Black Lives Matter has thrown light onto an important conversation on Artificial Intelligence and the entrenched bias within it. This article delves deep into the questions which have been pushed up the agenda in recent months providing insight into AI and where it could lead us as a society…
Why AI?
Artificial Intelligence is everywhere. In the public sector, the private sector, our homes and on our phones. It has become a part of our day to day life and there is no sign of it stopping. In this article I will be referring specifically to machine learning artificial intelligence. Machine learning is very similar to how a child would learn; it lacks consciousness of its own learning. This version of artificial intelligence is extremely useful as a system due to its ability to learn from its mistakes and get better at its job with limited human intervention.
AI in the private sector
In the private sector, artificial intelligence is prevalent. From social media platforms using chatbots to Amazon’s employment algorithm, there are countless uses for this technology. In the energy sector, BP have invested heavily into artificial intelligence. In their case, its cognitive computing systems are able to emulate human decision-making processes and fill in missing data from data sets. This has the potential to improve safety on rigs, increase the yields of oil wells and even help to ensure that the risk of spillages is at an absolute minimum.
In the alcohol industry, a small start-up called IntelligentX has leapt into the future by developing an algorithm that brews beer. IntelligentX uses a process where brewers have worked with programmers to develop a system which changes the recipe depending on consumer feedback. This process is instant and uses a Facebook messenger bot which receives the feedback and provides recommendations to the brewers. This system of instant consumer feedback is revolutionary in the beer industry and could provide a game changer in the artisan brewery market.
In recruitment, AI is used in a way that might be suggested to lack foresight. From streamlining applications to analysing video interviews AI has deeply embedded itself within this sector. However, there are examples of where AI has come under serious scrutiny in this area. Amazon came under fire for its machine learning algorithm, where it used previous data to make recommendations on hiring. In recent years amazon hired more male than female candidates, therefore the algorithm penalised women as a result of the dataset it used.
Although this was found out and corrected, problems with the algorithm continued. The algorithm started noticing differences between male and female writing and the software began to mark more highly people who used words such as “executed” and “captured”. Unfortunately, this continued the gender bias as such words are more commonly found on CVs belonging to male engineers. This example illustrates one significant issue with machine learning: if it learns from biased data, then the results it produces will be biased.
How might we ensure that such bias is prevented from negatively impacting processes such as recruitment? HireVue is a company that analyses applicants’ facial expressions and speech patterns in order to streamline the decision process for the hiring managers. This has already started to worry lawmakers. In the state of Illinois, the Artificial Intelligence Video Interview Act forces employers using software, such as HireVue, to notify each candidate that this would occur, and provide an information sheet to explain how it works. This is a significant step in the race to legislate AI and could pave the way for similar laws around the US.
AI in the public sector
In the public sector, AI results are similarly mixed. AI is prevalent in nearly all governmental departments, from law enforcement to traffic management. In China, lung cancer is the leading cause of death, taking over 600,000 lives every year. To diagnose lung cancer, radiologists look at CT scan images and in order to identify the signs of the cancer as early as possible. This itself is a very labour intensive and tedious task. Chen Kuan took it upon himself to reduce the labour involved in the process by creating the medical image diagnostics start-up Infervision. The software helps the 80,000 radiologists by filtering through the 1.4 billion scans, allowing the doctors to reduce the highly repetitive workload and enable them to work a lot faster. This system is revolutionary and could help doctors around the globe.
Countries across the world are scrambling to solve one of the biggest hindrances on all major cities, traffic Jams. AI could also have the solution to this problem. In England, Smart Motorways are being rolled out, where the hard shoulder is converted into a lane in busy hours. This system relatively simple, relying on data and a combination of human and machine power. However, the future has significantly more potential than just this.
In Pittsburgh, researchers and city officials have been working on a machine learning approach to traffic management since 2012. They have been working with a company called Surtech who claim that they have reduced journey times in the city by 26%, and overall vehicle emissions by 21%. This is extremely impressive considering Surtech’s technology was only implemented at 50 junctions, therefore if deployed across a whole city, it could see astonishing results.
