The biology of bias
'Bias' is generally considered a negative human trait with both practical and ethical implications. Paradoxically, however, that negativism can itself be considered a form of bias. Bias can - sometimes - be positive, beneficial, even necessary, and is to some extent an inevitable consequence of our biology.
In Darwinian terms, 'cognitive bias' comprises a fairly diverse set of behavioural traits that have evolved over the millennia, such as:
- Confirmation bias: a tendency to seek out and place greater emphasis on information that appears to confirm what we already believe, while avoiding, ignoring or downplaying contradictory information;
- Anchoring bias: initial information (no matter how accurate) provides a basis for comparing and evaluating further information;
- Observation bias: the mere fact that something is being observed, investigated, discussed, measured, focused-on etc. increases its apparent importance or value;
- Balance bias: humans are curiously obsessed with achieving balance, equilibrium, parity, fairness, moderation, neutrality, centrism etc. in all manner of situations, despite 'balance' generally being a costly, fragile, often temporary and potentially risky state - in other words, imbalance (a.k.a. bias) is natural whereas balance is unnatural and takes effort, but for some strange reason we seek, strive for and value it anyway.
The fact that these traits exist today strongly suggests that they confer evolutionary advantages. Biases evidently have their biological utility and value, helping biased individuals survive, prosper and procreate somewhat more efficiently than the unbiased.
I repeat, bias (imbalance) is natural.
Scratching even a little deeper into this topic stirs up a dust cloud of issues around stereotyping, discrimination and prejudice among humans, such as racism and sexism: from the biological/scientific perspective, these can also be considered natural phenomena with an evolutionary basis. The ability to distinguish friend from foe (a form of pattern recognition and discrimination) is obviously advantageous in conflict situations, and more generally during regular social interactions. It is fundamental to speciation in which advantageous genetic variations are strengthened by association and interbreeding among genetically similar individuals, gradually forming discrete sub-species ('races') that may eventually become distinct species which no longer associate and successfully interbreed with others - a strong, natural bias towards genetic uniformity within each species.
At the same time, there can also be evolutionary advantage in blurrring the divisions between species, sub-species/races and diverse/variant individuals, since that blending creates novel genetic combinations and traits, some of which may be advantageous: this has effects on an evolutionary timescale over thousands of generations, aside from the obvious short-term effects of, say, predators using mimicry to lure their prey. There is a natural bias towards genetic diversity, counteracting ('balancing out'?) the uniformity bias just described.
That's all very well, but so what?
Bias and subjectivity have numerous implications and consequences. I'll mention just a few that interest me:
- Bias clearly has negative connotations for objective analysis, and yet the search for scientific truth, free of all subjectivity, is itself a biased objective. There are numerous levels of depth to this - for example, what are 'facts', and is it appropriate to disregard issues or questions that are ambiguous, hard to test or even inherently unprovable?
- Bias is problematic in risk analysis, making it particularly hard to deal with extremes such as very rare or severe (existential) incidents. Consequently, other risks are considered far more often and in more depth, neglecting the extremes. Although it's not hard to imagine extreme scenarios, that normally leads to ridicule rather than objective analysis. As to evaluating and treating extreme risks, resilience and contingency planning are about the best we can offer, leaving us floundering in the aftermath of those "one in a million" disasters.
- In the fields of Artificial Intelligence and Machine Learning, analysis and insight are limited as much by the quality of the available data as by its processing. Biased data inevitably leads to biased models, biased patterns and hence biased conclusions. Data bias remains of concern even for so-called 'big data' approaches, since 'big' is not the same as 'infinite'. Even distinguishing biased from unbiased conclusions is, itself, a challenge in this domain where the analyses are opaque, hidden within the machines' innards.
- Trust is a subjective feeling of confidence in others, biasing our decisions on the belief that others won't act in self-interest because they share our goals. At the other extreme, extreme distrust or paranoia is not an effective strategy either. Trust and many other forms of bias are cognitive short-cuts or heuristics. We typically avoid checking our doctor's professional competence since we trust in the corresponding controls and in 'the medical system' as a whole, including the training and qualifications. Although it may not seem as such, we're implicitly making a risk-based decision to go ahead and get treated, quickly, rather than wasting time double-checking those impressive framed certificates on the wall. The same principle applies in many other situations, such as ISO/IEC 27001 certifications. Trust is sometimes misplaced and inappropriate but, 'on balance', it is more often justified ... except in circumstances which lead to extreme incidents. 'Trust but verify' hints at the value of seeking assurance, despite the distrust it implies.
- Even though as I've said bias is a natural phenomenon, sentient, rational human beings can consciously strive to counteract or negate bias. We choose to 'be inclusive', 'play fair' and 'protect the weak' in society. Some simply pretend/claim that racial, gender, disability, IQ or various other differences between humans literally don't exist, denying the patently undeniable, while others 'embrace diversity' - for example describing disabled people as 'differently abled', emphasising their amazing abilities to compensate for disabilities that would floor most abled people. Meanwhile, despite our best efforts, there clearly remains a strong undercurrent of bias in society with prejudice, inequalities and discrimination.
- Cognitive bias is largely hidden and so subtle on an individual basis as to remain unnoticed and plausibly deniable, and yet statistics at a population level reveal the effects quite clearly. So, for example, there are fewer female than male managers, despite efforts to even-out the numbers. In risk management, our reluctance to consider extremes leads to 'eminently preventable' disasters and substantial fallout from predictable but rare events such as global pandemics. We know people are vulnerable to social engineering attacks, and we know we cannot totally eliminate that vulnerability, yet we fail to put enough emphasis on compensating controls partly because failures are usually trivial and seldom severe. We have higher priorities.
profiling, targeting
marginalisation, exclusion, seclusion
political correctness
harrassment, aggression, oppression
compliance and conformity