How AI can perpetuate historical inequalities—and what we can do to stop it

From IBM CTO to founder of a startup fighting bias

After 25 years at IBM—including four years on the AI ethics board—Calvin Lawrence, the company’s former CTO for artificial intelligence, left to found IntegrityFirst.ai in 2025. The new company offers the XTegrity platform, which is designed to identify potential harms caused by AI systems, test risks, measure financial and business exposure, audit operational models, and establish ongoing governance.

Why documenting policies isn’t enough

Lawrence explains that during his time at IBM, organizations already had many principles and frameworks in place, but lacked practical tools to show where a model might cause harm and what was being done about it. IntegrityFirst.ai fills this gap by enabling companies to identify, test, and fix concrete problems, rather than just producing documents.

When accuracy doesn’t mean justice

One striking example cited by Lawrence involves a system used to predict criminal recidivism. The model used the number of previous arrests as an indicator. Although the data appears objective, it reflects disproportionate policing patterns: more heavily monitored communities accumulate more records, which leads the algorithm to classify them again as risk areas. In this case, the bias is not a code failure, but the mathematical transformation of historical inequalities into future decisions.

We need to ask: who needs it? Who is it wrong for? And who pays the price when the model makes a mistake?

Bias testing needs to go beyond global metrics

Lawrence warns that declaring that a system has been “tested against bias” is meaningless unless it specifies which groups were examined, with what data, and what the false positive and false negative rates were for each group. He recommends analyzing results broken down by race, gender, age, and other relevant characteristics, and ensuring that corrective actions are taken when disparities are identified.

Regulation is still lagging behind adoption

In Europe, significant parts of the AI Act came into force on August 2, 2026, imposing transparency obligations and requirements for general-purpose models. In the United States, the absence of a comprehensive federal law on AI leaves a regulatory vacuum, while in Brazil, Bill 2338/2023 is advancing in Congress. In all cases, the speed of AI system deployment outpaces the ability to create rules, requiring companies themselves to take responsibility for governing their technologies before regulators intervene.

Empowering the next generation of creators

In addition to his technical work, Lawrence founded AI 4 Black Kids, an initiative that aims to teach Black children not only how to use AI, but to become creators and critics of technology. The goal is for them to learn to program, think critically about automated decisions, and consider entrepreneurship as a path to autonomy.

Looking ahead

Lawrence says he is hopeful but not comfortable. The debate about bias and accountability has advanced, but the pace of AI evolution—especially with generative and autonomous models—requires agile and continuous governance. The central message is that human protection must keep pace with technological progress, and not be left behind.