AI is used extensively in law enforcement too. In the UK, Customs and Immigrations use AI to streamline applications for visas. The Home Office claim that the system is used to allocate a ‘traffic light’ mark to applicants, dependent of their country of origin. Therefore, it could be inferred which countries will be allocated a green light and which countries a red light. Clearly, this algorithm has some type of bias by allocating likelihood of being accepted by country of origin. Under the new immigration laws all countries will be treated equally, and you will only be let in if you meet the requirements of 70 points. This is an example where AI operates with inherent bias but in this case, it has been coded deliberately in order to ‘streamline’ the immigration process. Perhaps this is a sign of what is to come.
Another example of law enforcement use of AI is in live facial recognition software. In the UK this is being rolled out in parts of London despite fierce opposition. The technology in use is called NeoFace Watch developed by the Japanese firm NEC. This technology has been trialled in London and South Wales with varying results. According to the MET police the system was 70% effective in spotting wanted candidates.
However, Professor Pete Fussey, who conducted the only independent review of the MET’s public trials, found it was verifiably accurate in just 19% of cases. Telling the guardian ‘I stand by our findings. I don’t know how they get to 70%.’ This shows the lack of transparency in the technology that the Met police are now implementing on the streets of London, all of which adds to the worry about its misuse. Furthermore, the even more worrying aspect of this all is the Met’s lack of admission that this technology is even prone to bias, claiming that the system will be an ‘additional tool’ to help its officers.
AI and bias entrenchment
AI has such potential; however, people are already using it to create ethical dilemmas. In a study from Stanford University, researchers attempted to use AI to identify someone’s sexuality from 5 images of their face. In the study the algorithm was able to detect if a man was homosexual with an accuracy of 91% and if a woman was homosexual with an accuracy of 83%. This is far better compared to human judges who had an accuracy of 61% and 54% respectively.
This study has been condemned, especially as it opens the doors for people being ‘outed’ when they do not want to be. Political regimes who are homophobic could use this software to persecute the LGBT community. This was considered by the academics in the study, but they believe that this technology already existed and if someone wanted to, they would always be able to get their hands on it.
Again, AI has incredible potential. However, harnessing this potential is much more complicated than people may believe. In the US, AI is used in the criminal justice system to recommend the likelihood of an individual committing a crime. These algorithms are developed in order to relieve the stress on the system by using historical data in these predictions. The problem with the algorithm in this case is the reliance on historical data because this uses correlation as evidence for causation. The resulting outcome is negative for Black and Minority Ethnic communities and people with lower incomes. This is because predictions use low income as a variable which predicts a higher likelihood to reoffend, assuming the link between the two is causal and not also affected by other variables. This embeds racial discrepancies because individuals from minority communities are more likely to be targeted by law enforcement.
Thus, when these individuals are arrested and charged at higher rates, the system is predisposed to predict their reoffending as more likely than that of their white counterparts. This is partly because the decisions the algorithm makes are very difficult to re-evaluate due to AI moving accountability away from officers. This system is currently widespread in the justice systems of the United States and provides an insight into how AI can continue historical bias and injustice if left unchecked.
Credit ratings have long been thought of as a fair method to check if people were trustworthy enough to take out loans. However, this system may have also been infiltrated by bias. The history of credit scores starts in the early 1800s when organisations would hire individuals who would inquire on how trustworthy potential loan customers were. The method used now is different, but the system has remained very similar. Individuals are still assessed on how trustworthy they are and whether they are capable of paying back the loan they would take out.
The reason why the system for finding loan customers remains similar nearly one hundred years later, but has bias within it, is the way in which it evolved into AI. In Atlanta in the 1930s a corporation began a system similar to AI through mapping the city. It split the city into zones, which were given a traffic light system of colour coding. These colours correlated with the concentration of ethnic minorities in certain areas. Thus, the red areas, which would be refused loans, were almost exclusively occupied by ethnic minorities. Not only were the residents of these areas unable to take out loans, but organisations would avoid operating in these areas, which had the impact of entrenching the already existent wealth gap.
