Ethics in AI: Biases & Responsibilities • Michelle Frost & Hannes Lowette

In a conversation with Hannes Lowette, Michelle Frost, an AI advocate at JetBrains, delves into the multifaceted ethical challenges inherent in artificial intelligence, emphasizing that the field is still evolving and requires collective effort to navigate.

The Value Alignment Problem in AI

The core ethical challenge in AI is the “value alignment problem,” which involves finding a middle ground between conflicting human values that are embedded in AI systems. Every decision involves trade-offs, and there is no one-size-fits-all solution. The right approach always “depends” on the specific context, including the company, product, industry, and regulatory environment.

Key examples of these conflicting values include:

• Fairness vs. Accuracy: Improving fairness in models trained on historically biased data can sometimes lead to a decrease in model accuracy. The question then becomes which value to prioritize.
• Privacy vs. Security: Technologies like facial recognition highlight this conflict. Enhancing national security through surveillance comes at the cost of individual privacy.

US Deregulation and a Shift in Responsibility

The discussion highlights a significant shift in the US regulatory landscape. The Biden administration had taken steps to promote responsible AI, introducing the AI Bill of Rights and a significant AI executive order in 2023, with NIST (National Institute of Standards and Technology) leading the creation of frameworks for bias evaluation.

However, the Trump administration immediately repealed this executive order, issuing a new one stating that “AI development should be free from social agendas.” This deregulation puts the onus of ethical development squarely on corporations and individuals. There is a risk that, similar to some DE&I initiatives, corporate efforts in responsible AI may become a “checkbox” exercise or be rolled back entirely. This leads to two major concerns:

1. Lack of Standardization: If each company creates its own rules, it creates a chaotic and inconsistent ethical landscape.
2. Pressure on Consumers: Consumers will have to exert pressure on companies to align with their values.

A Socio-Technical Problem

AI’s challenges are not purely technical; they are socio-technical. The issues we see in AI, such as bias, are amplifications of existing human problems. Therefore, solutions cannot come from just one domain. Meaningful advancement requires a holistic approach that integrates technology, policy, regulation, and law. Focusing on future problems like AGI can sometimes be a “distraction from problems that we still have that need to be addressed” in current machine learning systems.

Market Disruptions: DeepSeek and Project Stargate

The AI landscape is being constantly disrupted. The recent emergence of DeepSeek, a model from a Chinese company, is a prime example. It claims to achieve higher benchmarks at a fraction of the cost of previous models (reportedly trained for around $5.5 million), challenging the notion that cutting-edge AI requires massive capital.

This development casts a new light on initiatives like Project Stargate, a $500 billion AI investment project involving OpenAI and SoftBank. The success of a cheaper, more efficient model like DeepSeek raises questions about the necessity of such colossal investments and opens the door for innovation outside of a few dominant tech giants. This could lead to a shift towards smaller, more specialized, and more efficient models (Artificial Narrow Intelligence) rather than just large, general-purpose ones.

Measuring and Mitigating Bias

Bias in AI is a persistent problem stemming from imbalanced and historically biased source data. Michelle Frost points to the “Adult” dataset from the 1990s as a classic example used to demonstrate how imbalanced data leads to biased outcomes, such as predicting income based on gender.

While we can address bias, the process is complex:

• Measurement: There are over 25 different mathematical metrics to measure fairness, making it difficult for teams to choose the right approach. These metrics can measure individual fairness, group fairness, and predictive equality.
• Mitigation: Techniques exist at various stages:
  • Pre-processing: Correcting and balancing the dataset.
  • In-processing: Making the model itself aware of potential bias during training.
  • Post-processing: Adjusting outcomes after the fact.
• Continuous Management: The most crucial aspect is continuous monitoring and management throughout the AI product’s lifecycle, including transparent reporting to stakeholders.

Accountability in the “Wild West”

A key unanswered question is accountability: “who is responsible?” when an AI system causes harm. The legal and ethical landscape is currently a “wild west,” with various lawsuits testing these boundaries. Responsibility could fall on the AI developer, the company deploying the tool, the end-user, or even a parent in the case of a minor. While organizations like NIST, Microsoft, and IEEE have proposed frameworks, there is no universally accepted standard, creating a “collapsed bridge that we have to build up.”

A Call for Responsible Innovation

Despite the challenges, the outlook is one of “pragmatic optimism.” The potential for AI to improve lives is immense, but achieving it requires intentional, focused effort. Technologists are encouraged to focus on use cases that provide clear human benefits, such as:

• Healthcare: Developing medical devices for epilepsy detection or tools to help people with “locked-in syndrome” communicate.
• Accessibility: Creating tools to help visually impaired individuals with daily tasks, like identifying expired food or sorting laundry.

The final message is a call to action for developers and companies to “not be evil” and consciously choose to work on responsible systems that create a better future.

• Tags:
• LLM
• Feature
• Responsible AI
• AI Ethics
• AI Regulation
• Value Alignment
• Bias
• DeepSeek