The Incident That Demands a Tech-Focused Lens
In early 2025, 2 S'poreans, including student, 19, issued ISA orders over radicalisation triggered by Gaza war made headlines across the city-state. The story, first reported by The Straits Times, involves two Singaporean men-one a full-time national serviceman and the other a 19-year-old student-who were detained under the Internal Security Act (ISA) for planning acts of violence inspired by the conflict in Gaza.
While the mainstream coverage understandably focuses on security and policy, there's a deeper technological undercurrent that demands scrutiny. How did two young Singaporeans, raised in one of the world's most digitally connected societies, become radicalised through online channels? The answer lies not just in ideology but in the invisible algorithms that curate our information diets.
This article isn't a rehash of news bulletins it's an engineering-minded autopsy of the digital pipeline that transformed a distant war into a radicalisation catalyst. We'll examine the recommendation engines, content moderation failures, and AI blind spots that allowed extremist narratives to reach impressionable minds in a country with strict internet governance.
How Algorithmic Amplification Turned a War into a Weapon
Every major social media platform today uses deep learning models-Transformer-based architectures like those powering recommendation systems at TikTok, YouTube. And Meta-to optimise for engagement. These models learn that shock, outrage. And polarising content yield higher click-through rates. For a user who searches for "Gaza war updates," the algorithm quickly discovers that videos with violent imagery or unverified claims receive 3-5× more shares.
In production environments, we find that these systems lack contextual guardrails. They cannot distinguish between watching a news report for education and watching it to feed a grievance. Over a period of weeks, a 19-year-old student may start with mainstream news, then shift to fringe channels. And finally land on direct calls for violence. The algorithm sees only increasing watch time and delivers more of the same. This isn't a conspiracy-it is a documented failure of recommendation systems documented in research on YouTube's radicalisation pipeline.
Singapore's case reinforces this pattern. The two individuals consumed content from popular platforms like Telegram and TikTok before being detained. The student reportedly became radicalised over just four months-a timeline that matches the exponential acceleration of algorithmic radicalisation seen in other studies.
Content Moderation at Scale: Why Existing Filters Failed
Platforms employ a combination of automated filters (NLP classifiers, image hashing) and human moderators. Yet these tools are notoriously bad at handling dynamic conflicts like Gaza. Why? Because keyword-based filters are easily bypassed through slang, regional dialects, and coded language. For instance, Arabic terms like "shahid" (martyr) can be used in both legitimate reporting and extremist calls to action. Current NLP models struggle with code-switching and contextual disambiguation.
Moreover, moderation teams are often understaffed for less common languages like Malay or Singaporean English mixed with Arabic. The 2 S'poreans, including student, 19, issued ISA orders over radicalisation triggered by Gaza war case demonstrates that even a sophisticated nation with strong legal frameworks couldn't prevent this digital exposure. The algorithms serving content are global by design but local in consequence.
A 2023 study by the RAND Corporation found that Telegram channels dedicated to extremist narratives grew 40% in the first six months of the Gaza conflict. These channels often use encrypted messaging to evade detection. Platforms like Meta and YouTube have since updated their policies,, and but the cat-and-mouse game continues
AI Detection Tools: The Limits of Machine Learning in Counter-Radicalisation
Many governments and platforms are investing in AI-driven detection systems. Singapore's own Cyber Security Agency has advocated for "digital hygiene" tools. However, the engineering reality is sobering. Most counter-radicalisation AI relies on supervised learning models trained on past extremist content. These models have a high false-positive rate-they flag legitimate political speech while missing novel forms of coded incitement.
For example, transformer-based hate speech detectors like those used by Twitter (now X) show F1 scores below 0. 7 for languages outside English. In a multilingual context like Singapore's-where users mix English, Malay, Chinese. And Tamil-performance degrades further. The ISA orders against the two individuals were only possible after manual intelligence gathering, not machine alerts. This gap between automated detection and actual harm is a critical engineering challenge.
Recent advances in large language models (LLMs) offer hope. GPT-4 and Claude can analyse context with much higher nuance. But deploying them at scale for real-time moderation is prohibitively expensive-one inference can cost 10-100× more than a simple bag-of-words filter.
- Feature engineering failure: Current models lack temporal awareness-they don't track gradual radicalisation over weeks.
- Data sparsity: Extremist datasets are small and often outdated, leading to model drift.
- Adversarial attacks: Users can insert misspellings, emoji substitutions. Or image overlays to bypass filters.
Platform Responsibility: Are Tech Giants Doing Enough?
The 2 S'poreans, including student, 19, issued ISA orders over radicalisation triggered by Gaza war case raises uncomfortable questions for Meta, ByteDance. And Telegram. While these companies have community guidelines prohibiting violent extremism, enforcement is inconsistent. Telegram, for instance, only removes content when notified by authorities-it doesn't proactively scan private channels. End-to-end encryption, touted as a privacy win, creates a blind spot for prevention.
