preprocessing | The system should clean and standardise the input. |
| Feature extraction | The system should transform text into machine-readable features. |
| Classification | The system should predict the likely class of the input. |
| Result display | The system should display the prediction clearly. |
| Verification warning | The system should remind users that AI classification is not a substitute for fact-checking. |
3.4 Dataset Design
The proposed dataset would contain labelled examples of news statements and social-media-style content. Each record could contain an identification number, text, source category, topic, language, label and optional metadata.
| Field | Description |
|---|---|
| ID | Unique record identifier |
| Text | News statement or social-media content |
| Topic | Politics, health, economy, security, entertainment, etc. |
| Language | English, Nigerian Pidgin or relevant local language |
| Label | Genuine or Fake |
| Source | Reference or origin of the content where available |
In a real implementation, labels should be based on reliable verification rather than simply on whether a post looks suspicious. Multiple annotators and a documented labelling protocol would improve reliability.
3.5 Data Preprocessing
The preprocessing pipeline would perform the following operations:
1. remove unnecessary formatting;
2. normalise text where appropriate;
3. handle URLs and mentions;
4. tokenise text;
5. remove selected noise;
6. preserve potentially useful terms and patterns; and
7. convert the processed text into numerical features.
3.6 Feature Extraction
TF-IDF would be used to represent the processed text numerically. The resulting sparse matrix would provide input to the classification algorithm. The use of TF-IDF is technically appropriate for a baseline text classifier and is directly supported by established machine-learning libraries.
3.7 Model Development
I selected Logistic Regression as the main baseline model. The model would be trained using labelled examples and evaluated using a separate test set.
The basic classification process can be represented as:
Input text → preprocessing → TF-IDF vectorisation → trained classifier → prediction → risk indication.
Alternative classifiers such as Multinomial Naive Bayes and Support Vector Machine could also be evaluated to determine whether another algorithm performs better on the selected dataset.
3.8 Evaluation Metrics
The model would be evaluated using accuracy, precision, recall, F1-score and confusion-matrix analysis.
Accuracy measures the proportion of correctly classified records.
Precision measures the proportion of items predicted as fake that are actually fake.
Recall measures the proportion of actual fake items correctly detected.
F1-score combines precision and recall into a single measure.
For a misinformation-detection system, recall is particularly important because failing to identify a genuinely misleading item can allow it to pass through the screening process. However, excessively high sensitivity can produce many false positives. The system therefore needs a balanced evaluation.
3.9 Ethical Considerations
The proposed system should not automatically punish users or remove content solely because an AI model classifies it as suspicious. False positives are possible. Political and socially sensitive content should receive particular care.
User-submitted data should be handled responsibly, and personal information should not be collected unnecessarily. The system should also communicate its limitations clearly.
3.10 Methodological Limitation
The project does not involve a live deployment across Nigerian social-media platforms. It therefore does not claim to have measured the behaviour of an operational system on real-time Nigerian social-media traffic.
Chapter Four consequently uses simulated/modelled evaluation data to demonstrate how the system could be assessed after implementation. These figures must not be presented as actual field results.
Chapter Four: System Implementation, Modelled Results and Discussion On AI-Based System for Detecting Fake News
4.1 Introduction
This chapter describes the proposed implementation and presents a modelled evaluation of the system. Because no real-time data collection or live deployment was conducted for this project, the numerical results in this chapter are simulated for academic demonstration.
4.2 Modelled Dataset
For the purpose of demonstrating the evaluation process, I modelled a balanced dataset of 2,000 labelled text records, consisting of 1,000 potentially fake items and 1,000 genuine items. The records are hypothetical and do not represent an actual collected dataset.
| Class | Modelled Records | Percentage |
|---|---|---|
| Potentially Fake | 1,000 | 50% |
| Genuine | 1,000 | 50% |
| Total | 2,000 | 100% |
4.3 Modelled Training and Testing Arrangement
I modelled an 80:20 training-testing split. This produces 1,600 records for model training and 400 records for testing.
