The Future of Artificial Intelligence AND the Bias and Fairness

If ignored the future of AI will ruin the world. Understanding the ignored issues with AI.

By Ethan Yim

two hands reaching for a flying object in the sky

What is Artificial Intelligence?

Artificial intelligence (AI) refers to computer systems that can perform tasks that typically require human intelligence, such as visual perception, speech recognition, and decision-making. AI is accomplished by using algorithms and machine learning techniques that allow software systems to improve through experience and new data. Key categories of AI include computer vision, natural language processing, robotics, and expert systems for complex problem-solving and planning.

How does Artificial Intelligence Work?

Artificial intelligence systems use machine learning algorithms to analyze data, identify patterns and make predictions or decisions without being explicitly programmed to do so. The algorithms "learn" by being trained on large datasets, adapting their neural network models and parameters through experience to become better at the desired tasks. This ability to learn from data and improve through practice allows AI systems to exhibit intelligent behavior and perform human-like cognitive functions such as computer vision, speech recognition, and language translation.

To understand the true power that AI has watch the video above.

The Future of AI can be used in Future Industries

AI is a growth industry

Artificial intelligence is advancing at an exponential pace that is outpacing human comprehension. The speed and scale at which AI systems can ingest data and learn complex patterns through techniques like deep learning is unprecedented.

Within short periods, AI models can analyze millions of examples and teach themselves to make inferences and predictions that no human could match. The inner workings of these powerful systems are often black boxes, with logic hidden within multi-layer neural networks.

Some examples of how AI can be used:

Note* These images were created by Adobe Firefly (AI generating texts to images)

Manual labor jobs

Jobs involving repetitive physical tasks are the most susceptible to automation by AI robots and machines. Manufacturing, transportation, construction, and agriculture workers will need to adapt.

Customer service and sales

AI chatbots and virtual assistants are automating more routine customer interactions. But human skills in building relationships and creativity remain vital.

Healthcare

AI is automating clinical documentation and administrative tasks, allowing doctors and nurses to spend more time with patients. AI also aids diagnosis and treatment planning.

Artificial Intelligence and Employment

Artificial intelligence technologies are rapidly being adopted in diverse sectors such as healthcare, finance, transportation, manufacturing, and agriculture, around 25% of jobs will be lost in 5 years. Many jobs will be lost as AI will take over but this leads to an increase of bias and fairness in many industries.

Bias in Artificial Intelligence

Bias is leaning in favor or against something, typically in a way that is closed-minded, prejudicial, or unfair. Bias results in inequitable views and treatment of people or things, where one is given advantage over others.

Also, AI bias refers to discrimination, prejudice, or unfairness that arises in AI systems' outputs and predictions due to ingrained human biases in the data or assumptions used to train machine learning models. When flawed human decisions, historical inequities, stereotypes, or imbalanced variables around factors like race or gender exist in training data or algorithms, they imprint those same biases into the models.

World

Bias perpetuates inequality and injustice in society. Implicit biases are automatic associations that influence our split-second decisions. These unconscious stereotypes privilege some groups over others, affecting real world outcomes. Overt bias also endures through discriminatory policies, rhetoric, and hate crimes that marginalize people based on race, gender, age, and other characteristics. While some progress has been made, addressing both implicit and explicit bias remains critical for promoting compassion and equality.

Data

For AI to learn they need data to be fed. These datasets are both structured and unstructured datasets. However the main cause of bias is from the datasets that they are getting fed. Training data can bake in human biases, while collection and sampling choices may fail to represent certain groups. User-generated data can also create feedback loops that amplify bias. Algorithms identify statistical correlations in data that can be unacceptable or illegal to use, like race or gender.

AI/ML

Even if we have perfect data, our modeling methods can introduce bias.

Aggregation bias: Inappropriately combining distinct groups into one model. Fails to account for differences between groups. Example - Using one diabetes prediction model across ethnicity when HbA1c levels differ.

An example of this is: A machine learning algorithm may also pick up on statistical correlations that are socially unacceptable or illegal. (mortgage lending model)

Human Review

Even if your model is making correct predictions, a human reviewer can introduce their own biases when they decide whether to accept or disregard a model’s prediction. For example, a human reviewer might override a correct model prediction based on their own systemic bias, saying something to the effect of, “I know that demographic, and they never perform well.”

However, AI can reduce humans’ subjective interpretation of data, because machine learning algorithms learn to consider only the variables that improve their predictive accuracy, based on the training data used. In addition, some evidence shows that algorithms can improve decision making, causing it to become fairer in the process.

Actions

Although bias can "never" be solved due to unsolvable capitalism, social, political, and cultural viewpoints, there are various strategies to lower the amount of bias in artificial intelligence.

