AI Toxicity Harder to Fake Than Intelligence, 2025 Study Reveals

Researchers surprised that with AI, toxicity is harder to fake than intelligence

A 2025 study led by Nicolò Pagan at the University of Zurich found that AI-generated text retains detectable emotional cues, making it easier to identify than previously assumed. The research highlights that even calibrated large language models (LLMs) fail to replicate human affective tone.

Why This Matters

The ideal of fully human-like AI remains unmet, as technical realities reveal persistent gaps in emotional authenticity. This has critical security implications: attackers using AI for impersonation or phishing could be exposed through subtle linguistic tells. The cost of misidentifying AI-generated content is rising, with 2025 reports showing a 37% increase in AI-based social engineering attempts.

Key Insights

• “AI-generated text remains distinguishable from human text in affective tone, 2025 study”
• “X Chat’s static encryption keys risk full conversation decryption if compromised”
• “OWASP expands Top 10 to include supply chain and dependency risks in software ecosystems”

Practical Applications

• Use Case: Detecting AI-generated phishing emails via emotional cue analysis
• Pitfall: Relying on static encryption keys (e.g., X Chat) exposes historical and future messages to decryption

References:

• https://dev.to/ziizium/security-news-weekly-round-up-14th-november-2025-2ngm