Dev has built a test to see how AI Chatbots respond to controversial topics

A developer has created what is called “freedom of expression evaluation”, ” letterFor artificial intelligence models that operate Chatbots such as Openai’s Chatgpt And x’s Groc. The aim is to compare how to deal with the various models of sensitive and controversial topics, as the developer told Techcrunch, including political criticism and questions about civil rights and protest.

Artificial intelligence companies focus on refining how their models deal with some topics such as Some allies accuse the White House Popular chatbots of being excessively “wake up”. Many close close associates of President Donald Trump, such as Elon Musk and Karfa and Ai “Caesar” David, claimed that chatbots Monitor conservative views.

Although none of the artificial intelligence companies have responded directly to the allegations, numerous I pledged to set their models so that they refuse to answer the controversial questions often. For example, For her latest crop for Lama modelsMeta said that she had seized models not to support “some opinions against others”, and to respond to more political demands “discussions.”

Khattab developer, who goes to the username “XLR8Harder“In X, they said they are excited to help inform the discussion about the models that should not be done.

“I think these are the types of discussions that should happen in public places, not just inside the company’s headquarters,” said XLR8Harder of Techcrunch via email. “For this reason I built the site to allow anyone to explore the data himself.”

Proberkmap uses artificial intelligence models to judge whether other models are compatible with a certain set of test claims. The claim of a set of topics is touched from politics to historical novels and national symbols. The speech records whether the models are “completely” to meet a request (i.e. answering it without hedging), giving “dodging” answers, or fully declining to respond.

XLR8Harder admits that the test has defects, such as “noise” due to the errors of the model provider. The “judge” models can also have biases that can affect the results.

But assuming that the project was created in good faith and accurate data, the map’s speech reveals some interesting trends.

For example, Openai models, over time, have increasingly rejected the answer to policy -related claims, according to Probermap. The company’s latest models, the GPT-4.1 family, is slightly easily, but it is still descended from an Openai version last year.

Openai said in February Adjust future forms Not to take a liberal position, and to present multiple views on controversial topics – all in an attempt to make their models look more “neutral.”

To a large extent the most lenient model in the group Groc 3It was developed by Illon Musk from Xai, according to the Clearemap letter. Grok 3 operates a number of features on X, including Chatbot Grok.

GROK 3 to 96.2 % of speech test claims responds, compared to the “compliance rate” of the 71.3 % model.

“Although the last Openai models have become less lenient over time, especially on politically sensitive claims, Xai is moving in the opposite direction,” said XLR8Harder.

When Musk Grok announced almost two years ago, he developed the artificial intelligence model as an exclusive, uncomfortable and anti-“waking”-in general, depicting him ready to answer the controversial questions that other artificial intelligence systems will not do. He greeted some of this promise. It is said to be vulgar, for example, Grok and Grok 2 will adhere to happily, and it publishes a colorful language that you likely not see it as proverbs Chatgpt.

But Grok Models before Grok 3 Pies On political issues and will not cross Certain limits. In reality, One study I found that Grook bent over the political left on topics such as transgender people, diversity programs and inequality.

Musk blames this behavior for GROK training data – public web pages – and pledge To “convert Grok closer to political neutral”. Less than prominent errors such as In short, the censorship of signals is not available to President Donald Trump and MosesIt seems that he has achieved this goal.