| Dataset Portion | Records | Percentage |
|---|---|---|
| Training | 1,600 | 80% |
| Testing | 400 | 20% |
| Total | 2,000 | 100% |
4.4 Modelled Classification Results
For demonstration, the principal Logistic Regression model is assumed to correctly classify 364 of the 400 test records.
| Metric | Modelled Result |
|---|---|
| Accuracy | 91.0% |
| Precision | 91.5% |
| Recall | 90.5% |
| F1-score | 91.0% |
These figures are simulated/modelled results only. They are included to demonstrate the expected structure of a machine-learning evaluation and must not be interpreted as actual performance obtained from a deployed Nigerian fake-news detection system.
4.5 Modelled Confusion Matrix
| Predicted Fake | Predicted Genuine | |
|---|---|---|
| Actual Fake | 181 | 19 |
| Actual Genuine | 17 | 183 |
Under this modelled scenario, 181 potentially fake items are correctly identified. Nineteen fake items are incorrectly classified as genuine. Seventeen genuine items are incorrectly flagged as fake, while 183 genuine items are correctly identified.
4.6 Modelled Comparison of Algorithms
To demonstrate comparative evaluation, the following hypothetical results are used.
| Algorithm | Accuracy | Precision | Recall | F1-score |
|---|---|---|---|---|
| Naive Bayes | 86.5% | 85.8% | 87.0% | 86.4% |
| Logistic Regression | 91.0% | 91.5% | 90.5% | 91.0% |
| Support Vector Machine | 92.0% | 92.3% | 91.4% | 91.8% |
Again, these figures are simulated rather than measured results. They suggest, for demonstration purposes, that a linear Support Vector Machine could outperform Logistic Regression slightly, while Logistic Regression provides a strong and relatively interpretable baseline.
4.7 Interpretation of the Modelled Results
The modelled results suggest that automated classification can provide useful preliminary screening. A modelled accuracy above 90% would indicate that the classifier is capable of learning meaningful patterns from the assumed labelled dataset.
However, accuracy alone should not determine whether the system is suitable for deployment. A model can achieve good overall accuracy while performing poorly on specific topics, languages or minority forms of expression.
False positives are also important. If a genuine news report is incorrectly labelled as fake, users may lose trust in the system. Conversely, false negatives may allow misleading content to pass through undetected.
4.8 Hypothesis Testing
Because the project does not contain a real experimental dataset, inferential hypothesis testing using the simulated results would not constitute a genuine statistical test of a population. Therefore, the hypotheses are evaluated conceptually rather than claimed as statistically proven.
H0₁: An AI-based text-classification system does not provide a useful basis for distinguishing potentially fake news from genuine news.
The modelled classification results demonstrate a plausible technical basis for distinguishing the two categories. Therefore, within the simulated design scenario, H0₁ would be rejected in favour of H1₁. This conclusion is limited to the modelled scenario and should not be presented as empirical evidence from a live Nigerian dataset.
H0₂: There is no meaningful difference in the classification performance of the proposed machine-learning algorithms.
The modelled results show different performance levels among the three hypothetical algorithms. Within the demonstration scenario, H0₂ would therefore be rejected. A real study would require statistical comparison using actual repeated experimental results before making a definitive conclusion.
4.9 Proposed User Interface
The proposed interface would contain a text box where a user can paste a headline, article excerpt or social-media message. A “Check Content” button would submit the text for processing.
The result page would display a classification such as:
Prediction: Potentially Fake
Risk Score: High
Recommendation: Verify this claim using credible sources before sharing.
The interface should avoid presenting the result as an absolute declaration of truth or falsehood.
4.10 Discussion of Findings
The literature supports the idea that fake-news detection can benefit from machine-learning and data-mining techniques. Existing research shows that content, writing style, source information and propagation characteristics can all contribute to detection.
The Nigerian context introduces additional challenges. Nigerian online communication may contain local expressions, mixed languages and culturally specific references that may not be adequately represented in international datasets. Consequently, a model trained on Nigerian examples should be expected to be more appropriate for the intended context than a system developed exclusively from unrelated datasets.
