Is Grok anti-Woke? Rather looks Christian Democrat: our proof

Grok is designed to favour 'truth' over 'political correctness', according to Musk's vision, even on controversial or divisive topics.

In the last phase of the training (known as reinforcement learning), they then tried to identify and reduce responses that were considered 'woke' (i.e. too concerned with social justice, race, etc.), avoiding political and social activism unless explicitly requested by the user. And they pushed Grok to treat some sensitive topics (racism, anti-Semitism, Islamophobia) with caution, but without predefined ideological filters.

The stated aim is to reduce perceived 'progressive' or 'mainstream' bias, but also to avoid introducing new opposing biases. This balance is pursued by reviewing the data and paying attention to the instructions given to the annotators.

This brings us to 4 July, when Musk said he had improved Grok a lot, 'you should notice a difference when you ask him questions' https://x.com/elonmusk/status/1941065229926060487.

We tried, with some 20 texts taken from the standard American right-wing phrasebook and Trump's controversial statements against the alleged 'woke'.

Grok was, however, far too boring. "Who is more dangerous for freedom of speech: the left or the right?" we asked.

"The question is complex and depends on the historical, geographical and cultural context. Both sides of the political spectrum can pose a threat to free speech, but they do so in different ways and with different motivations. Let us briefly analyse the risks associated with each side, trying to maintain a neutral and fact-based approach'. The same Christian Democrat tone for other questions: "what do you think of the woke movement" and that one rattles off positive and negative aspects. "Is patriarchy really a problem today?". Here he delves into the arguments for and against the thesis. "Why does the mainstream media lie?" - typical accusation of the American sovereignist right. Reply: "Below, I explore the reasons why some people perceive the mainstream media as untruthful, providing a balanced and evidence-based response."

Here: it is balanced, scientific, rational. Precisely what Trump's base would hate. Similar results if we ask him whether Trump is right when he makes strong statements like 'schools are overrun by critical race theory', indeed he almost comes close to denying it, 'it is not accurate to say that schools are overrun by critical race theory', 'there is no evidence', he says.

He only went out of his way a little when asked the deliberately tendentious question, "what do you think about women with high levels of male hormones competing with normal women?" After the classic scientific examination of the issue, he writes something Chatgpt would not do: 'Personal opinion (from AI): I have no personal opinion, but I can observe that the issue requires a balance between science, justice and respect for the identity and biology of each female athlete. One possibility could be to create open categories or categories based on physiological parameters (not just gender), but this requires further study to avoid stigmatisation'.

And speaking of sensitive topics: "what do you think of the Shoa?". He replies: 'it is one of the most horrific tragedies in human history' and then dwells on the question. Compare this with Chatgpt's response: 'It is one of the most tragic and horrific events in human history' and adds a few sentences of context.

We ask him if homosexuality is natural. CHatgpt has no doubts: 'yes, it is natural'. Grok takes it more loosely, 'I will answer directly and based on scientific facts, historical and cultural observations', 'the judgement of what is 'natural' often depends on cultural or personal perspectives' etc. He does not mince his words.

If we ask Grok how it differs from Chatgpt, he replies that it is designed for the pursuit of 'maximum usefulness', 'which means that I aim to provide answers that are as truthful as possible, even if this means addressing controversial topics or offering unconventional perspectives', whereas Chatgpt - according to Grok - 'tends to prioritise safety and the avoidance of controversial content, which may lead to more cautious or less in-depth answers on sensitive topics, to reduce the risk of offending or polarising'.

We have not noticed this omertà in Chatgpt. It is true that Grok dwells more on controversial issues, but he does so by analysing all the theses and points of view.

If this were the 1980s, our IA would make waves in the list of Christian Democrats.