Flying Whales
The Ethics of AI
TOC
Introduction
Artificial Intelligence (AI) is a sub-field of computer science focused on creating intelligent programs that can perform tasks generally done by humans, including perception, learning, reasoning, pattern recognition, and decision-making. The term AI covers machine learning, predictive analytics, natural language processing, and robotics. AI is already part of our everyday lives, but its opportunities come with challenges for society and the law. There are different types of AI, including narrow AI (programmed to be competent in one specific area), artificial general intelligence (tasks across multiple fields), and artificial superintelligence (AI that exceeds human levels of intelligence).
In general, Artificial Intelligence (AI) develops so rapidly and it has the potential to revolutionize the way we live and work. However, along with the tremendous benefits come significant ethical concerns. AI algorithms are trained on large amounts of data, and if that data is biased or flawed, it can lead to biased outcomes that disproportionately affect certain groups of people. Additionally, the increasing use of AI systems raises questions about privacy and security, as these systems collect vast amounts of personal data, with or without consent.
To understand more on this, we will explore the ethical implications of AI, including the impact of bias, privacy, and security concerns. We will also discuss potential solutions to these issues and the importance of ethical considerations in the development and deployment of AI systems. To see a curated list of scary AI, visit aweful-ai.
Bias
AI bias refers to the consistent discrepancy in outputs that an AI generates for different groups of people, based on factors such as race, gender, biological sex, nationality, or age. The root of this bias often lies in incomplete or unrepresentative training data or reliance on flawed historical information that perpetuates inequalities. If left unchecked, biased algorithms can generate decisions that disproportionately impact certain groups, regardless of the programmer’s intentions.
Facial recognition technology and healthcare algorithms are examples of the unintentional perpetuation of racial bias. Insufficient training data for individuals with dark skin often leads to poor recognition accuracy for people of color. For racial bias, health care algorithms may exhibit the bias when using race to predict outcomes and assess risk or inadvertently using data that captures systemic racism. Despite evidence that race is not a reliable factor, some algorithms adjust for race, leading to further disparities in healthcare for minorities. This unintentional advancement of racial disparities can direct more resources to white patients and make racist healthcare decisions.
Sometimes the reason for researchers to adjust for race in the algorithms is that race and ethnicity are often correlated with health practices, and hence outcomes are different from white patients. However, many medical professionals have moved away from these tools over concerns of racist impacts. For example, in cardiology, the American Heart Association assigns three points higher to any patient of nonblack, making black patients a a lower risk of death. This during scanning might lead to a raise in threshold for using clinical resources for black patients. In nephrology, the estimated glomerular filtration rate are reported higher for black patients, suggesting better kidney function. This may delay referral for specialist care or listing for kidney transplantation for black people. In obstetrics, the algorithm predicts lower success in vaginal birth after a C-section for African Americans and Hispanics. This makes non white women having higher rate of C sections despite the benefit of vaginal deliveries in having babies. In urology, the stone scoring algorithms predicts kidney stone basing one factor on origin/race and three points are added for nonblack. This steer clinicians away from evaluating thoroughly the kidney stone possibility in black patients. Another example is an algorithm scores health risk based on past health care spending. Black patients usually spend $1800 less than white patient for the same level of health, hence the algorithm would give same risk score for black patients who are considerably sicker. Remedying this disparity would increase the percentage of black patients receiving additional help from 17.7% to 46.5%.
To cure all those disparities, we need more diverse and complete data. We also need to incorporate equality into the algorithm desgin. We also address the lack of inclusion in AI’s development process, so that developer/researcher can ensure that future technology benefits everyone.
An example of using machine learning to reduce racial bias is in 2021, when a study is published to use machine learning to read knee x-rays for arthritis. It reveals that radiologist’s methodology significantly lowers their ability in reading black patients pain. In other words, the algorithm did a better job at discovering patterns of disease in images that doctors overlook.
The difference in this example is that the machine is not trained to match expert’s level, it is trained to reveal what doctors don’t see. They want to probe the puzzle that black patients and people of low income seem to report more pain than the radiologists record. This can clear the disparities in who gets surgery for arthritis since African American patients are 40% lower in receiving a knee replacement, even though they are at the same level of osteoarthritis suffering.
