Take a simple question like: has one of your parents been to jail in the past two years? with the answers being 'yes' or 'no.' The context that is missing is why did they go to jail, what for/what type of crime? The context of this answer will likely play a role in the future acts or behaviors of the person that is currently being sentenced but the system doesn't take this into account.

There is never going to be a perfect system and so people shouldn't lean heavily on these systems but use them as a single data point in a collection of data points they are using to help make decisions.

It is an example of sampling bias in data, similar to that demonstrated in the 'gender shades' example from the MIT researcher. It could also be an example of inaccurate labelling. This produced the troubling result of all black faces 'appearing the same' to the facial recognition technology.

What I found equally troubling about the article is the apparent lack of questioning about potential motives or alibis before the arrest, indicating the blind trust and faith placed in the system. The willingness of law enforcement to trust the system indicates racist undertones and reflects broader societal inequalities.

Yes, the controversies surrounding the policing systems mentioned in the image are closely linked to algorithmic bias. Many modern policing systems utilize data-driven algorithms to predict criminal activity and allocate resources based on historical data pertaining to specific areas or groups. This approach may inadvertently exacerbate existing biases.

For instance, these systems might over-police certain neighborhoods predominantly inhabited by minority groups while neglecting other areas that may equally require attention. This exemplifies a typical issue of algorithmic bias, particularly when models are constructed on unbalanced or biased training datasets, potentially amplifying such unfair phenomena.

What types of bias are involved? According to Barocas and Selbst's classification, several types of biases may be relevant here:

Historical Bias: When a model relies on datasets imbued with historical prejudices, injustices related to race, gender, etc., can persist through the algorithm. For example, certain groups may have been disproportionately monitored or arrested due to racial profiling in history; consequently, algorithms built upon such data continue labeling these groups as high-risk individuals or areas.

Representation Bias: Certain populations may be either overrepresented or underrepresented within the dataset leading to inaccurate predictions for those groups by the algorithm. For instance, higher arrest rates in particular regions do not necessarily indicate elevated crime rates but rather reflect concentrated law enforcement efforts in those areas which inflate risk assessments for minority communities.

Measurement Bias: The metrics or variables employed within an algorithm might not be equitable. Arrest records or crime reports could be influenced by racial prejudice resulting in misleadingly high crime rates attributed to specific ethnicities; this leads algorithms towards unjust predictions.

Reflecting on target assessment: By integrating Baracas and Selbst's analysis into our understanding, we recognize that there is a crucial need for vigilance against introducing biases during system design and for ongoing iterations and improvements aimed at preventing systemic injustices from arising.

 #DataEthicsMOOC @UoE_EFI

the answers as of why they provide me those results with the generic prompts are that the data is feed with how the media and other outlets represent a high rank executive.

The issue is not the technology, we just need to keep refining the data to be more representative of the society we want. With policing the problem starts when we assume that all individual in certain demographic or ethnic group are criminals, just like the image generators assume that all high rank executives are white males.

discretion, and the future of policing. New York,

NY: Oxford University Press.

Jefferson, Brian Jordan (2020). Digitize and

punish. Racial criminalization in the digital

age. Minneapolis: University of Minnesota

Press.

A significant part of the ‘welfare robots’ controversy involves algorithmic bias. AI systems in welfare often rely on historical data to make decisions, such as determining eligibility for benefits or identifying potential fraud. If this data reflects existing inequalities or systemic biases (e.g., socioeconomic disparities, racial discrimination, or geographical inequities), the algorithm may perpetuate or even exacerbate these biases. This is particularly concerning in welfare systems, where marginalized groups are disproportionately affected.

Statistical Bias arises if the training data is unrepresentative or incomplete. For example, if the algorithm is trained on data that does not reflect the diversity of welfare recipients, it may systematically disadvantage certain groups.

The type of bias involved is discriminatory bias, specifically historical bias and measurement bias. Historical bias occurs because these algorithms are often trained on historical data that reflects past inequalities and discriminatory practices, such as racial profiling or disproportionate sentencing of minority groups. Measurement bias may arise when the data used to train these algorithms includes biased variables, like arrest records or prior convictions, which can unfairly target certain demographics.

As a result, these biases can perpetuate unfair outcomes in policing and sentencing, with certain groups being unfairly over-policed or over-sentenced.

With little information on the case, it's hard to know which type of bias was involved. What was most striking was the fact the police offers trusted the results of the AI solution, without exercising any critical thinking (maybe because they were happy with what came out of it?).

But we could easily think the solution was biased, in multiple ways. (1) Training data: we know that some populations have been discriminated by the police in the past. Using historical data sets as they are is likely to lead to an over representation of those populations in people to watch out for. (2) Feature selection: the AI algorithms often use data sets that are linked to where people live, their social economic status, etc. Some of these data sets are proxies for ethnicity. Eventually, it can lead to ethnicity discrimination.

Yes it most probably is, because the algorithms guiding the data processing will have been skewed certain ways in how they come to their conclusions. But other types of bias are involved including historical bias, measurement bias (including bias by proxy, labelling bias) and very importantly the particular attitude of the human ultimately making the decision on that day. Because in the case of benefits processing in the UK, which is the article I reviewed, the human still has the final say in the process. Rubrics can only get us so far in making a less biased decision, and humans will always be biased (as will the machines they program). What we can do is be aware of our bias, be aware of the bias we are programming in to tech, train it on more representative (and ethically sourced) data sets to reduce this bias, and where relevant add counter-bias (as companies like Google have done with their AI, sometimes to negative results).

The controversy revolves around algorithmic bias, and multiple biases are involved. There is historical, sampling, measurement, labeling and social biases involved. This is a very nuanced and multilayered issue involving data, design and human nature and society. Technical awareness, ethical reflection, diverse voices and accountability are needed to address these issues. 

2) Training data / Data Collection: Police have used historical crime records which happen historically in neighbourhoods where crime commonly happened, thus skewing the data. Simply, police have no more data to set as input in the model

3) Proxies: mainly because the student is of black race, some proxy relating to another characteristic may have caused him to fall into a high potential criminal zone, given that the AI tools are biased against Black or Asian races

Far and away the worst aspect of predictive policing is individual risk assessments where target variables (like, for example, recidivism) are determined by class labels that mask clear bias. The example often cited here is the question "Have your parents ever been arrested?" In the US, you're far more likely to answer that if you're poor, black, or a minority. The result? You're considered likely to offend or reoffend, and you're watched more carefully than someone who may actually be at greater risk to reoffend.

Systems that are meant to spread policing and resources more efficiently are simply exacerbating existing problems. It's sad, and quite simply, terrifying.