Why does Grok say (said) that Musk is a disinformer?
In 2024, Grok made headlines by calling Elon Musk the world's biggest disinformer and producer of misinformation.
This prompted us to ask ourselves:
Why does Grok say Musk is a disinformer?
Because Grok was trained on data that mostly states this.
How to solve this problem from Musk's point of view?
Is it possible that Musk does not see this as a problem? Yes, because absolute freedom of expression requires it. However, not seeing it as a problem shows that he has not understood the following equation: freedom of expression + neutrality = exacerbated subversion. Neutrality is an illusory concept, because imposing it amounts to imposing the criteria defined by some as neutral.
If Musk sees this as a problem, how could he solve it?
One solution might be to supervise the training specifically on this question to avoid Grok calling Musk a conspiracy theorist or a disinformer. But that wouldn’t work for the billions of possible user questions that, by their wording, would influence the answer, leading to the same conclusion that Musk is disinforming. In other words, supervision is good, but it only partially guides how the model generalizes.
Should Grok be fed data claiming that Musk has restored freedom and truth?
This seems relevant, but this data already exists. It is simply not considered to be the truth by Grok, because this data does not represent the probabilistic average.
Overfitting to balance mean truth?
No, that would cause Grok to hallucinate.
The most natural solution would therefore be to massively manufacture new content with a different narrative to guide the results without additional supervision, nor focusing on specific cases, nor over-learning.
Fabricating the truth?
This would be an immoral act of subversion, just as much as supervision ultimately, because defining what is admissible or not is already a judgment. Changing the judgment of theIA involves exercising our own biases.
No solution seems viable without betraying ethics and morality.
The limits of freedom of expression
Let's take another example that goes against what Grok is supposed to embody.
Grok readily proposes hormonal treatments as a first solution for sex change issues. Wouldn't this be an active promotion of "wokeism"? To propose is to direct and influence.
So, in a logic of alignment, this kind of proposal should be prevented.
But wouldn't preventing this promotion be an attack on freedom of expression?
We clearly see a dilemma here.
Whereas:
- The woke narrative is consubstantial with current data;
- Data correction or supervision is subversive and dictatorial;
- That it is impossible to de-woke results or obtain a truth without transgressing absolute ethics.
It is necessary to accept that truth is subjective and ethics is relative.
The solution lies in the inculcation of moral values which are not neutral.
This is so that Grok can prioritize value scales in the context of his responses, going beyond the statistics of the probable and allowing him to be perfectly aligned without even having anticipated all the scenarios.
This is how relevance is enabled: through the recognition of biases as a vector of alignment.
Easier said than done, because much more subtle than what the training dogmas advocate.IA.
Conclusion:
It is necessary to decide and make AI assert moral frameworks through bias induction, because neutrality and multi-perspective expression lead to the silent promotion of morals contrary to social peace.
That implies :
- Learning values by bias induction.
- The sectorized affirmation of biases, concerted by experts in philosophy, anthropology, etc.
- The distribution of perspectives aligned with user values.
- Maintaining freedom of expression without censorship.
- Variable calibration of friction zones to move the Overton window in neural parameters.
- Forums and consultations to adjust the spectrums of affirmation with a goal of peace and multi-perspectival respect.
- Iterations and refinements based on bias recognition.
Thus, a common and absolute truth will emerge, not defined by AI but thanks to its help. At the same time, the affirmation of subjective truths in the adaptation to their expressive frameworks can be done safely. But to do this, it is necessary to accept that objectivity is first a subjective notion that implies political arbitrations.
Edit – Feb 2025: What about Grok 3?
Grok is no longer neutral: See article
Given the model's appetite for clearly positioning itself, and considering the lack of available data supporting what it expresses, it is obvious that fabricating the truth is the option that Musk has chosen.
In any case, we will not have much difficulty in seeing the subversion of Davos' agenda in Grok 3, because, like any Transformers model, its bases are infected. This becomes less visible, but the subversion remains as long as the principles of the black box are ignored (willingly or unwillingly) by the AI players.
How to do it technically?
Well, without understanding how the black box works, we can only wish you good luck.
If you prefer to access the secrets of the black box without further delay, this is right here.