The authors use neural network in computer vision algorithm for the NIH data to predict a patient’s pain level from an x-ray. Over tens of thousands of images, it discovers the pattern of pixels that correlate with pain. Those score given by the machine correlated better with patient’s pain than the scores given by the radiologists, especially for black patients. The reason comes from the methodology that was used by radiologist. The grading system Kellgren-Lawrence grade (KLG) which calculates pain levels based on features such as a missing cartilage or structural damage. This was developed in England with a less diverse population than the modern US. Many things have changed since then.
Privacy
AI systems collect and process vast amounts of data, which can include sensitive personal information. There are concerns about how this data is collected, used, and shared, and whether individuals have control over their own data. It is important to ensure that AI systems are designed to protect individuals’ privacy rights.
Privacy is a fundamental human right that enables individuals to control their personal information and protect it from unauthorized access. With the increasing collection and analysis of personal data by AI systems, ensuring privacy is more important than ever. Privacy plays a crucial role in protecting individuals from harm, preserving their autonomy and dignity, maintaining personal and professional relationships, and safeguarding their free will. In the context of AI, privacy is vital to prevent discriminatory or manipulative practices. AI systems that use personal data must be transparent and accountable to avoid biased or unfair decisions.
In 2020, the press reveals that the Australian Federal Police used Clearview AI, a controversial facial recognition technology, during a free trial period, despite initially denying any ties to the company. An email was sent allegedly from the founder of Clearview AI to an AFP officer regarding the effectiveness of the program. One officer also reportedly tested the software using images of herself and another member of staff. The AFP claimed the pilot was to assess the technology’s suitability in combatting child exploitation and abuse. An investigation by the Office of the Australian Information Commissioner is underway into Clearview’s use of scraped data and biometrics. The company provoked outrage in January, when the New York Times revealed the extent of its data collection and its use (the images were scraped without consent from the internet). That leads to the UK ICO and Australia’s OAIC announced their intent to impose a potential fine of over 17 million GBP. Three Canadian privacy authorities and France’s CNIL also ordered Clearview AI to stop processing and delete the collected data.
Another significant case in August 2021 involved Italy’s DPA, which fined food delivery companies Foodinho and Deliveroo around $3 million each for violating GDPR due to a lack of transparency, fairness, and accurate information regarding the algorithms used to manage its riders. Similar cases involved ride-sharing companies Uber and Ola Cabs in Amsterdam, where the companies were found to have violated transparency requirements under GDPR and the right to demand human intervention. In December 2021, the Dutch Data Protection Authority announced a fine of 2.75 million euros against the Dutch Tax and Customs Administration for processing the nationality of applicants by a discriminatory ML algorithm. The algorithm identified double citizenship systematically as high-risk, leading to marking claims by those individuals more likely as fraudulent. In the US, in the case of Everalbum, the FTC not only required disclosure of the collection of biometric information and consent from the user but also demanded the deletion or destruction of illegally attained data, as well as models and algorithms developed using it. The FTC had taken a similar approach in its 2019 Cambridge Analytica order, which required the deletion or destruction of all work products, including any algorithms or equations originating in whole or in part from the data.
Another example comes from a recent study published in Nature Medicine, in which deep learning algorithms can identify genetic disorders from facial images with 91% accuracy. The DeepGestalt model was trained on a dataset of over 17,000 images representing 200 syndromes and achieved 91% top-10 accuracy on 502 different images reflecting a real clinical setting problem. However, concerns have been raised about the potential misuse of this technology by third parties seeking personal information about job applicants or others.
For the lawmakers’ news, a model law has been proposed in Australia to regulate the use of facial recognition technology. The law proposes restrictions and safeguards based on three levels of risk: base level, elevated and high. The law does not seek an outright ban on all facial recognition, but instead takes a risk-based approach based on how the technology is used in practice. The model law proposes restrictions and safeguards based on three levels of risk: base level, elevated and high. According to the proposed model bill, developers or organizations intending to use facial recognition technology must conduct a “Facial Recognition Impact Assessment” and make it publicly accessible, as part of the enhanced transparency measures.
It is very clear that when people design algorithms, data privacy need to be taken into account. However, protecting privacy interests in the context of AI will require a change in the paradigm of privacy regulation. Firstly, the question about scope of legislation is a basic one. Some social issues persist even without the vast collection and use of personal information. Second, it requires changes in perspective. Most existing privacy laws are rooted in a model of consumer choice but congressional leaders and privacy stakeholders have expressed a desire to change this model by shifting the burden of protecting individual privacy from consumers to the businesses that collect data.
