Does removing unfair bias in AI applications entail promoting fairness?

Short reflection on the following sentence from the Ethics Guidelines for Trustworthy AI of the High-Level Expert Group on Artificial Intelligence, set up by the European Commission.

If unfair biases can be avoided, AI systems could even increase societal fairness. (p. 12)

If an AI system is trained on biased data and makes recommendations or decisions based on that biased data, this can sustain or even increase bias in society (the latter happens especially when the biased decisions are used as support for further biased decisions, either by humans or as new training data, and thus create feedback loops).

If instead you find methods of ridding bias from AI-systems (either by countering bias in the training data or in some other way), then potentially these systems can promote fairness in society and maybe even leverage above-mentioned feedback loops for the better (fairer decision-making is a precedent for more fairer decision-making ).

But to me it seems that this depends on where the bias comes from. The bias in the data may be caused because the data is not a representative sample. It may also be that the data is biased because it is a representative sample of a biased society. In the latter case, a unbiased AI system would hypothetically do better than biased human decision-making in a biased society. In this case, methods for removing unfair bias ( a negative problem formulation ) would end up promoting fairness.