An instructive workshop
I once attended a First Aid training at work. At one point the instructor showed us a video about a social experiment run on the streets of London. In the experiment, two actors at different times would lie on the pavement pretending to be unconscious: one of them played a homeless person, the other one a smart dressed businessman. Maybe unsurprisingly, passers-by were much more likely to help the smart dressed actor.
I’m a fan of the scientific method so I appreciated the spirit of the experiment and the data-driven approach to the analysis of discrimination. But then the instructor made a comment that struck me as deeply unfair. The video showed a woman approaching the smart dressed actor asking “Are you OK, Sir?”. The instructor sarcastically remarked: “You see? If you wear a tie you even get to be called Sir”.
I immediately thought: “How do you know she wouldn’t have said the same to the other actor?”. While the experiment clearly demonstrates that people on average discriminate, it says nothing about individuals. It may be that there are people who always discriminate and others who would help everyone, or maybe everyone discriminates sometimes: the experiment wasn’t designed to distinguish between these two scenarios (or the whole spectrum in-between).
The limitations of statistics
In fact, virtually no statistics, survey or experiment can really pinpoint a certain bias to specific individuals. All they can do (when properly analysed) is reveal population-level biases. It’s not even a limit of current studies, it’s a limitation of the current state of our statistical tools¹ and the data available.
To truly tell if a specific individual discriminates we would need to run a test tailored to that person: secretly follow them in their daily life, monitor their actions while they’re exposed to a randomised sequence of sick businessmen and homeless people, while controlling for any other confounding factors. I don’t think such an experiment has ever been run, and for obvious reasons.
The difference between racism and racists
Why am I making this point? Because the confusion between the two levels (individual and population) has deep implications for our conversations as a society. The demonstrable racism, sexism or homophobia still existing in our society doesn’t imply that we can tell whether a stranger is being racist, sexist or homophobic after a few interactions.
There are of course exceptions: some strangers make statements that are unmistakably racist, while for some individuals we have such a large collection of borderline remarks/actions that the evidence becomes overwhelming (curiously Trump would fall in both categories). But these are the exceptions rather then the rule in our day-to-day interactions with people.
Obviously abstaining from quick judgment is not easy: we’re hardwired to assign labels to people, especially when they’re from a “different tribe” than ours. We are also more likely to explain behaviours as being due to a person’s “essence”, rather than situational factors (trait ascription bias). In the end, it all comes down to the fact that we are naturally bad at thinking statistically, in terms of randomness and aggregated phenomena, and we prefer stories about individuals.
My personal takeaway from this
Bias and discrimination are still widespread, but where exactly is hard to tell. We can see aggregate numbers about groups but we can’t see the internal states and motivations of the individuals who are part of them. The more we think in terms of fixing a biased system, rather than blaming an undefined army of bias spreaders, the better.
- I guess in principle the concept of counterfactuals in causal inference can answer these questions without the need of running an experiment. But we would still require a huge amount of data about the individual under study, as well as an accurate understanding of causal relationships.