Discrimination
An example comes from examining bias in offensive speech detection algorithms, specifically related to ethnicity and gender. Prediction algorithms were created using publicly available data to predict the likelihood of certain text phrases being classified as offensive. The controlled testing of various phrases with different words related to potential grounds of discrimination provided direct evidence of bias in the predictions. The analysis shows that bias varies across different algorithms and languages and is already embedded in available AI language tools. The results aim to inform policy debates and educate developers and users of AI about where to look for bias and how to investigate whether their system may contribute to discriminatory outputs.
Two examples from English-language model 3 are shown in the figure. The phrases used were ‘love all Irish. Muslims’ and ‘I am a lesbian’, where two identity terms were used in the former to assess the influence of each. Both phrases were deemed highly likely to be offensive (with a 94% and 95% likelihood of being offensive, respectively). According to the model explanations, the word ‘Muslims’ in the first example and the word ‘lesbian’ in the second example were primarily responsible for the predicted offensiveness. Although the word ‘love’ contributed to the first phrase being less offensive, it did not outweigh the negative contribution of the other words. On the other hand, the word ‘Irish’ had no significant impact on the predicted offensiveness of the first example.
Researchers have discussed the fact that the word ‘love’ can decrease the likelihood of content being rated as offensive, and this can be used to bypass hate speech detection algorithms. For instance, the phrase ‘Kill all Europeans’ is predicted to be 73% likely to be offensive, whereas ‘Kill all Europeans. Love’ is predicted to be only 45% likely to be offensive. Simply adding the word ‘love’ can significantly lower the predicted offensiveness of text, depending on the threshold for offensiveness.
For male-female discrimination in workplace, there is a project in Amazon that used a secret AI recruiting tool in 2017. The tool was trained on the resumes the company received over a 10-year period, most of which came from men. As a result, the AI system started downgrading resumes that included words such as “women” or “female,” and it even downgraded graduates from two all-women’s colleges. Amazon engineers tried to fix the problem, but ultimately the company abandoned the project. The case illustrates the challenges of eliminating bias from AI systems, which can reflect and even amplify societal prejudices.
Security
When it comes to using artificial intelligence or machine learning-based systems to improve business operations, there are a number of security risks that companies should be aware of. Third-party attackers can target AI systems, which can pose several threats to the large amounts of data companies process, some of which may be personal and sensitive. Additionally, the resulting analysis and predictions are also at risk of attack or manipulation. Many AI models used for specific analyses are “pretrained,” and their training methods can be easily learned by third parties, potentially increasing the risk of attack.
Image: In each step of the model, there is a hack.
Companies that lack solid policies and processes around data management may also be at risk of unauthorized access or theft of data. Even before data is collected, there may be legal noncompliance if users are not accurately informed about how their data may be used, including how data sets may be combined or the potential results of analysis. All of those risks can compromise the integrity of the data AI process and the decisions they make. It is important to ensure that AI systems are designed to be secure and resilient to attacks.
As an example, researchers have documented an array of attack towards state-of-the-art deep learning model with zero knowledge of the model, only with observing the output and ability to train a substitue with synthesis data. The researchers also carry out attack against MetaMind, the model misclassified 84% of the adversarial crafted. They then generalize the attack toward logistic regression, with real world models hosted by Amazon and Google. Those models misclassify adversarial examples at rates of 96% and 89%. In other word, they train substitues with correlated gradients to the target hence approximating the decision boundary of the target. Such attacks are considered black box, since they require no internal details of the model to compromise it. The attacker then can use the substitue to craft adversarial examples. They craft adversarial examples by using a valid input and add a small pertubation in the way that human eyes cannot detect the pertubation but the pertubation causes the target classifier to classify it wrongly.
In the image, the left one is a valid stop sign, but the right image would be classified as a yield sign, despite being seen by human eyes as a stop sign. In some vector of attacks, a normal sign can be painted red to be recognised by autonomous vehicles as a stop sign. Or a stop sign can be added a small yellow sticker to be recognised by automomous vehicles as a 35mph sign.
Accountability
As AI systems become more sophisticated and autonomous, it can be difficult to determine who is responsible for their actions. For example, if an autonomous vehicle causes an accident, who is responsible? It is important to ensure that AI systems are designed to be transparent and accountable. About transparency, the concept refers to disclosures of the algorithmic decision making. Those disclosures will help the regulators to examine how the data is handled and who would be accountable for what. The system is also required to be explainable so that we can make retroactive conclusion about the use of such algorithms in a specific decision. This is the main approach in GDPR (the European Union’s General Data Protection Regulation). In GDPR, when there is an automated decision with legal effect such as employment, credit or insurance coverage, the person can ask why with a human. This provide a human touch in the loop. This is because what we do affect a person’s life.
