GenAI tools are being used rapidly in academic writing, which has caused both benefits and problems. GenAI humanizers, for example Quillbot, GPT-Humanizer, and Undetectable.ai, help make AI text seem like it came from a human author. Even as they vow to enhance personalization and speech, such instruments cause major concerns about tone manipulation, semantic changes, reduced originality, and the dishonest marking of AI-written work. As a result of these new risks, academic papers and the main principles of research integrity may become questionable. This research introduces an original study into the harm caused by using humanizers in academic writing, using a custom-built dataset that simulates both the humanized changes and the ethical problems related to them. We gather and process two types of features: linguistic and semantic ones, including similarity scores, the chance of being an AI response, the accuracy of cited sources, detection of changing tone, and a measure of extra mental effort needed. The selected dataset is trained on the eight algorithms: Logistic Regression, Random Forest, Support Vector Machine, Decision Tree, Naive Bayes, K-Nearest Neighbors, Multi-Layer Perceptron, and XGBoost. 10-fold stratified cross-validation combined with SMOTE is put to use. The highest weighted F1 score, at 0.93, was obtained by XGBoost and this helped it identify subtle shifts caused by the humanizer tools better than traditional classifiers. It appears that AI-based writing may not be recognized by most AI detectors and yet may harm the truthfulness of what is written. Therefore, we suggest that researchers, editors, and institutions adopt guidelines for using AI responsibly, create clear disclosure standards, and introduce security strategies to lessen the chances of misusing technologies that make robotics seem more human.