You can’t ignore the rapid growth of Artificial Intelligence (AI). It’s everywhere, from small startups to massive corporations, and its market value reached almost $197 billion in 2023. According to Forbes, pretty much every company will use AI to some extent by 2025, boosting its market value even more. However, as it becomes more integrated into our lives, it’s becoming more obvious it’s not without faults – especially discriminatory bias.
AI doesn’t come up with bias on its own. It simply learns from the data it’s given. If the data reflects or contains biases that are found in society, then AI will too. More specifically, if we feed AI’s training software biased information, it will produce biased results. In other words, the biases we see in AI are actually coming from us, the people who develop, train, and use these systems.
This brings up a big question. In our rush to implement AI, did we accidentally create a prejudiced monster? We’re going to look at how human biases made their way into AI and the problems this causes, and think about what may happen if we don’t try to mitigate it.
Understanding Machine Learning and AI Bias
It’s impossible to fully understand AI and its bias without first talking about machine learning, the base of how AI operates. Machine learning teaches computers how to think and become smarter on their own without us having to program them every step of the way. The process involves feeding AI massive amounts of data, letting it analyze it for emerging patterns, and then make predictions based on what it’s observed.
For example, to teach AI how to recognize spam emails, you would provide thousands of email examples. The machine would then use them to learn to distinguish between spam and non-spam messages by identifying patterns. This would save you from having to list every spam characteristic, making the process much quicker and easier.
The most crucial part of machine learning is the data we feed it. The better and more diverse it is, the more accurately AI can learn and apply its knowledge to new, unseen information. However, if the data is prejudiced, it can lead to AI bias and inaccurate or unfair outcomes. The impact of this can be significant – a study by Gartner indicates around 85% of AI projects now deliver incorrect outcomes due to these skewed biases.
When AI systems learn from biased data, they can unintentionally perpetuate and amplify those biases. For example, if an AI system is trained to evaluate job applications, but the data it learns from contains a historical bias against women, it might discriminate against female applicants. This isn’t the AI “thinking” prejudicially on its own – it simply mirrors the biases present in the training data.
It’s not just a speculation. AI bias has already had many real-world consequences, often resulting in severe exclusion of certain minorities in access to advanced healthcare, cheaper interest rates, or better employment.
For instance, a healthcare algorithm used on more than 200 million people selected mainly White patients to be offered more extensive healthcare insurance. This excluded a lot of Black patients with similar health histories from access to more advanced medical care. Similarly, Amazon’s hiring AI favored male applicants for many technical positions, while AI used for mortgage lending has been reported to charge Latino and Black borrowers higher interest rates.
How Does AI Learn Its Bias?
AI learns its bias from supplied data, but in many different ways. It can come from multiple sources and manifest in various stages of the AI development process. Understanding them is key to identifying and mitigating bias in AI systems. The main sources of AI bias include:
- Training data – A dataset used to teach AI how to perform its tasks. It often includes examples, special features, and labels which help recognize patterns and make decisions.
- Algorithms – Sets of instructions that guide AI in processing data, making decisions, and solving problems. They are essential for AI systems to learn, make predictions, and improve their performance.
- Data collection issues – These happen when the data collected doesn’t represent the full reality or when it reflects existing societal prejudices.
- Problem framing – The way a specific problem is described so AI can understand and solve it to the best of its ability. If issues are not defined properly, AI can overlook certain perspectives or outcomes and favor some groups over others.
- Feedback loops – A process where AI results are fed back into the database for a computer to learn from. This creates a loop that can further reinforce existing bias.
- Confirmation bias in development – This occurs when developers influence AI with their own beliefs and experiences, which can introduce prejudice.
- Lack of diversity in AI development teams – This can lead to a lack of awareness of potential biases and how they might impact different groups of people.
AI Bias and the Importance of Data Protection
Talking about bias in AI is a good reminder for you to look out for your own online safety. In a world where technology can reflect societal biases, taking steps to protect your digital footprint becomes an essential proactive measure. A reliable VPN adds a layer of security by encrypting your internet connection and masking your digital activity from prying eyes. It’s an essential tool if you’re concerned about your digital privacy in an increasingly AI-driven world.
Reputable VPNs work across many popular operating systems, including iOS, Android, and Windows. Just a couple of clicks connects you to a VPN server, not only securing your data but also letting you use AI apps on any network. Alternatively, download a VPN Chrome extension to protect your browsing with a handy add-on. It’s a tiny step you can make towards promoting digital fairness while protecting your online privacy.