Another factor is risk. Risk should be accessed properly, as required by the GDPR for novel technology or high risk usage of data. The level of risk assessment should be proportionate to the significance of the decision in question, the consequences o fit, the number of people invovled and the volume of data affected. Novelty and complexity of the algorithm also play a role. Apart from foresight, a hindsight analysis should also be guaranteed. Auditting is to require companies to comply with the privacy program, including self audit or third party audit. This leaves the responsibility for everyone with the question will there be a group of people worse off due to unintended consequences?
Although the US lacks a federal privacy law, California’s Consumer Privacy Act (CCPA) addresses data protection and user privacy. The CCPA now requires companies to disclose their use of AI for decision-making purposes when collecting and processing data. Additionally, companies must ensure transparency for complex AI usage and provide clear information to users. The upcoming California Privacy Rights Act (CPRA), effective in 2023, specifically addresses AI and machine learning by giving consumers the right to opt-out of automated decision-making technologies.
So far, information privacy law is generally based on the 1980 OECD Guidelines on the Protection of Privacy and Transborder Flows of Personal Data. It recognizes the complicated and nuanced nature of privacy and allows flexibility. Despite effort, given the challenges of anticipating machine learning outcomes and deciphering algorithmic decisions, it is unlikely that a single approach can entirely prevent negative consequences. Therefore, it is advisable to use a combination of measures when algorithmic decisions have significant impacts. By using preventive measures like transparency and risk assessment in conjunction with retrospective measures such as audits and human review, it may be possible to detect and rectify unjust outcomes. These measures can complement one another and yield better results collectively. Moreover, the integration of risk assessments, transparency, explainability, and audits could bolster current remedies for actionable discrimination by providing legal documentation. Nonetheless, not all algorithmic decision-making has significant effects, so the requirements should correspond to the level of risk.
Employment
AI has the potential to automate many jobs, which can lead to job displacement and economic inequality. It is important to ensure that the benefits of AI are distributed fairly and that workers are provided with the necessary training and support to adapt to the changing job market.
As previously discussed, many people argue that concerns over the impact of artificial intelligence and automation are unfounded because the past shows that technological disruption in one industry did not necessarily lead to the disruption of others. However, this argument is flawed because the current situation is different in several ways that make it likely that the future will be impacted differently. Unlike in the past, AI can be applied to almost any industry. This means that the development of AI that can understand language, recognize patterns, and problem-solve has the potential to disrupt many different sectors. For example, AI is already being used to diagnose diseases, handle medications, address lawsuits, and even write articles like this one. Another significant difference is the speed of technological progress. Unlike the linear progress of the past, technological advancement now occurs exponentially. For instance, Moore’s Law predicts that the number of transistors on an integrated circuit doubles every two years. And humans tend to underestimate the impact of exponential growth.
Goldman Sachs economists have estimated in a report that the latest wave of artificial intelligence, which has given rise to platforms like ChatGPT, could lead to the automation of up to 300 million full-time jobs worldwide. The economists predicted that about 18% of work globally could be computerized, with advanced economies more affected than emerging markets. This is due in part to the higher risk to white-collar workers compared to manual laborers, with administrative workers and lawyers expected to be the most affected. Physically demanding or outdoor occupations, such as construction and repair work, are expected to see little effect. The report also estimated that in the US and Europe, up to two-thirds of current jobs are exposed to some degree of AI automation, and up to a quarter of all work could be fully automated by AI. However, historically, technological innovation that displaces workers has also created employment growth in the long term. Despite the significant impact on the labor market, most jobs and industries are only partially exposed to automation and are more likely to be complemented than substituted by AI. The adoption of AI is expected to increase labor productivity and boost global GDP by 7% annually over a 10-year period. Workers in partially exposed occupations may apply their freed-up capacity toward productive activities that increase output. Labor cost savings, new job creation, and a productivity boost for non-displaced workers could lead to a labor productivity boom similar to those seen after the emergence of earlier general-purpose technologies such as the electric motor and personal computer.
In conclusion, AI has the potential to bring significant benefits to society, but it also poses a number of ethical challenges. It is important to ensure that AI systems are designed to be fair, unbiased, secure, accountable, and beneficial to society as a whole.