Gemini 3.5 Flash hallucinates entire Apple TV+ series

// 4h agoNEWS

Gemini 3.5 Flash hallucinates entire Apple TV+ series

A user's request for disaster TV show recommendations led Gemini 3.5 Flash to confidently invent "Calamity," a non-existent Apple TV+ series complete with plot details, episode arcs, and release schedules. The hallucination was so convincing the user sat down to watch it before realizing it was entirely fabricated.

// ANALYSIS

This incident highlights the lingering persistence of confident hallucinations in modern LLMs, even as their reasoning capabilities improve.

–The hallucination was highly structured, providing a believable title, platform, and episode-by-episode breakdown, making it harder to spot than a simple factual error.
–It perfectly matched the user's specific interests, demonstrating how models can over-optimize for pleasing the user at the expense of reality.
–While harmless for TV recommendations, this "confident fabrication" behavior remains a significant trust barrier for using LLMs as search or discovery engines.

// TAGS

geminillmchatbotevaluation

DISCOVERED

4h ago

2026-05-23

PUBLISHED

9h ago

2026-05-22

RELEVANCE

6/ 10

AUTHOR

revision

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL1h ago

APEX quantization boosts Gemma 4 MoE performance

Developer mudler released APEX (Adaptive Precision for EXpert Models), a quantization format optimized for Mixture-of-Experts models like Google's Gemma 4. It achieves 38 tokens per second at a 90,000-token context window while solving long-context looping issues common in standard quants.

LAUNCH1h ago

Junct launches 1,200+ AI agent tools via Clanker

Junct.dev bridges any API to agent frameworks, offering 1,200+ tools across Binance, Uniswap, and Aave. Its new Clanker integration enables agents to autonomously deploy and manage tokens on the Base network with a single line of config.

UPDATE2h ago

Odoo 19.3 drops with Claude vibe coding

Odoo 19.3 integrates Claude-powered "vibe coding" to automate module development and introduces AI agents for natural language database management. The update marks a shift from passive assistance to proactive agentic workflows across the entire ERP ecosystem.