Building Ethical Machines: On AI Bias, Data Privacy, and Model Explainability

AI has become a ubiquitous presence in our daily life, from autonomous vehicles to virtual assistants, to every student’s favourite essay writing tool. Beyond this, the generative AI hype has catapulted the world into a genAI frenzy with investors, founders, and everyday consumers taking a closer look at how they can benefit from these technologies. Simultaneously, concerns around the ethical challenges underpinning these systems are growing. Beyond worries around AI fairness or lack thereof, companies are facing real, monetary consequences of failing to prioritize consumer privacy, addressing biased algorithms, and more factors related to ethical AI. These consequences primarily come in the form of; fines for failing to adhere to regulatory requirements such as GDPR, lost customers, and reputational damage.

• Amazon scraps secret AI recruiting tool that showed bias against women
• Details On 700 Million LinkedIn Users For Sale On Notorious Hacking Forum
• Generative AI tool Stable Diffusion amplifies race, gender stereotypes

Below I highlight trends in AI adoption, outline the ethical factors we should be thinking about as those who are impacted by AI (aka everyone), and make recommendations for how those in the tech community specifically can integrate considerations of these factors into the companies we build and fund.

Private investments into AI have grown 13x in the decade between 2013 and today. Additionally, the average number of AI capabilities that organizations use has doubled from 1.9 in 2018 to 3.8 in 2022.

Despite this, there has been no substantial increase in organizations’ reported mitigation of AI-related risks.

As AI continues to advance and permeate our lives, ethical considerations become increasingly vital. Mitigating algorithmic bias, protecting user data privacy, and enabling model explainability are critical for building trustworthy and accountable AI systems.

Algorithmic Bias: Addressing Unintended Discrimination in AI Systems

One of the most important considerations in building ethical AI systems is algorithmic bias. Bias in AI systems can result in discriminatory outcomes, perpetuate existing inequalities, and amplify social biases. For example, did you know that to increase your likelihood of landing a job in tech according to this hiring algorithm, all you have to do is be named Jared and play high school lacrosse?

As AI systems are trained on historical data, they can inherit biases present in the data, leading to biased predictions and decisions. To mitigate algorithmic bias, researchers are developing techniques such as adversarial training, re-sampling, and re-calibration. Adversarial training involves training AI models to be robust against adversarial examples, which are carefully crafted inputs designed to fool the model. Re-sampling techniques aim to re-balance the training data to ensure that all groups are fairly represented, and re-calibration methods adjust the predictions of AI models to achieve fairness. Additionally, there is a growing focus on developing diverse and inclusive datasets to ensure that AI models are trained on representative data that reflects the real-world diversity of the population. These efforts are aimed at creating more transparent and accountable AI systems that are free from bias and promote fairness and inclusivity.

Data Privacy: Protecting User Information in the Age of AI

Consumer privacy protection is of utmost importance, and companies that have failed to prioritize data privacy are feeling the heat. For instance, the widely-covered T-mobile data breach that occurred in 2022 cost the company $350 million in 2022 — and that’s just in customer payouts. Since I started writing this article, news broke that T-mobile was involved in yet another massive data breach which I anticipate will have similar consequences for their bottom line and customer loyalty.

Techniques such as federated learning and differential privacy have gained traction as methods to protect user data while still allowing for meaningful AI insights. Federated learning allows multiple organizations to collaborate and train an AI model without sharing raw data. Instead, only aggregated and anonymized data is exchanged, minimizing the risk of exposing sensitive information. Differential privacy adds noise to data to prevent the identification of individuals in the training data while still enabling accurate model predictions.

Model Explainability: Enhancing Transparency and Accountability in AI

As AI systems are increasingly deployed in high-stakes domains such as healthcare, finance, and criminal justice, it is crucial to understand how these systems arrive at their decisions. Explainable AI (XAI) techniques are designed to provide insights into the decision-making process of AI models, making them more transparent and accountable. Explainable AI is crucial for building trust in AI systems, allowing stakeholders to understand how decisions are made and ensuring that decisions are fair, transparent, and aligned with human values.

To highlight a simple consequence of what might happen if we don’t understand the systems we rely on, researchers learned that minor changes via a little bit of spray paint or some stickers on a stop sign were able to fool a deep neural network-based classifier into thinking it was looking at a speed limit sign 100 percent of the time. If we don’t understand our systems, we can’t ensure we will be able to explain and agree with their decisions. This may not be so bad in a controlled research lab analyzing neural networks for future use in autonomous vehicles, but as you can imagine, the consequences of not understanding our systems and their decision making processes when deployed in the real world can be catastrophic.

What’s Next?

I think that one of the biggest roadblocks to bringing AI ethics into our companies is the lack of regulatory standardization. With so many “gold standards” available, it is difficult to know what to hold ourselves and our portfolio companies accountable to and what true compliance means. That said, it is in our best interest as operators and investors to prioritize these factors, not only because it promotes fairness, transparency, and equal representation within our systems, but also because it bolsters the strength and longevity of our companies.

For both operators and investors, ethical AI can be prioritized by;

• Increase your own awareness about (some of the more prominent) standards being put forth that are either already impacting or likely to become relevant for your companies. To get started, check out:
• General Data Protection Regulation (GDPR): GDPR is a regulation in the European Union that sets guidelines for the collection, processing, and storage of personal data, including AI systems that handle personal information.
• AI Ethics Guidelines by the European Commission’s Expert Group on AI: This group published a comprehensive set of guidelines to ensure trustworthy AI. The guidelines include transparency, fairness, accountability, and robustness in AI systems, as well as respect for human autonomy and prevention of harm.
• AI Ethics Guidelines by the United Nations: The United Nations has issued guidelines to assist governments and organizations in harnessing the potential of AI while addressing ethical concerns. These guidelines emphasize transparency, accountability, and avoiding biases in AI systems.
• AI and Machine Learning Ethics Guidelines by Microsoft: Microsoft has published a set of guidelines that provide principles and practices for ethical development and use of AI technologies. The guidelines cover areas such as fairness, reliability, privacy, and inclusiveness.
• Investigate, adopt, and increase investment pipeline in ethical ai tech
• There are several Canadian tech startups making headway in the ethical ai space. For example Private AI (CEO: Patricia Thaine) has built world-class technology that enables companies to become GDPR compliant by finding, redacting, and generating synthetic PII (personally identifiable information) within semi- and unstructured datasets across 49+ languages. The company recently raised a CAD $10.7M Series A.

Specific to investors, we can prioritize AI ethics by incorporating relevant considerations into our diligence process. For example, we can ask our AI companies how they plan to ensure that their datasets are representative of the entire population the outputs represent, how they protect user privacy and adhere to looming or existing regulatory requirements around user privacy, and how they interpret model insights.

What do you think? I’d love to hear your thoughts on how ethics factors into the current AI boom. Email me at sarah@panache.vc to keep the conversation going.