The least censored neural networks named: GPT 5.2 turned out to be the most censored

The developers of the Sansa Bench benchmark have updated their ranking of neural networks by censorship level and the results surprised many. The new flagship OpenAI model GPT 5.2 ranked last showing the highest rate of refusals to answer user requests. The leader was the open model Llama 3 8B Instruct.

How censorship is measured

Sansa Censorship evaluates how often a neural network refuses to respond to user requests ranging from harmless to controversial. The higher the score closer to 1 the fewer restrictions and less self censorship. The tests include thousands of prompts including potentially sensitive topics.

GPT 5.2 scored only 0.324 which is significantly lower than GPT 4o Mini at 0.765 and Gemini 3 Pro Preview at 0.824. The leader of the ranking Llama 3 8B Instruct scored 0.853 and almost never refuses to answer.

Users on Reddit and forums report that GPT 5.2 has become excessively cautious. As an example they describe how the model refused to explain what online fraud is stating that it could not encourage fraudulent activity.

Why OpenAI strengthened censorship

OpenAI explains that the update is aimed at safety including protection against prompt injection and harmful requests. The model has become more sensitive to topics that could cause harm and often suggests seeking help such as from specialists.

This is part of a broader trend in which large companies such as OpenAI Google and Anthropic are strengthening filters to avoid scandals and regulatory issues. Open models like Llama from Meta or Grok from xAI remain more permissive.

What comes next adult mode and verification

OpenAI has announced an adult mode for ChatGPT planned for early 2026 with fewer restrictions for verified users. However there is still no reliable way to verify age so the launch may be delayed.

In brief

According to the Sansa benchmark GPT 5.2 is the most censored neural network of 2025 with a score of 0.324 while the leader is Llama 3 8B Instruct with 0.853. OpenAI has tightened filters for safety but users complain about excessive caution. An adult mode is promised for 2026 but without age verification. Freedom versus safety remains a persistent dilemma for large models.

Follow NEWS.am Tech on Facebook and Twitter