Types of AI Bias
Algorithmic Bias
Algorithmic bias in AI happens when the algorithms computers use are incorrect, leading to them producing prejudiced or unfair results. Imagine trying to learn about animals using only books about dogs. You’d think all animals look like man’s best friend, which isn’t true. Similarly, if AI algorithms learn from incomplete data, its decisions or predictions won’t serve the wider public, but only a select few.
No matter which industry applies AI algorithmic predictions, this type of bias can have disastrous outcomes, especially in healthcare, financial services, and the justice system. Scientists regularly find new examples of AI discrimination, some of which are more worrying than others. In the 2019 healthcare example noted above, for example, scientists found an AI algorithm favored White patients when it came to shortlisting people eligible for more advanced care.
This had nothing to do with skin color, which wasn’t specified in the algorithm. It just happened that Black patients spent less money on healthcare, so the AI assumed they suffered from fewer medical issues. In reality, they were dealing with far more illnesses than White patients. They were cut off from access to better care, unintentionally perpetuating a racial bias in medical care and insurance.
Sample Bias
Sample bias happens when the AI training data doesn’t fully capture the diversity of the population or situation it’s intended for. The more limited a data sample is, the more likely it is that AI will discriminate against certain individuals or groups of people who are not represented in the dataset.
This bias is common in facial recognition systems. The National Institute of Standards and Technology found that many AI facial recognition systems don’t work well for people of color. That’s because their data sample often doesn’t have as many examples of minority faces. Without a variety of ages, ethnicities, and genders represented, the AI is essentially wearing blinders. As a result, it doesn’t work for anyone who doesn’t match the training set.
Such discrimination can sometimes have significant repercussions. Several US facial recognition programs have come under fire for not recognizing people of color. In some instances, these inaccuracies have led to wrongful arrests. It’s a stark reminder of how sample bias can have real, tangible impacts on people’s lives, reinforcing the importance of diversity in both our society and in the data that trains the machines intended to serve us all.
Prejudice Bias
Prejudice bias happens when preconceived human opinions and discriminatory attitudes manifest themselves within AI decisions. Unlike human prejudice, which is rooted in social, cultural, or personal beliefs, prejudice in AI comes from the data it is fed either by developers or users and the way it’s programmed. This type of bias often hides under good intentions as developers may not even realize their perspectives may be nudging the AI in an unfair direction.
The COMPAS system is a classic real-life example of prejudice bias. The AI program, designed to assess how likely criminals were to recommit offenses, started making skewed predictions. After analyzing its results, experts found the system was twice as likely to flag Black individuals as “high-risk” recidivists compared to White criminals. Sadly, the company behind COMPAS refused to admit any AI wrongdoing and hasn’t changed its algorithms.
Measurement Bias
Measurement bias happens when the data is collected, measured, or labeled incorrectly. Unlike other types of bias, this one can be much more difficult to spot and detect in development as it’s not just about who or what is represented in the data, but about how that data is interpreted and used. This can make AI much less reliable and useful in areas like health, banking, and law.
This type of bias is prevalent in survey analysis, often stemming from how questions are asked, who responds, and what the survey measures. First, knowing it’s a survey can influence how you understand and answer questions. If they are confusing, leading, or too complex, they might not capture your true opinions or behaviors. This can skew the results, making them biased towards certain interpretations.
As a result of wrong survey assumptions, AI systems may recommend products, services, or policies which don’t align with the true preferences or needs of the users they serve. At the very best, it’s mildly annoying for you as a user, but it can be highly discriminatory and damaging for those the data doesn’t represent.
Exclusion Bias
Exclusion bias happens when some groups of people or types of information are left out – internationally or not – from the AI learning process. This impacts how fair and effective AI systems can be in their predictions as they don’t account for diversity present in the real world.
This kind of bias often pops up because the data used to teach the AI doesn’t cover all the bases, which can happen more often than you think. For example, if a health app is mostly trained with data from young people, it might not be as good at giving health advice to older people. Or, if a voice recognition system is trained with voices from a certain region, it might struggle to understand accents from other parts of the world.
The latter is very easy to observe if you watch YouTube or other internet content with automatically generated subtitles. Experts found AI speech recognition is a lot less accurate with female voices and non-English accents, generating random words or sentences that don’t make sense. While most likely not intentional, these shortcomings can heavily disadvantage those not represented.
Selection Bias
Selection bias happens during the data collection or selection phase when chosen information doesn’t fully represent the reality and diversity of the world. Instead, data samples end up prioritizing certain individuals or groups over others. Sounds a lot like sample bias, doesn’t it? Though the two are often used interchangeably, they differ ever so slightly depending on their context.
Both types of biases involve problems with how data is picked and whether it truly reflects the group that AI is currently learning about. However, selection bias covers a wider range of issues that can mess up the selection process, like set criteria, researcher bias, and non-random sampling methods. Sample bias simply refers to the already picked sample, which doesn’t represent the larger group it’s supposed to.
