The official blog of BNP Paribas Asset Management

Addressing bias in artificial intelligence

Artificial intelligence (AI) has great potential, but care needs to be taken to ensure it doesn’t continue to propagate inherent biases that exist in society and hamper the achievement of diversity and inclusion goals.[1]

AI is a mega-theme that can create significant value across business sectors. It could contribute up to USD 15.7 trillion to the global economy in 2030 – USD 6.6 trillion from increased productivity and USD 9.1 trillion from consumption effects. [2]

While the technology holds tremendous potential, it has a problem with impartiality. If not tackled, this could continue to hinder efforts to enhance diversity, equity and inclusion.

At its core, AI algorithms train on specific datasets and find solutions for real-world problems. As society and organisations increasingly adopt algorithmic decision-making, we must be cognizant of the harm that could arise from algorithmic bias.

Exhibit 1: Automated decision-making using sensitive data such as information on race, gender or familial status can affect individuals’ eligibility for housing, employment, or other core services – the table lists the various spheres of life where automated decision-making can cause injury and notes whether each sort of harmful effect is illegal or unfair

Addressing bias in artificial intelligence


Why do bias related issues surface with AI?

There are several instances of AI-powered systems acting in a discriminatory manner. In one of the most viral examples, a prominent credit card issuer set a woman’s credit limit 20x lower than that of her husband, even though she had better credit scores and a similar financial history.

This brings up a key issue: although the central goal of an optimisation algorithm is not to solve for societal factors, it is critical to understand how the algorithm makes decisions, the input factors it uses and their impact on the final outcome. Serendipitous discovery can take you only so far.

The data sets used to train an AI algorithm influence the efficacy of decision-making. Facial recognition technology, for example, has been found to do more poorly on darker-skinned individuals as image data used for training is skewed towards lighter-skinned individuals.

The results from the gender shades project, which evaluates the accuracy of gender-based products using AI for computer vision, show that facial recognition accuracy is at its worst for dark-skinned females.

What steps are regulators taking to make AI-based algorithms fairer?

The concept of fairness itself is rooted in societal norms, and the trade-offs that people are willing to accept depend on values espoused in the society. In the most recent example, contact tracing, which would have been considered a major privacy violation in the pre-pandemic era, is now widely accepted as the norm given the societal health benefits. Thus, the context is critical when considering issues around gender, ethnicity, age, privacy, etc.

That said, it is essential to have a framework of expected standards when developing AI algorithms and regulatory bodies have started to weigh in on the issue.

The Algorithmic Accountability Act in the US requires periodic assessments of high-risk systems that involve personal information or make automated decisions such as systems that use AI or machine learning. High-risk systems are those that may contribute to inaccuracy, bias or discrimination or facilitate decision-making about sensitive aspects of consumers' lives by evaluating consumer behaviour.

In the EU, the Digital Services Act includes provisions for an ethics framework for AI as well as a future-oriented civil liability framework to help adjudicate AI related issues.

What steps can developers take to eliminate bias in AI-based algorithms?

Given the known issues with AI-based algorithms, developers need to think through the project intent, the impact of system design and possible limitations linked to data availability.

The intent of any new project must be a key consideration. A thorough assessment may provide insights into the unintended consequences and basic human ethical values that could be impacted. For example, in a study published in 2018, algorithms were trained to distinguish the faces of Uyghur people, a predominantly Muslim minority ethnic group in China, from those of Korean and Tibetan ethnicity. This raised concerns in the scientific community as such studies could be used to train surveillance algorithms.

When evaluating the design and output of AI algorithms, developers must assess correlation – the movement of one variable with the other – and causation (cause and effect). The image below depicts how correlation could imply inaccurate linkages.

For AI algorithms, it is critical to identify the variables affecting the outcome and any relationship between them to ensure that biases is not encoded into the decision tree.

Exhibit 2: Explaining correlation and causation

Addressing bias in artificial intelligence


Finally, the output of any AI algorithm is as good as the input data set. In many cases, good data sets are limited outside of the majority sample. Several data de-biasing techniques exist today and many show promising results in reducing bias. However, we must understand the limitations of any technique based on information we have today, and must continuously test and monitor the output samples.


In summary, AI has the potential to provide innovative ways to enable progress on environmental and social issues as well as human welfare. We need, however, to take a step back before we build a system that mirrors the diversity and inclusion issues encountered in the real world. We should be using the latent power of AI to aim for better than the status quo.

Also read:

More on sustainability

Any views expressed here are those of the author as of the date of publication, are based on available information, and are subject to change without notice. Individual portfolio management teams may hold different views and may take different investment decisions for different clients. The views expressed in this podcast do not in any way constitute investment advice.

The value of investments and the income they generate may go down as well as up and it is possible that investors will not recover their initial outlay. Past performance is no guarantee for future returns.

Investing in emerging markets, or specialised or restricted sectors is likely to be subject to a higher-than-average volatility due to a high degree of concentration, greater uncertainty because less information is available, there is less liquidity or due to greater sensitivity to changes in market conditions (social, political and economic conditions).

Some emerging markets offer less security than the majority of international developed markets. For this reason, services for portfolio transactions, liquidation and conservation and conservation on behalf of funds invested in emerging markets may carry greater risk.

[1] Also read Focusing on the ‘S’ in ESG – How disclosure and action can aid diversity on Investors’ Corner

[2] Source: Sizing the prize,

Related articles

Weekly insights, straight to your inbox

A round-up of this week's key economic and market trends, and insights on what to expect going forward.

Please enter a valid email
Please check the boxes below to subscribe