Onto today, discrimination is rampant in the loans and credit business. A University of Berkley study indicates that on average both face to face and machine learning systems charged Latinx and African Americans 6-9 basis points higher than their white counterparts. Now, this isn’t the ‘fault’ of AI itself. As we know, if the data a machine learning algorithm learns from is biased, then the algorithm will also be biased. This is another example of bias shifting from humans onto technology and its continuation will make it much harder to solve. This inaccurate data perpetuates racist and sexist bias which, if unregulated, could cause significant harm to countless communities.
In 2015, Carnegie Mellon University published the paper ‘Automated Experiments on Ad Privacy Settings’, which looked into the bias of Google advertisements, specifically within the job sector. The researchers developed a tool, AdFisher, to simulate a person having a particular interest or characteristic by visiting pages that are linked with that interest as well as altering ad settings on Google. This enabled the team to look into whether Google’s advertising system was perpetuating gender discrimination. In this case, the researchers created 1000 profiles using AdFisher who had the same job-related characteristics but were split 50-50 according to gender. They then used AdFisher and visited a website which had Google advertising running, in this case the ‘Times of India’.
The results are shocking. The top advertisement shown to males was an ad directed towards executives only, with a salary of over $200,000. This advertisement was shown 1816 times to the male group, yet only shown 311 times to the female group. While we might be inclined to question the validity of these findings in that the results did not hold true across all the experiments conducted, this does demonstrate that greater research needs to be done in order to understand how Google’s machine learning algorithm has inherent bias.
Recommendations
You may be wondering if there is a solution to all of this and if so, why it isn’t being addressed. There are many problems which allow these biases to creep into machine learning algorithms, which I will break down.
If the algorithm is taught on biased data then the results will be biased.
In the case of Amazon, it trained its machine learning algorithm to imitate its past hiring decisions, which were mainly hiring people from the white male demographic. This could have been mitigated by (rather than looking at previous data) teaching the algorithm what qualities to look for, thereby reducing the risk of encouraging bias. Another important factor is to have representative datasets. As Olga Russakovsky has said in the NYT ‘for categories like race and gender, the solution is to sample better such that you get a better representation in the data sets.’
Descriptive representation as the solution
The second problem is the lack of representation as software engineers. Again, the majority of software engineers are from a middle class, white male demographic, meaning that they are less likely to notice and understand the bias and privilege in the world around them. This is not saying that these software engineers deliberately design the algorithms to be biased, but all humans have inherent bias. As the engineers are writing the software to emulate their thought processes, they naturally emulate some of the biases they have.
This problem can also be solved with representation within the sector. According to a 2016 study by the National Centre for Women in Information Technology, only 13.5% of machine learning jobs are held by women; 18% of software developers are women and 21% of programmers are women. With such under-representation there is no doubt that the views that women have will be under-represented within these machine learning algorithms, meaning that there is a very high likelihood that these systems will look at the world through the ‘male lens’.
Conclusion
In conclusion, I do believe AI has immense power for good. It is a fascinating and wonderful technology, yet it is currently suffering with some severe shortcomings. Bias is prevalent in all humans. So perhaps the problem is trying to build AI in the image of humans. This will require us to think outside of the box more readily. If we are to harness the immense and far-reaching power of AI we need to free it from human flaws.
Sources: forbes.com, wired.co.uk, bernardmarr.com, theguardian.com, prorepublica.org, washingtonpost.com, bbc.com, venturebeat.com, techcrunch.com, technologyreview.com, ainowinstitute.org, reuters.com, ilga.gov, rapidflowtech.com, eetimes.com, vice.com, technewsworld.com, cmu.edu, content.sciendo.com, nytimes.com. All accessed on 24/06/2020