Hate speech is an ever-increasing menace among social media and online platforms. This covers harmful and offensive language directed towards an individual or group on the basis of race, gender, religion, or other identities. The alarming spread of hate speech creates toxic environments that have a serious collateral effect on individuals, including mental wellness and online safety. Most platforms have installed automatic systems to detect and remove hate speech, but fitness is often lacking. Traditional machine learning models like LSTM (Long Short-Term Memory) have been in use, especially in hate speech detection. Although these were good models, they seem to struggle to understand deeper meaning in most of their words and sentences and especially when the given speech features sarcasm or indirect hate. We propose improved approach in our project using the BERT (Bidirectional Encoder Representations from Transformers) model- an state-of-the-art Natural Language Processing model, and unlike LSTM which processes the words in a sequence, BERT reads an entire sentence in one go and understands both ways, thus making detection of hate speech that much easier even in the most complex and trickiest of sentences. BERT was trained on the social media comments dataset where both hate, and neutral languages are used. Thus, with these results, this comparison of BERT to LSTMs shows that hate speech can be identified more accurately with less error using BERT. It can find those more nuanced patterns of hate speech that traditional models usually won't pick up. Achieving online safety is therefore the main aim of this project: installing a system with a more trustworthy detection scheme specific for the detection of hate speech. BERT can help platforms minimize harmful content more effectively, creating a more secure digital space for users. This work underlines the essence of adopting modern AI techniques to address real-world issues and improve communication on the web.