Can Artificial Intelligence Models Detect Suicidal Thoughts on Social Media?

Every year, more than 720,000 people die by suicide worldwide, making it one of the leading causes of death among young adults. Social media platforms, such as Reddit, have become spaces where individuals openly express their psychological distress, sometimes in indirect or metaphorical ways. A recent study evaluated the ability of various computational tools to identify these distress signals in online posts.

Researchers compared several approaches: classical machine learning models, deep neural networks, and an advanced language model based on artificial intelligence. Traditional methods, such as support vector machines or random forests, were tested using semantic representations of texts. Neural networks, such as CNNs and LSTMs, were also assessed for their ability to capture contextual and emotional patterns. Finally, a fine-tuned GPT-2 model was trained to classify messages into two categories: suicidal or non-suicidal.

The results show that the GPT-2 model achieved the best performance, with an accuracy of 98.25%. It effectively distinguished at-risk messages from those that were not, while minimizing classification errors. Among the neural networks, the CNN achieved an accuracy of 96.04%, closely followed by the LSTM with 95.97%. Classical models, when combined with semantic representations such as the Universal Sentence Encoder, also demonstrated good effectiveness, with accuracies exceeding 93% for the best combinations.

The study highlights that transformer-based models, like GPT-2, excel at capturing emotional nuances and long contextual dependencies. These capabilities are particularly useful for detecting indirect expressions of distress, often present in social media posts. Traditional models, while less performant, remain competitive and more accessible in terms of computational resources.

Researchers also verified the robustness of the results by repeating experiments with different configurations and data partitions. The performance of the models, particularly GPT-2 and CNN, proved to be stable and reproducible, confirming their reliability in this context.

This study underscores the importance of artificial intelligence tools for early detection of distress signals on social media. However, the results remain specific to the Reddit platform and the data used. Further validation on other networks and in different languages would be necessary to generalize these findings.

Media Sources

Reference Document

DOI: https://doi.org/10.1186/s43093-026-00867-w

Title: A unified comparative evaluation of machine learning, deep learning and GPT-2 for suicide ideation detection from social media

Journal: Future Business Journal

Publisher: Springer Science and Business Media LLC

Authors: Yasmeen Mohamed Saleh; Fahad Kamal Alsheref; Mahmoud Mohamed Bahloul