Only the quantum processes of the very small can produce high fidelity copies of DNA – the building blocks of life. But, even though there is an extremely low error rate of around one in one billion, it is not perfect, errors DO happen.
These errors are what has allowed life to adapt over hundreds of millions of years. Historical knowledge is codified in DNA – human and otherwise – and forms the basis for life on Earth, as we know it.
But what does this have to do with algorithms.
Biases defined in the past will be unavoidably coded into algorithms, it cannot happen any other way – without a time machine. The algorithm must learn – if left unchecked then the biases will continue ad infinitum. They must also unlearn as posited in this MIT Post. Algorithm authors must facilitate this learning by injecting errors to replicate adaptability. In the same way as enzymatic replication of DNA has.
Unfortunately, the indicative timescales that nature has provided us are not very human friendly, since;
Opponents of an algorithmic world cite these biases of the past as abhorrent. They they must be removed at all costs, NOW! The binary worlds of computing and mathematics evoke ideals of perfection, immediately. Particularly notable when current diversity aspirations are using data from 20th century society. Simply not tenable!
Proponents assert that mathematics and computing can improve on error rates found in the squishy world of life. But this is to avoid the point. Adaptation takes a long time to test, and exists on evolutionary timescales. Adapting too fast is mutation and often leads to death.
From either viewpoint above the notion of perfection is wrong. Whatever algorithms humans (or machines) produce will be imperfect. Things will improve, given more time and more data.
We need to strive for acceptable imperfection.