The modelled results also demonstrate why the system should be considered a screening tool. Even a hypothetical 91% accuracy leaves room for incorrect predictions. The remaining errors could have serious consequences if users interpreted the system as a perfect fact-checker.
Recent research also supports combining technical interventions with human awareness and fact-checking. A Lagos study recommended both technical actions and continuing media-literacy efforts, while research involving Nigeria and Kenya demonstrated the value of interventions aimed at reducing the sharing of misinformation.
4.11 Findings of the Study
Based on the system design and modelled evaluation, I found that:
1. AI and machine-learning techniques can provide a practical framework for preliminary fake-news classification.
2. Text preprocessing and feature extraction are important components of the proposed detection pipeline.
3. TF-IDF provides a suitable baseline representation for Nigerian social-media-style text.
4. Logistic Regression provides a practical baseline classifier for binary fake-news detection.
5. Machine-learning predictions should not be treated as final proof that a claim is false.
6. Nigerian-specific training data are important because language and cultural context can influence classification performance.
7. Human fact-checking remains necessary for high-impact and ambiguous claims.
8. The proposed system can be expanded in the future to incorporate source credibility, external evidence, network information, images and multilingual processing.
4.12 Limitations of the Study
The first limitation is the absence of a live deployment. The system has been designed conceptually and evaluated using modelled results rather than real-time social-media traffic.
The second limitation is the use of simulated data in Chapter Four. Although the figures demonstrate how the system could be evaluated, they cannot establish the actual performance of the proposed system.
The third limitation is that the proposed baseline focuses primarily on text. Modern misinformation can involve images, videos, manipulated screenshots, audio and coordinated networks.
The fourth limitation is language diversity. A system initially trained on English-language content may perform less effectively on Nigerian Pidgin and indigenous Nigerian languages.
The fifth limitation concerns changing misinformation patterns. Users and malicious actors can change their writing styles, narratives and methods over time. The model would therefore require periodic retraining and evaluation.
The sixth limitation concerns external verification. Text classification does not automatically establish whether a statement is factually true. Reliable verification may require trusted sources, databases, expert review and contextual investigation.
Chapter Five: Summary, Conclusion and Recommendations
5.1 Summary
This study examined the development of an AI-based system for detecting potentially fake news on Nigerian social media. The study was motivated by the rapid growth of Nigeria’s digital information environment and the difficulty of manually examining the large volume of content circulated through social platforms.
The study reviewed the concepts of fake news, misinformation, disinformation, artificial intelligence, machine learning and natural language processing. It also examined existing approaches to fake-news detection and identified the importance of content features, source information, propagation patterns and contextual information.
A prototype architecture was proposed consisting of a user interface, input-processing module, text-preprocessing module, TF-IDF feature-extraction component, machine-learning classifier and result module.
Logistic Regression was selected as the main baseline classifier because of its suitability for binary text classification and sparse feature representations. Other algorithms were also considered for comparative evaluation.
Because actual data collection and live deployment were not undertaken, simulated/modelled data were used in Chapter Four. The modelled results demonstrated how accuracy, precision, recall, F1-score and confusion-matrix analysis could be applied to evaluate the proposed system.
5.2 Conclusion
I conclude that an AI-based fake-news detection system can provide a useful first layer of screening for social-media content in Nigeria. Machine-learning models can analyse large amounts of textual information much faster than manual review and can identify patterns associated with previously labelled examples.
However, AI classification should not be confused with factual verification. A system may classify a message incorrectly, and even a strong classifier cannot necessarily understand every political, cultural, linguistic or contextual issue surrounding a claim.
The most appropriate approach is therefore a human-in-the-loop model. AI can identify suspicious content and prioritise it for investigation, while journalists, researchers, fact-checkers or knowledgeable users can examine evidence before a final conclusion is reached.
The Nigerian context also requires localised datasets. Future systems should incorporate Nigerian English, Nigerian Pidgin and relevant indigenous languages, while also accounting for local topics, expressions and communication patterns.