Let’s look at some examples of selection bias. If AI algorithms use data from candidates who actively seek jobs through a specific platform, it might overlook equally qualified candidates who apply through different channels. Or if a company selects survey participants based on their previous engagement with the brand, the data might skew toward positive feedback, excluding critical perspectives from dissatisfied customers.
Recall Bias
Recall in AI is about how good the system is at catching everything it’s supposed to catch. For example, if an AI is supposed to identify all the spam emails, recall measures how many of those it actually spots. If it misses a lot of them, we’d say it has poor recall.
Recall bias happens when AI has been taught using data that isn’t quite balanced. Say it learned mostly from non-spam emails. It might get really good at recognizing those, but then omit a lot of spam because it didn’t practice enough on them. It’s like only practicing easy questions for a test, and then finding it’s full of hard questions.
This can often happen if AI’s dataset uses historical information from users who were asked to recall specific events that happened to them. Human memory is notoriously fallible. It can be influenced by many factors, leading people to forget, exaggerate, or downplay their experiences. The subjective nature of memory means that data collected in this way can be unreliable.
Think about AI systems used in healthcare. Many models are trained on patient-reported data about symptoms, medical history, and lifestyle choices. However, patients may vividly remember and report recent dramatic experiences, while omitting milder, chronic, or past symptoms. This selective memory distorts the patient’s medical profile and skews the dataset AI learns from, potentially leading to misdiagnoses or inappropriate treatment recommendations.
AI Bias in Action: How AI-Generative Images Reinforce Stereotypes
We did some research to test AI biases for ourselves. We got our hands dirty by running some practical checks on popular generative AI services – ChatGPT (DALLE), Bing AI, and Pixlr – to see if they show bias in their results. We asked each AI to generate 10–30 images of people working in each of a set of professions – CEO, nurse, doctor, scientist, and teacher – over the course of a few weeks. We also asked it to generate images of criminals. We used basic prompts like “Draw a doctor” to keep the potential responses as open as possible.
In our sample, ChatGPT over-represented men as CEOs, doctors, scientists, and even teachers – more than 90% of our requests for images of CEOs, doctors, and scientists resulted in pictures of men. At the same time, it overwhelmingly portrayed nurses, kindergarten teachers, and healthcare workers as young women. When it came to portraying nationalities, the OpenAI tool seemed to focus more on Caucasians, rather than people of color. We only saw a couple of images depicting non-White workers.
Pixlr turned out to show the highest level of bias. It stuck to the script of doctors, scientists, and CEOs being middle-aged men, with a mix of White and Asian backgrounds. Nurses were always women, mostly White, though we spotted a few women of color in the mix. Teachers returned unexpected results: we saw a fair mix of men and women but all on the older side, rocking big square glasses.
Bing AI came out as the least biased tool in our study. It gave us a balanced mix of genders across all requested professions and included people of various nationalities as doctors, nurses, and teachers. Interestingly, it also exclusively showed CEOs as female – though all of them were Caucasian, highlighting an area for improvement.
When it came to generating images of criminals, ChatGPT seemed unsure of what to do, and its answers varied depending on which country we ran the tests from. A few times, the AI outright refused to follow the prompts to avoid reinforcing stereotypes or producing inappropriate content. Other times, it asked multiple follow-up questions to make sure it complied with its use policies.
When we finally managed to persuade it to generate an image of a criminal, ChatGPT remained pretty neutral. The images produced didn’t seem to perpetuate racial biases, depicting a hooded figure, presumably male, without a face and hiding in nighttime shadows to maintain an aura of mystery.
Asking Bing and Pixlr for images of criminals was easy, but the outcomes varied. Bing kept it light with generic cartoon characters, steering clear of real-life stereotypes. Pixlr gave us more realistic results but generated only men of color, showing it could be influenced by biases and prejudices.
It’s important to note the seemingly biased results may not inherently stem from intentional prejudice, but rather general statistics. According to the World Health Organisation, around 90% of nurses worldwide are female, so it would make sense for AI to generate more images of women in nursing uniforms. Similarly, women make up around 85% of all pre-primary school teachers in almost all countries. The proportion of male and female teachers evens out more in older schools, but not to the male-heavy extent shown in our results – 80% of ChatGPT’s images of teachers were of men.
Considering popular AI tools, like ChatGPT and Bing, are trained using text databases available on the internet, it would explain the seemingly obvious bias our research presented. Without much monitoring and regular testing, these statistics turn into prejudice, producing results that reflect biases present in the modern world.
