SIAgentNews.com
← All news
DevelopingGeminiResearch

Google's Gemini Decision 1 fails to detect 2025 AI voices

A new study shows that Google's multimodal decision model can no longer distinguish modern AI-generated speech from human recordings.

· First reported by echo.ls

Ink and watercolour: picture for the story 'Google's Gemini Decision 1 fails to detect 2025 AI voices'
AI-made ink and watercolour for this story. No real people or products are shown.
0.52AUC for 2025-2026 AI generators

The short version

  • Gemini Decision 1 identifies pre-2025 AI voices with an AUC of 0.67, but drops to 0.52 for 2025-2026 models.
  • The model shows a strong bias toward labeling audio as human-made, correctly identifying only 20 out of 1,020 AI clips.
  • The model failed to recognize its own 'Gemini 3.1 Flash TTS' voice with an AUC of only 0.58.
  • Human listeners also struggled with recent models like Cartesia Sonic-3, which had a 32% correct identification rate.

Why it matters

This research demonstrates that AI audio generation is evolving faster than the models designed to police it. For users, it means that identifying deepfakes or synthetic content is becoming increasingly difficult for both humans and automated systems.

The detection gap

Google released Gemini Decision 1 on November 11, 2026, as a multimodal decision model. This type of model is designed to make judgments based on various types of data, including audio. Researchers recently tested its ability to distinguish between human speech and synthetic AI voices.

The results show that the model is capable of identifying older AI voices with some success. When tested against generators from 2023 and 2024, the model achieved an AUC of 0.67. This metric, known as Area Under the Curve, measures how well a model can separate two different categories of data.

However, the model's performance falls off a cliff when facing newer technology. For generators released in 2025 and 2026, the AUC dropped to 0.52. An AUC of 0.50 represents a random guess, meaning the model is effectively unable to tell the difference between modern AI and humans.

A bias toward humanity

The study reveals that Gemini Decision 1 has a significant internal bias. Instead of being neutral, the model tends to label almost everything it hears as human-made. Out of a total of 1,020 AI clips provided in the test, the model only assigned a 50% or higher chance of being AI to 20 of them.

This suggests that the model is not just failing to see the AI; it is actively leaning toward human labels. While it can sometimes spot older synthetic voices, its default stance is to assume the audio is authentic. This makes it less reliable as a tool for filtering out deepfakes.

The researchers noted that the drop in performance is not a gradual decline. The data shows a sharp break at the end of 2024 or the very beginning of 2025. This timeline suggests that the underlying technology for AI voice generation underwent a major leap forward during that period.

Comparing man and machine

The research also looked at how humans compare to the AI model in these tests. In a 2026 listening study involving 35,532 real-or-fake guesses from 1,768 people, humans were often fooled by the same generators that tricked Gemini Decision 1. This indicates a shared struggle for both biological and artificial intelligence in identifying modern fakes.

Interestingly, some signals that humans can detect, the AI model cannot. For example, a signal-processing method from 1984 was recognized by humans 72% of the time, but the model only achieved an AUC of 0.53. This means the model misses certain technical artifacts that human ears can still pick up.

The model also struggled with its own technology. It achieved an AUC of only 0.58 when trying to recognize Gemini 3.1 Flash TTS. This suggests that even the creators of a model may find it difficult to distinguish their own synthetic voices from real human speech.

Testing for metadata

To ensure the results were accurate, researchers had to rule out simple tricks. They wanted to make sure the model wasn't just looking at hidden data tags, known as metadata. If the model were just reading a label that said 'AI-generated,' the results would be skewed.

The team used a rigorous cleaning process on all audio clips before testing. They re-encoded the files using ffmpeg to strip out metadata and ensured all clips were normalized for loudness. They also limited the audio to a specific frequency range to keep the testing conditions consistent across all generators.

By removing these external factors, the study confirmed that the model's failure is a result of the audio quality itself. The synthetic voices are now so high-quality that they lack the distinct markers the model relies on. This creates a significant hurdle for developing reliable automated detection tools.

The future of detection

The findings highlight a growing arms race between voice generation and detection. As generators like Cartesia Sonic-3 become harder for humans to identify, the tools meant to catch them are falling behind. This creates a landscape where synthetic content can be deployed with high levels of realism.

The sharp decline in performance starting in 2025 suggests that the industry is moving toward a point of 'indistinguishability.' When both humans and models perform at chance levels, it becomes nearly impossible to verify the source of a voice clip using current methods.

Moving forward, researchers will need to find new ways to identify synthetic audio. Since current models are failing to catch recent outputs, the focus may shift toward more complex multimodal cues or different types of behavioral analysis to spot fakes.

What to watch

  • Whether new models can be trained specifically on 2025-2026 data to recover detection rates.
  • If human detection rates continue to drop as synthetic audio quality improves.

Try it

Look for the 'Cartesia Sonic-3' voice generator online to see how difficult it is for humans to distinguish from real speech.

Sources and links

First reported by echo.ls. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news