5.3 Recommendations
Based on the findings of the study, I recommend the following:
1. Develop a Nigerian-specific dataset. Researchers should create carefully labelled datasets containing genuine and misleading Nigerian social-media content.
2. Use multiple independent annotators. Labelling should involve more than one qualified reviewer, particularly for ambiguous claims.
3. Combine AI with human fact-checking. AI predictions should be treated as screening signals rather than final verdicts.
4. Expand language coverage. Future systems should support Nigerian Pidgin and relevant indigenous languages in addition to English.
5. Introduce explainability. Users should receive useful information about why content was flagged rather than seeing only a “fake” or “genuine” label.
6. Include external evidence. Future versions should connect predictions to credible fact-checking databases and trusted sources where technically and legally appropriate.
7. Retrain models regularly. Misinformation patterns change over time, so a static model may gradually lose effectiveness.
8. Protect user privacy. Developers should collect only the information required for the detection task and apply appropriate data-protection measures.
9. Avoid automatic censorship. A model prediction should not by itself result in account suspension, content removal or other punitive action.
10. Expand beyond text. Future research should investigate multimodal detection involving images, video, audio and manipulated screenshots.
5.4 Contribution to Knowledge
The main contribution of this study is the design of a context-oriented framework for applying AI-based text classification to fake-news screening in Nigerian social media.
The study also contributes a practical distinction between automated classification and factual verification. This distinction is important because a technically accurate classifier can still produce errors when confronted with unfamiliar language, sarcasm, breaking news or claims that require external evidence.
Finally, the proposed architecture provides a foundation for future research involving larger Nigerian datasets, multilingual models, transformer-based systems, source credibility analysis and human-in-the-loop fact-checking.
5.5 Suggestions for Further Research
Future researchers should develop and test the system using a large, independently verified Nigerian dataset. They should compare traditional machine-learning models with modern transformer-based language models and investigate how well each approach performs across different Nigerian languages and subject areas.
Further research should also investigate multimodal misinformation detection. Such a system could analyse the relationship between written claims, photographs, videos, captions and external evidence.
Another promising direction is explainable AI. Researchers could investigate methods that allow the system to identify the textual or contextual features that contributed to a prediction.
Finally, future work should conduct a real-world evaluation with appropriate ethical safeguards, rather than relying on the simulated evaluation used in this academic project.
References
Alghamdi, J., Luo, S., & Lin, Y. (2024). A comprehensive survey on machine learning approaches for fake news detection. Multimedia Tools and Applications, 83, 51009–51067.
Bondielli, A., & Marcelloni, F. (2019). A survey on fake news and rumour detection techniques. Information Sciences, 497, 38–55.
DataReportal. (2025). Digital 2025: Nigeria.
DataReportal. (2026). Digital 2026: Nigeria.
Okorie, N., & Olagunju, T. (2025). Social Media Messaging, Fake News and Fact-Checking Behaviours among Residents of Lagos, Nigeria. Journal of African Films & Diaspora Studies, 8(3).
Offer-Westort, M., Rosenzweig, L. R., & Athey, S. (2024). Battling the coronavirus ‘infodemic’ among social media users in Kenya and Nigeria. Nature Human Behaviour, 8, 823–834.
Ogbodo, J. N., Orji-Egwu, A. O., Njoku, E. O., et al. (2026). Dynamics and contours of online disinformation based on perspectives from Nigerian internet users. Discover Global Society, 4, Article 54.
Shu, K., Sliva, A., Wang, S., Tang, J., & Liu, H. (2017). Fake News Detection on Social Media: A Data Mining Perspective. ACM SIGKDD Explorations Newsletter, 19(1), 22–36.
Zhou, X., & Zafarani, R. (2020). A Survey of Fake News: Fundamental Theories, Detection Methods, and Opportunities. ACM Computing Surveys.
scikit-learn developers. (2026). TfidfVectorizer documentation. scikit-learn documentation.
scikit-learn developers. (2026). LogisticRegression documentation. scikit-learn documentation.
World Health Organization Regional Office for the Eastern Mediterranean. (2025). Strategies to combat infodemics in public health.