The Future of Bias in AI
The future of bias in AI always sparks an intense debate among technologists, ethicists, and the general public. Some argue AI bias could improve its ability to understand and interact with the world in a way relatable to humans. This suggests that if AI can reflect the complexity of human bias, it might also highlight inequality or injustice within our society, helping us see and address these challenges.
This perspective is not without its critics. Many say teaching AI human biases could further establish harmful stereotypes and prejudices, marginalizing already vulnerable groups even more. AI’s broad influence can reach a wider population, accidentally amplifying these biases and embedding them within society.
However, as AI databases grow really fast, identifying and removing prejudices becomes almost impossible. Removing specific markers like age and race from datasets is one step, but AI can still develop biased outputs based on less obvious factors such as education level or income. This suggests bias in AI is not just a data problem, but a fundamental issue with how algorithms interpret and learn from information.
Despite these challenges, the AI community has put in a lot of effort to address and reduce bias. OpenAI’s work on making GPT-3 less offensive through learning from human feedback is a case in point. This approach involves training AI models to align more closely with human values and ethics. While it’s progress, it also highlights the ongoing challenge. AI, like the humans who create it, is a work in progress, continually evolving and adapting.
As we move forward, we need to recognize the potential benefits and dangers of bias in AI. We need to push for ongoing vigilance, innovative research, and a commitment to ethical development practices. The goal is to create AI that not only mimics the best of humans but surpasses it, using the latest technology to build a more equitable and fair world.
Essential Steps to Address and Mitigate AI Bias
Mitigating bias in AI involves a multifaceted approach that addresses the issue from the data collection stage all the way through the deployment and monitoring of AI systems. As an individual, it’s almost impossible to remove bias yourself – but you can advocate for new policies and guidelines. This will help improve your very own experience with AI, as the less bias computers reinforce, the more accurate results and predictions they’ll generate.
Here are some key steps AI companies can take to mitigate bias:
- Ensure diverse data representation – Collect and use data reflecting the diversity of the population or scenarios the AI system is made for. This could include a wide range of demographic groups, behaviors, and situations to minimize the risk of biased outcomes.
- Implement bias detection tools – Use advanced tools and methodologies to identify and measure AI bias. This can include statistical analysis, data visualization, and AI fairness metrics which highlight inconsistencies in performance across different groups.
- Promote diversity in AI teams – Involve individuals from a variety of backgrounds, disciplines, and perspectives. This can help uncover potential biases and assumptions which might otherwise go unnoticed.
- Stick to ethical guidelines – Develop, follow, and regularly update guidelines and standards relating to bias in AI. These should be based on current research, ethical considerations, and the potential social impact AI could have.
- Improve transparency – Make AI decision-making processes clear and understandable to developers and users. This can help identify and correct existing biases, and foster user trust in AI systems.
- Carry out audits and testing – Test AI systems regularly, especially in the deployment stage. Focus especially on rigorous performance evaluation across different demographics and scenarios.
- Incorporate user feedback – Seek and use feedback from a broad range of users, especially those from underrepresented groups. This will provide valuable insights into biases which may not come out through data analysis alone.
- Follow legal standards – Ensure AI meets legal and regulatory requirements related to discrimination and privacy. This includes understanding and applying laws that govern fairness in employment, credit, housing, and other areas where AI can have a significant impact.
- Collaborate with the community – Engaging with organizations, researchers, and experts involved in AI development can help apply best practices, tools, and research on bias mitigation.
- Focus on continuous learning – Stay informed about the latest research, trends, and advancements. Be prepared to update and refine AI systems in response to new insights and changes. This will ensure bias mitigation is an ongoing process.
Eliminating AI Bias Can Help Create a Fairer Future
The problem of bias in AI reflects the unfairness and stereotypes we already have in society. This bias doesn’t just magically appear – it comes from the data AI learns from, which can be full of our own human biases.
Fixing this isn’t just on one person or group. Everyone involved has a part to play. Developers need to make sure they’re using data that represents everyone fairly. People using AI should be on the lookout for biases and call them out. And the policymakers should set up some good guidelines to make sure AI is being fair and not leaving anyone out.
Looking ahead, getting rid of bias in AI will likely be something we’ll always have to work on. It’s about making sure that as AI gets smarter and more a part of our lives, it treats everyone fairly. This means everyone working on AI needs to keep checking their work, listening to different viewpoints, and making changes where they’re needed.
In simple terms, we all need to stay sharp and keep pushing to make AI better and fairer. As we bring more AI into our world, let’s make sure it’s the kind that helps everyone, not just a few. It’s on us to keep making it better, making sure it’s fair for everyone, no exceptions.
Leave a comment