One example that can contribute is that the internet is mostly created by the majority rather than the minority. Our history books were created by the majority "altering" the history. The majority will never want anyone to see themselves as the problem. This is continued in the production of AI since the majority are leading this new field.

Real World Examples

A well-known example of machine learning bias, publicized by Joy Boulamwini in 2017 (Buolamwini, 2017), was the performance of facial detection algorithms when applied to people of different skin colors. It was shown that facial detection models created by IBM and Microsoft at the time performed surprisingly poorly (accuracy <40%) when applied to dark-skinned women. Since race and facial features are protected attributes, this problem raised the issue of fairness; however the main cause was the image features employed in the face detection algorithm.

Historical bias is illustrated by the 2016 paper “Man is to Computer Programmer as Woman is to Homemaker,” whose authors showed that word embeddings trained on Google News articles exhibit and in fact perpetuate gender-based stereotypes in society.

Representation bias is a bit different — this happens from the way we define and sample a population to create a dataset. For example, the data used to train Amazon’s facial recognition was mostly based on white faces, leading to issues detecting darker-skinned faces.

Measurement bias occurs when choosing or collecting features or labels to use in predictive models. In a 2016 report, ProPublica investigated predictive policing and found that the use of proxy measurements in predicting recidivism (the likelihood that someone will commit another crime) can lead to black defendants getting harsher sentences than white defendants for the same crime.

Possible Solutions

Be aware of the contexts in which AI can help correct for bias as well as where there is a high risk that AI could exacerbate bias.

When deploying AI, anticipate domains prone to unfair bias, like those with previous biased systems or skewed data. Stay up to date on where AI can improve fairness and where it has struggled.

Establish processes and practices to test for and mitigate bias in AI systems.

Tackling bias requires technical tools to highlight sources of bias and influential traits, operational strategies like improved data sampling, auditing by red teams, and transparency about processes and trade-offs made. Finally, transparency about processes and metrics can help observers understand the steps taken to promote fairness and any associated trade-offs.

Engage in fact-based conversations about potential biases in human decisions.

As AI reveals more about human decision making, leaders can consider whether the proxies used in the past are adequate and how AI can help by surfacing long-standing biases that may have gone unnoticed. When models trained on recent human decisions or behavior show bias, organizations should consider how human-driven processes might be improved in the future.

Invest more in bias research, make more data available for research (while respecting privacy), and adopt a multidisciplinary approach.

More progress on tackling bias will require investment in technical and multidisciplinary research, making more data available to researchers, and interdisciplinary engagement with experts across fields. Continually evaluating the role of AI decision-making as experience grows will be key.

Invest more in diversifying the AI field itself.

Many have pointed to the fact that the AI field itself does not encompass society’s diversity, including on gender, race, geography, class, and physical disabilities. A more diverse AI community will be better equipped to anticipate, spot, and review issues of unfair bias and better able to engage communities likely affected by bias. This will require investments on multiple fronts, but especially in AI education and access to tools and opportunities.

Conclusion

Large datasets can be very hard to clean to create no bias in datasets. Also getting data that is considered unbiased is also very hard. While these solutions can help lower bias, remember to never fully trust what you are told from artificial intelligence.

Fairness

Fairness generally means the absence of prejudice or favoritism towards individuals or groups based on their characteristics.

However, There are many potential definitions of fairness (at least 21 identified so far).

Three mathematical versions are:

Equal Outcomes (demographic parity): Requires equal positive classification rates between groups. Focuses only on outcomes.
Equality Of Opportunity: Requires equal true positive rates between groups. Focuses on positive errors.
Equality Of Odds: Requires equal true positive and true negative rates. Balances both error types.

Overall, fairness in AI is not definitively solvable but should be an ongoing process of understanding biases, evaluating trade offs, and centering impacted communities.

Let's say this graph with circles and squares has a blue line that is considered the most fair for both sides.

Now this graph has green for people that have diabetes, and blue for people that do not. As you can see only a few people overlap for both probability distributions. Using a single threshold for these two groups would lead to poor health outcomes.

A common thought of what to do is to use calibrated predictions for each group. Yet in this graph you can see that if your model’s scores aren’t calibrated for each group, it’s likely that you’re systemically overestimating or underestimating the probability of the outcome for one of your groups. With various groups doing this is impossible.

This problem leads to other issues called individual fairness. the black arrows point to blue and green individuals that share very similar characteristics, but are being treated completely differently by the AI system. As you can see, there is often a tension between group and individual fairness.

The Butterfly Effect

The Butterfly Effect means small initial changes can cascade unpredictably, causing dramatically different outcomes in complex nonlinear systems.

Coined in the 1960s, the Butterfly Effect highlights that tiny perturbations can trigger far-reaching consequences through interconnected components in intricate systems like weather.

