How can CIOs course-correct AI for inherited human biases?

4 Min Read

Though AI can be a powerful tool, it can be plagued with the very human biases of the developers who craft its algorithms, as well as the users who feed it information and tailor the prompts it learns from. That means it is up to CIOs and other C-suite tech leaders to ensure team members understand the risks their prejudices and preferences can create.

Despite being born from code, human bias does exist within AI, said Ran Aroussi, CTO at VarOps, an AI transformation and agentic AI implementation firm. "An AI system reasons over what got written down, and what gets written down isn't usually neutral," he said.

For example, directions from senior leadership can become muted, misremembered or misinterpreted as teams jot down or discuss those decisions in company chat groups.

"The system inherits all of that and then answers with confidence, as if it were looking at the whole company," he said.

Related:Cisco chief automation officer on the power of 100x engineers

As an AI system learns from human data and human decisions, it inherits human blind spots, said Justin Kappers, CIO at network security firm Infoblox.

"That can't be fixed with a one-time bias audit," he said.

At Infoblox, every consequential AI-assisted decision has a named business owner, with oversight scaling as autonomy increases. For instance, a tool meant to summarize documents needs only a light review. On the other hand, an AI agent that influences decisions, such as processing a financial transaction, requires human approval, Kappers said.

"The part we potentially miss is that busy teams tend to trust a confident AI answer — we need to guard against such bias by building human judgment in the process," he said.

Human bias is a baked-in problem

AI human bias is pervasive because large language models and AI systems train on historical human data sets, from human-generated data, human-designed objectives and human-defined processes, said Fábio Caversan, CTO at AI and IT consulting firm Stefanini Group.

"They fundamentally act as a mirror to human culture, capturing our institutional assumptions, implicit preferences and historic blind spots," he said.

Caversan noted that bias detection requires end-to-end telemetry paired with continuous behavioral auditing. "The strongest strategy combines technical testing with operational monitoring and human review," he said.

He also suggested that organizations evaluate AI systems before deployment and continue to do so afterward, because data, users, prompts and business conditions change.

Steps to minimize bias

Nate McCarthy, CIO and chief digital information officer at ECG Management Consultants, said it's essential to establish strong processes, particularly in governance and controls. "Educate your people on biases, and recruit for the right mindset, attitude, aptitude and understand human-centered design," he said.

Related:Kraft Group CIO: Training facility revamp brings VR to the gridiron

The next step, McCarthy said, should be to examine how AI technology is viewed within the organization. "You're building AI and the solutions you have embedded in the AI with a way to mitigate bias risk and intervene when it trends in the wrong direction," he said

This approach can also help minimize bias in the culture and workforce for humans and non-humans. "You can then hammer everything home with learning, development, simulation and governance checks with your AI and IT teams."

Caversan stressed the importance of employee education.

"Organizations should ensure that employees know how to do more with the AI they already have, rather than constantly jumping to the next shiny model," he said. He added that better prompting, clearer task definition and stronger verification often produce more value than simply switching models.

When engineers and business users learn how to craft hyper-specific prompts, establish clear guardrails and stress-test midlevel models, they develop a sharper eye for detecting subtle output shifts and biases, Caversan said.

Related:InformationWeek Podcast: Keeping shadow AI agents in check

The root of AI bias

The biggest mistake is treating AI bias as a one-time model problem, rather than an ongoing systems and governance issue. "CIOs can't treat AI bias only as a software engineering problem that can be solved by buying a better model or installing a turnkey tool," Caversan said.

There can be a tendency among CIOs to presume that swapping models or adding static filters can fix a lack of human governance, but that might not suffice. Caversan warned that adding new complexity without training the staff to exercise continuous judgment, or holding human operators accountable, will only accelerate inefficiency and mask underlying risks. The real root of AI bias is human context, data quality and operational design, he added.

Ultimately, AI bias is a leadership and systems-engineering issue, Caversan said. "Organizations shouldn't wait for a perfect governance framework or a perfect AI harness before acting," he said. "They should establish practical controls now."

That includes machine identities, least privilege principles, segmentation, end-to-end tracing, human checkpoints, continuous testing and reliable shutdown mechanisms, he said.

About the Author