Approaches to Reducing AI Bias Continue to Evolve

Earlier this year, the Stanford Institute for Human-Centered Artificial Intelligence (HAI) released the AI Index 2022 Annual Report. The report is a comprehensive study of the current state of AI, with a focus on progress made in the past year.

Here are some key takeaways:

"Bigger and More Capable" AI Systems are More Likely to be Out of Line with Human Values

The report found that "as AI systems become more capable, there is a risk that they could become less predictable and more opaque in their decision-making process." In other words, as AI systems get better at making decisions or predictions, they may also become less transparent in their decision-making process, which could lead to them making decisions that have disastrous consequences.

This is because AI systems often rely on large data sets to learn and operate. These data sets can be biased, either intentionally or unintentionally. As a result, AI systems can inherit these biases and produce results that are unethical or harmful—at a massive scale. This can be prevented by being aware of the potential biases in AI systems and designing them to reduce these risks.

The report recommends that organizations pay close attention to this trend and take steps to ensure that their AI systems are designed in a way that will prevent them from producing outputs contrary to desired (and, in some cases, legally required) values. One way to do this is to design AI systems that are "transparent by default." That is, design AI systems in a way that makes it easy for humans to understand how they work and why they make the decisions they do. This will help ensure that AI systems remain accountable to humans and do not stray from their nobler purposes.

Reduced Costs Driving Rapid Adoption

In recent years, there has been a boom in the use of artificial intelligence (AI) technologies. This is largely due to the decrease in cost and training time for things like image classification models and artificial neural networks (ANNs).

For example, in 2012, it cost $1 million to train an ANN on ImageNet—a large dataset of images used for training computer vision models—using GPUs. By 2018, that cost had decreased to $0.2 million. And by 2020, it had decreased again to $0.06 million—a 95% decrease from 2012 levels. This decrease is due largely to advances in hardware and software technologies. As these technologies keep getting better, the cost of training ANNs is likely to keep going down as well. Meanwhile, the report states that the fastest possible training times for mature AI models in general have dropped by a factor of 27.

While this has led to the rapid adoption of AI in multiple industries, it has also exacerbated the potential risks.

Poorly trained or biased models can have negative consequences for the people and businesses that rely on them. This could lead to the misclassification of items, which could have cost and efficiency implications for manufacturing, supply chain, retail, and e-commerce, and more serious implications in fields such as medical diagnosis or law enforcement. A biased model may also generate inaccurate results that favor one group over another. This could create unfairness in areas such as credit scoring or job recommendations.

While the benefits of AI are clear, it is important to be aware of the risks associated with its use. Proper training and testing of models is essential to ensuring that they perform as intended and do not cause unintended harm.

Interest in AI Ethics Grows Dramatically Among Industry Players

Researchers with industry ties contributed 71% more publications on fairness-focused AI ethics at industry conferences in 2021 compared to 2020, according to a FastCompany article. This may be because people are paying more attention to AI algorithms and the effects that biased decision-making could have on society.

AI ethics is a hot topic among tech companies, with many large companies establishing ethics boards or publishing internal ethical guidelines. However, there is still a long way to go in terms of ensuring that AI systems are ethically sound. As AI becomes more common, it is important that people in the industry continue to think about the ethical implications of their work.

Another key finding from the report is that there has been a dramatic increase in interest in AI ethics among industry players. In 2019, only 33% of respondents to a survey said that their organization had an ethical framework for AI development. In 2020, that number jumped to 63%. This increase represents a significant shift in the way industry players view AI ethics. It shows that more and more organizations are recognizing the importance of ensuring that their AI systems are ethically sound.

This change is probably due to a mix of things, such as high-profile scandals involving unethical use of AI (like facial recognition technology), more people being aware of the risks of unethical use of AI, and more rules about how AI is developed and used. Whatever the reasons, it's clear that organizations are taking AI ethics seriously and are working to ensure that their own practices align with ethical standards.

Methods for Reducing Bias Still Evolving

At the moment, there is a debate in the industry about the best way to get rid of toxicity and bias from AI training data. The debate centers on whether it is better to carefully curate training data or to increase dataset size so that “good” data pushes bad content to the margins.

Synthetic data offers a promising solution here.

Synthetic data is generated by algorithms and can be used to train AI models. Synthetics have several advantages over real-world data, including being more diverse, representative, and controllable.

How is synthetic data more controllable than real-world data? With synthetic data, you can carefully control the characteristics of the data set, which is not always possible with real-world data. This makes synthetic data a valuable tool for debugging and validating AI models.

Synthetic data can also be generated in large quantities, which is not always possible with real-world data. As a result, synthetic data offers a compelling alternative to real-world data for training AI models.

Invite More People Into the Conversation

I've talked about this before: When everyone in the room looks or sounds like you, it's time to get a bigger room and invite more people in. Diversity among those programming AI technology itself is a critically important issue, not just for ethical reasons but also for economic ones.

When it comes to explainability, various points of view improve the development of artificial intelligence, and inclusion has been shown to boost profitability generally. Companies in the top quartile for gender diversity are 21% more likely to be profitable than the average, while ethnic and cultural diversity correlates with a 33% increase in performance. In other words, diversity is good for business.

Conclusion

Tech companies still talk a lot about the ethical implications of AI, and more and more organizations are realizing how important it is to make sure that their AI systems are ethically sound. There is still a long way to go in terms of ensuring that AI systems are ethically sound, but industry players are starting to pay greater attention to the issue.