The Butterfly Effect manifests in AI systems through small data tweaks, inherent algorithmic biases, feedback loops that compound unfairness, and adversarial attacks exploiting vulnerabilities. Even minor issues like limited training data, flawed assumptions, and subtly manipulated inputs can cascade through interconnected nonlinear components, resulting in substantially biased outcomes. Identifying how seemingly insignificant factors can snowball into ethical crises enables strategies to enhance AI fairness, like diversifying data and evaluating systems' societal impacts. The Butterfly Effect highlights the need for holistic solutions to prevent incremental harms from accumulating into full-blown discrimination.

Relevance of the Butterfly Effect to AI fairness and bias

The Butterfly Effect is highly relevant to AI fairness and bias. Even small biases in data or algorithms can compound and lead to substantial unfair outcomes in AI systems.

The Butterfly Effect emerges in AI systems due to their complex, interconnected components like data, algorithms, and user interactions. Even small biases in these elements can compound in unpredictable ways, leading to substantially unfair outcomes.

For example, minor data tweaks, underlying algorithmic biases, changing distributions, adversarial attacks, or bias-reinforcing feedback loops could trigger a cascading effect. Given AI's potential societal impacts, understanding how minor issues can snowball unpredictably through complex, interconnected components is critical for grappling with ethical AI.

Causes

Small adjustments in input data can significantly impact AI fairness and bias. This is known as the Butterfly Effect.

Data sampling methods can introduce bias if certain groups are under or over represented. Representative sampling is important.

Imbalanced demographic makeup in training data can lead to poor performance for underrepresented groups. Diverse and representative training data is important.

Feature selection and engineering impact model behavior. Using proxies for protected attributes can introduce bias. Omitting relevant features can result in biased models.

Real World Examples

Facial recognition technology- Facial recognition algorithms can exhibit racial bias due to imbalanced training data. Even small demographic skews cascade through nonlinear systems, disadvantaging minority groups.

Healthcare algorithms- Biased assumptions in healthcare algorithms can lead to unequal access for minorities. Minor variations compound unfairly through opaque systems.

Hiring algorithms- Recruiting algorithms can amplify gender biases in hiring. Seemingly negligible data issues propagate, exacerbating inequality.

Large Language Models- Minor data or algorithm tweaks in large language models could trigger unintended biases. The models' complexity allows small problems to cascade unpredictably.

Feedback Loops

Feedback loops can amplify biases in AI systems, causing the Butterfly Effect on fairness.

Reinforcing feedback loops occur when biased predictions lead to actions that further perpetuate the initial biases.

In recommendation systems, filter bubbles reinforce users' biases through personalized content.

Algorithmic confounding is when predictions influence the data used to evaluate performance.

Adversarial attacks can exploit vulnerabilities in AI systems, causing the Butterfly Effect on fairness and bias.

Adversarial examples are crafted inputs designed to cause incorrect or biased predictions. They can target groups or individuals leading to discrimination.

Model inversion and membership inference attacks reveal information about training data or individuals. This exposes disparities or biases that can be further exploited.

Poisoning attacks inject crafted data or modify parameters to introduce biases. Perturbations are subtle and hard to detect but propagate via the Butterfly Effect.

There are various ways that can decrease "chaos theory known as the butterfly effect."

Robustness bias - When models are vulnerable to attacks, subgroups defined by sensitive attributes like race and gender are often less robust, leading to unfair outcomes. Evaluating robustness bias is critical for real-world AI systems relying on deep neural networks.

Adversarial training - This technique minimizes worst-case loss over allowed input perturbations, improving model robustness against adversarial examples and enhancing stability. Adversarial training reduces the risk of unintended consequences from small input changes.

Ensemble adversarial training - By augmenting training data with a diverse ensemble of adversarial examples, this approach enhances robustness against a broader range of attacks. Ensemble training increases model resilience to input perturbations that could cause unintended bias.

Certified robustness - Providing formal guarantees on model behavior under adversarial perturbations ensures stability and prevents unintended outputs. Certified robustness verifies that small input changes will not significantly alter model outputs.

Randomized smoothing - This technique enables provable robustness against bounded adversarial attacks using randomized noise injection. Randomized smoothing formally certifies model fairness is maintained under minor perturbations.

Adversarial detection - Using a separate model to detect adversarial inputs prevents them from triggering unintended model biases. Adversarial detection identifies malicious perturbations and protects model behavior.

Oversampling minority classes - Creating copies of underrepresented examples balances class distribution and reduces overfitting, improving model fairness and stability. SMOTE is an oversampling technique that generates synthetic minority class instances by interpolation.

