I don't believe that. In a sense, it's like the classic NRA defense, that guns don't kill people, etc.
In reality, certain technologies also come with specific tendencies -- ascribed in their very design and formulation.
In the very least, certain technologies within specific societal contexts tend to have specific outcomes. Which makes placing the blame to "people using that infrastructure" problematic, because those people also operate within that societal context, and are influenced by it, they are not totally free agents.
I didn't mean to put all the blame on the users of the infrastructure. Actually I was thinking about the case where the user is involved in creating the infrastructure in the first place, and then pulls the "It's an algorithm!" excuse.
If the infrastructure is made available for someone else to use, then part of the blame for abuse also lies with the person who decided to let them use it. But even in that case it's not the infrastructure that is the problem, it is the lack of access control and accountability.
You may be interested in "Fairness in Machine Learning" (https://fairmlclass.github.io/)
From what little I know - In the same way that vision systems pick out edge and feature detection without being explicitly told what those things are, the kind of algos you're talking about could pick up on biases... and just like DeepDream shows how you can then investigate the features encoded in the system, maybe you could then investigate the biases.
In other words, if these kinds of models encode the existing social biases, it also means those biases are now somewhere very, very examinable.
We could attribute that impact to:
1. Military leaders - but their hands were forced, their only choice was to deploy machine guns or see their troops mown down by them.
2. Gun inventors - but surely if Richard Gatling hadn't invented the Gatling Gun, someone else would have. It was an idea whose time has come.
3. "The system" - but isn't that an unsatisfying cop-out? Doesn't this let the above claim rewards for the upsides of their choices while disclaiming responsibility for the downsides, and by so doing creating perverse incentives?
When a machine learning model determines that according to the available data, blacks are more likely to default on loans, it's not unfairly biased against blacks. It is merely reflecting that the world is unfairly biased against blacks.
The fault inevitably lies with the decision-makers who look at an algorithmic prediction and decide to maximize their profit. Because they confuse what is with what ought to be.