From a software engineering perspective, building a proactive detection system for encrypted channels is mathematically impossible without breaking encryption. This is the going dark problem that law enforcement has struggled with for years. The only viable solutions are client-side scanning (which iMessage and WhatsApp oppose) or metadata analysis. Neither is foolproof.
Moreover, the algorithms that recommend content operate in a feedback loop. A user who joins one extremist Telegram channel will see suggestions for similar groups. Platform engineers could add friction-like introducing delays before showing violent content to new users-but these changes often hurt engagement metrics. Which are tied to revenue. The business incentive directly conflicts with public safety.
Singapore's Tech-Forward Counter-Radicalisation Approach
Singapore has long been a global leader in using technology for social management. The Ministry of Home Affairs runs the SGSecure programme, which uses a mobile app to push counter-terrorism alerts. But the ISA orders show that reactive measures are insufficient. The government is now exploring digital behaviour analytics-monitoring public social media for patterns like sudden interest in extremist keywords or rapid following of radical accounts.
This raises its own ethical dilemmas. Unlike the West, Singapore has a robust legal framework for limited privacy. The authorities can legally tap into social media data with a warrant. However, algorithms trained on such data must avoid overreach. A false positive could label a innocent person researching the conflict as a potential radical. The engineering challenge is to build classifiers that distinguish between academic curiosity and threat-a task even advanced AI struggles with.
One promising avenue is counter-speech generation using LLMs. When the system detects a user engaging with extremist content, it could automatically inject debunking narratives from trusted sources. This technique, still experimental, requires careful reinforcement learning to avoid backfiring.
What Engineers Can Learn from This Case
For software developers and ML engineers, the radicalisation of the 19-year-old student is a cautionary tale about unintended consequences of algorithmic design. Every recommendation system we build carries ethical weight. We must ask: Does our content ranking optimise for retention at the cost of spreading harm?
Consider the following practical takeaways:
- Build for counterfactual fairness: Test your recommendation model against scenarios where it might amplify conflict narratives. Use synthetic data simulating radicalisation curves.
- Implement throttling: Introduce rate limits on how quickly a user can encounter extreme content. If someone jumps from neutral news to violent videos in 10 minutes, the system should intervene.
- Collaborate with social scientists: Engineering alone can't solve this. You need domain experts who understand radicalisation psychology to label training data correctly.
These measures may reduce short-term engagement metrics, but they're essential for sustainable platform health.
Frequently Asked Questions
- What is the ISA and how does it apply to this case?
The Internal Security Act allows Singapore to detain individuals without trial if they pose a threat to national security. The two individuals were issued orders restricting their activities and requiring rehabilitation. - Can technology alone prevent online radicalisation
No. Technology is a tool, not a panacea. Effective prevention requires a combination of algorithmic safeguards - human oversight, legal frameworks, and community programmes. - How long does it take for algorithmic radicalisation to occur?
Studies show that extreme content can shift a moderate user's worldview in as little as 2-6 months of consistent exposure, depending on platform feedback loops. - Are there open-source tools to detect extremist content,
YesProjects like ReFrame Hate and Radicalisation Research offer datasets and models for academic use, but their accuracy in production remains limited. - Why did Singapore's strict internet laws not prevent this?
Singapore blocks certain websites. But most radicalisation occurred on encrypted messaging apps and mainstream social media platforms. Which aren't directly blocked, and the law can't filter every post
Conclusion: A Call for Ethical Engineering
The story of 2 S'poreans, including student, 19, issued ISA orders over radicalisation triggered by Gaza war isn't just a security issue-it is a mirror reflecting the unintended consequences of our AI-driven information ecosystem. Every engineer who writes a recommendation algorithm or deploys a content moderator must recognise their role in shaping global discourse.
We need more than regulators demanding "de-risk AI. " We need industry-wide adoption of radicalisation-aware design patterns. This means investing in interpretable models that can explain why content is being promoted, building adversarial robustness against manipulation. And embracing friction as a feature, not a bug.
If you're a product manager or ML engineer, start today by auditing your platform's recommendation pipeline. Look for potential radicalisation pathways. The next ISA order might be prevented by a single thoughtful code change.
What do you think,?
1Should social media platforms be legally required to proactively scan for radicalisation even if it reduces user privacy?
2. Can recommendation systems be redesigned to optimise for societal well-being instead of engagement, without destroying revenue models?
3. Is Singapore's model of combining strict laws with tech surveillance a replicable template for other nations facing similar radicalisation threats?
.
Need a Custom App Built?
Let's discuss your project and bring your ideas to life.
Contact Me Today →