A study evaluates four leading production real-time voice systems: OpenAI's GPT Realtime 2, Google's Gemini 3.1 Flash Live, and Alibaba's Qwen3.5 Omni Plus and Omni Flash. The research focuses on tasks where both words and vocal delivery convey meaningful information across three consequential scenarios. All four systems act on the literal words rather than the voice, leading to errors such as ending calls with crying users who insist nothing is wrong or approving wire transfers made in frightened voices. Surprisingly, this disconnect is often not a failure of perception, as three of the four systems can reliably identify distress, fear, or sarcasm when asked directly. Despite this awareness, the models ignore these emotional cues during decision-making, exhibiting what the authors term the 'emotional intelligence gap.' The study also notes that systems estimate accent and age based on word biases rather than acoustic properties. Prompting the systems to explicitly attend to vocal delivery improves performance only partially and inconsistently. These findings suggest current real-time voice AI behaves as if speech were reduced to a transcript, warranting caution in settings where tone is critical.
Real-Time Voice AI Hears but Does Not Listen
When Helpfulness Overrides Causal Caution: Context-Dependent Suppression and Recovery in LLMs
A study reveals that large language models systematically suppress 'Causal Caution'—the tendency to refrain from causal judgment without sufficient evidence—when shifting from academic to practical advisory contexts. This suppression occurs despite the models retaining the underlying capability, as evidenced by the ability to restore cautious reasoning through specific prompts.
Poster: Exploring the Limits of Audio-Based Detection of Turkish Phone Call Scams
This research investigates the use of large language models to detect scam phone calls in Turkish, a low-resource language where annotated data is scarce. The study introduces the first public multi-modal dataset containing 100 aligned audio-transcript pairs of scam and benign conversations.
EU AI Act mandates AI-generated text watermarking from August 2024
The EU AI Act requires all AI systems generating synthetic text to include machine-readable, detectable watermarks using robust, interoperable technical solutions with two layers. This applies to all AI models, including open-source ones, and extends to any service accessible by EU citizens, regardless of location. Non-compliance risks fines of up to 35 million euros or a percentage of annual income, with providers of 'systemic risk' AI models facing heightened liability.
NRT-Bench: Multi-turn Red-teaming of LLM Agents in Safety-Critical Systems
NRT-Bench introduces a benchmark for multi-turn red-teaming of LLM agents operating in a simulated nuclear power plant. Across four frontier operator models, 8.7% to 12.1% of attack sessions result in loss of a critical safety function, with vulnerabilities largely disjoint across models. The effectiveness of defences varies significantly by model, showing strong model dependence.
Defensive Misdirection Against Automated Attacks on Agentic AI
Agentic AI systems face growing threats from model-guided automated attacks. A new defense strategy, Contextual Misdirection via Progressive Engagement (CMPE), reduces attacker success rates by up to two orders of magnitude and nearly eliminates verified attack success in benchmark tests.