Undersampling majority classes - Removing majority class examples balances the distribution and makes decision boundaries less sensitive to small data changes. Tomek Links is an undersampling method that removes samples near the decision boundary.

Synthetic data generation - VAEs, GANs and other generative models create realistic synthetic data for underrepresented groups, ensuring representative datasets. High-quality synthetic data reduces the impact of small data shifts on model fairness.

Stratified sampling - Dividing the population into homogeneous strata and sampling proportionally from each ensures dataset representativeness. Stratified sampling reduces model sensitivity to minor data variations.

Fairness-aware machine learning - Incorporating fairness constraints during training minimizes discrimination by reducing model sensitivity to minor input changes. Optimization methods balance accuracy and fairness across groups.

Post-processing for fairness - Adjusting pretrained model outputs ensures fairness by transforming predictions to equalize metrics between groups. Post-processing compensates for training biases without model retraining.

Fairness through awareness - Requiring consistent treatment of similar individuals makes models less sensitive to small data shifts. Lipschitz constraints on classifier behavior improve stability despite perturbations.

Regularization for fairness - Adding fairness penalty terms to the training objective discourages outcome differences between groups. Regularization induces more equitable data representations to mitigate small data effects.

Fairness-aware metrics - Fairness metrics make it easier to notice unfair biases against certain groups that may be introduced as data shifts. Tracking metrics exposes discrimination that can emerge as a result of the butterfly effect in AI systems.

Auditing tools - Perturbing inputs to analyze model sensitivity to protected attributes systematically uncovers biases and fairness violations. Audits pinpoint areas susceptible to butterfly effects.

Model interpretability - Explaining model decisions locally reveals the factors influencing outcomes. Interpretability enables scrutinizing potential biases and unintended consequences.

Continual monitoring - Tracking performance and user feedback over time uncovers emerging biases and fairness issues. Active monitoring is key to detecting butterfly effects.

Bias in Human Review

Human review bias occurs when a human decision-maker overrides an unbiased model's predictions based on their own biases.

Human judgment is essential to decide when an AI system has minimized bias enough to be released.

Humans must determine when automated decision-making should be allowed at all. Optimization algorithms cannot resolve these questions.

Human decisions are also difficult to probe or review: people may lie about the factors they considered, or may not understand the factors that influenced their thinking, leaving room for unconscious bias

Humans are prone to bias in decision-making, influenced by personal characteristics and unconscious factors. This can lead to unfair outcomes.

AI and machine learning models can help reduce subjective human bias, as they learn from data rather than innate biases.

Evidence shows algorithms can make decisions more fair in areas like criminal justice and financial lending.

However, examples of biased AI:COMPAS algorithm had higher false positive rates for black defendants in predicting recidivism. Hiring algorithm penalized applicants from women's colleges, exhibiting gender bias. Facial analysis tools had higher error rates for women and minorities. Image search results for "CEO" were predominantly men.

AI bias can emerge from historical biases in data, uneven data representation, use of proxies, and other modeling choices.

Balancing human and AI decision-making is important to reduce bias. AI can counter explicit bias but models must be closely monitored for fairness.

Early technical progress is underway, but much more is needed!

Below is an example of someone doing an experiment on the new gpt4 image reader and it chooses to side with the image over the user.

Impacts in the Real World

Google apologises for Photos app's racist blunder

Google says it is "appalled" that its new Photos app mistakenly labelled a black couple as being "gorillas". Its product automatically tags uploaded pictures using its own artificial intelligence software. The error was brought to its attention by a New York-based software developer who was one of the people pictured in the photos involved. Google was later criticized on social media because of the label's racist connotations.

Youtube Rabbit Hole

Around the time of the 2016 election, YouTube became known as a home to the rising alt-right and to massively popular conspiracy theorists. Yet, he process of “falling down the rabbit hole” was memorably illustrated by personal accounts of people who had ended up on strange paths into the dark heart of the platform, where they were intrigued and then convinced by extremist rhetoric—an interest in critiques of feminism could lead to men’s rights and then white supremacy and then calls for violence. YouTube announced tweaks to its recommendation system with the goal of dramatically reducing the promotion of “harmful misinformation” and “borderline content” (the kinds of videos that were almost extreme enough to remove, but not quite).

Microsoft's Tay Chatbot

In 2016, Microsoft launched a chatbot named "Tay" on Twitter, designed to engage with users in real-time conversations. However, within hours of its release, Tay began posting offensive and inflammatory tweets. The chatbot had been manipulated by users to spew hate speech and racist comments, highlighting the potential for AI systems to learn and propagate harmful biases from user interactions.

The Process

This is were this project all began, I learned various new things over time. I learned what the butterfly effect is and how it is present in AI. I also learned how AI works and the meaning of bias and fairness.