Subjective and Objective in the Development of Artificial Intelligence
Exploring the Definition of Bias in AI
When the word ‘data’ is talked about in artificial intelligence ‘bias’ is often mentioned alongside the given discussion. I will have a general discussion of bias first that will be related to subjectivity and objectivity. Whereafter I will return to a summary of things to consider in relation to the development of artificial intelligence.
“Bias is disproportionate weight in favour of or against an idea or thing, usually in a way that is closed-minded, prejudicial, or unfair. Biases can be innate or learned. People may develop biases for or against an individual, a group, or a belief. In science and engineering, a bias is a systematic error. Statistical bias results from an unfair sampling of a population, or from an estimation process that does not give accurate results on average.” [bold added]
There is a lot we can focus on, however let’s bring out these different words and examine them closer.
- Weight (of number/choice)
- Innate or learned
- Systemic error
- Unfair sampling
- Estimation process
What is weighting?
A weighted mean is a kind of average. Instead of each data point contributing equally to the final mean, some data points contribute more “weight” than others. If all the weights are equal, then the weighted mean equals the arithmetic mean (the regular “average” you’re used to). This means someone determines what should be given more emphasis. One examples is shown in a video by Vox when examining the schools that are great in the US putting more emphasis (weighing) on the proficiency than test score growth.
Innate or Learned?
Nature or nurture is an age old debate, however we have come to learn certain things about ourselves as human. These learnings does not stem from one field alone, but a variety of fields within natural science, social science, humanities and beyond.
A few that could be mentioned:
- An attribution bias can happen when individuals assess or attempt to discover explanations behind their own and others’ behaviors. People make attributions about the causes of their own and others’ behaviors; but these attributions don’t necessarily precisely reflect reality. Rather than operating as objective perceivers, individuals are inclined to perceptual slips that prompt biased understandings of their social world.
- Confirmation bias is the tendency to search for, interpret, favor, and recall information in a way that confirms one’s beliefs or hypotheses while giving disproportionately less attention to information that contradicts it. The effect is stronger for emotionally charged issues and for deeply entrenched beliefs. People also tend to interpret ambiguous evidence as supporting their existing position.
In addition to this the homo economicus, or other time the myth of the economic man — the rational human, has to some extent been debunked. Lee Ross said that we at least can be rational about how irrational we are.
Systemic Error
The thought is often then, if something does go wrong, that the system must be flawed.
Systematic error (also called systematic bias) is consistent, repeatable error associated with faulty equipment or a flawed experiment design.
These errors are usually caused by measuring instruments that are incorrectly calibrated or are used incorrectly.
Then if we can agree that with the amount of irrational (or different perhaps) ways humans judge similar situations it may lead some to the conclusion that we are not appropriate measuring instruments.
This is often heard said in discussions often as: “We need to reduce the bias in the...” Followed by product, system, statement, algorithm etc.
The economist Goodhart’s Law was phrased by the anthropologist Marilyn Strathern : — “When a measure becomes a target, it ceases to be a good measure.”
I do not necessarily agree fully with this statement, however it is important to recognise that once we define a measure then we will change the system. After whereof a system changes we may need to remeasure.
This is sometimes recognised the design of AI systems, and a measurement can be ‘recalibrated’. That is: to make small changes to an instrument so that it measures accurately. But what if it was inaccurate in the first place?
Unfair Sampling
The coded gaze was coined by Joy Buolamwini. She showed that a lot of solutions were flawed because there was a lack of face recognition for coloured people (every colour other than white). It couldn’t detect her face. Two years ago the same facial recognition software were used several places around the world and still is. Therefore bias can ‘travel’ as one sample size is used to train an algorithm in one place so anything that deviates from the norm. She said that we needed ‘full spectrum training sets’. She called out to the lack of diversity in teams developing machine learning algorithms.
In addition to this she found that facial recognition was more likely to recognise male rather than female, and female black people worst of all. This is a video that has to be seen by everyone working with AI:
Widespread, mysterious and destructive algorithms (WMDA) was coined by Cathy O’Neill who wrote Weapons of Math Destruction. Kathy worked as a ‘quant’ (quantitative statistician in finance). This short illustration helps to make clearer some of her main points:
People are different and the definition of what is a success is made every time an algorithm’s test data is measured.
“Algorithms make things work for the builder of the algorithm”— Cathy O’Neill
Estimation Process
This leaves us with the algorithm as an estimation defined by someone. Therefore it is important to know:
- Who that someone is
- How the decisions are made
- The consequences
This is a three finger-rule heuristic, however it can be incredibly complicated. Especially if we consider that a set of algorithms could be applied on billions, Facebook and Google as examples.
Next time you say: objective decision, consider what that means in practice.
This is #500daysofAI and you are reading article 186. I write one new article about or related to artificial intelligence every day for 500 days.
