Bias Is a Design Problem, Not a Bug: What UCLA's Digital Humanities Major Says About Building AI Responsibly

UCLA has launched a digital humanities major that requires students to study how AI systems encode bias and to help shape the technology responsibly. The program is not a computer science degree with a humanities garnish. It puts the questions of power, representation and harm at the center of technical work.

That matters because most AI bias failures are not caused by broken code. They are caused by choices made before a single line is written: what data to use, what to optimize for, who gets to review the output, and what counts as success.

What Actually Happened

The Daily Bruin reported that UCLA's digital humanities major is designed to counter AI bias and to prepare students to shape technology responsibly. The mechanism is curriculum design. Students learn to interrogate training data, to question what a model is optimizing for, and to think about who is affected when a system makes a decision.

This is not a product launch or a funding round. It is an educational response to a real and well-documented problem: AI systems that produce discriminatory outcomes because the people building them did not look for those outcomes early enough.

The Failure Mode This Addresses

Bias and algorithmic discrimination is a governance failure that starts long before deployment. It happens when a team treats fairness as a post-launch fix rather than a design constraint.

Three real cases show the pattern.

In 2016, Northpointe's criminal risk score showed large racial disparities in false-positive rates. The tool was used in sentencing decisions. The problem was not that the algorithm was malicious. It was that no one audited error rates across groups before the tool was used on real people. The fix required a fairness audit across groups, explainability, and human oversight. The EU AI Act now treats this kind of use as high risk and requires human oversight under Article 14, with Annex III covering criminal justice tools.

In 2019, Goldman Sachs faced public reports that women were receiving far lower credit limits than their spouses. That triggered a regulator probe. The mechanism was likely a model that had learned from historical lending patterns that reflected past discrimination. The fix is disparate-impact testing before launch, often using the four-fifths rule, plus the protections in the Equal Credit Opportunity Act. The EU AI Act's Article 10 requires data governance that addresses bias in training sets.

In 2023, iTutorGroup's recruiting software automatically rejected applicants over an age threshold. The Equal Employment Opportunity Commission reached a settlement. The mechanism was a simple rule that encoded illegal discrimination. The fix is a bias audit, a review against the Age Discrimination in Employment Act and the Americans with Disabilities Act, and human review of rejections. The EU AI Act's Article 14 again requires human oversight for high-risk hiring tools.

Why This Keeps Happening

The common thread is not bad intentions. It is missing controls.

Teams optimize for accuracy, speed or cost. They do not measure whether the model performs equally well for different groups. They do not test for disparate impact before launch. They do not give a human the authority to override a bad decision. And they do not document who checked what, when, or how.

That is a governance gap, not a technology gap. UCLA's program is trying to close it at the source by training people who will ask those questions before the model ships.

The Controls That Actually Work

If you run an AI system that makes decisions about people, you need three things.

First, a fairness audit across protected groups before launch. The four-fifths rule is a useful starting point: if one group is selected at a rate less than 80 percent of the highest-selected group, you have a problem worth investigating. You also want to measure the equal-opportunity gap, which is the difference in true positive rates between groups.

Second, explainability and human oversight. Someone needs to be able to see why a decision was made and to overturn it. Track how often human review overturns the model. If that number is high, the model is not ready.

Third, documentation. Record what data you used, what you tested for, what you found, and what you did about it. This is what regulators and courts will ask for.

The Obligations You Cannot Ignore

The EU AI Act classifies many of these systems as high risk. That means human oversight under Article 14 and data governance under Article 10. The high-risk rules apply from 2 December 2027 for Annex III systems and 2 August 2028 for embedded Annex I systems. Those dates were set by the Digital Omnibus, Regulation (EU) 2026/1744, which is in force.

In the United States, the Equal Credit Opportunity Act, the Age Discrimination in Employment Act and the Americans with Disabilities Act all apply to automated decisions in lending and hiring. The EEOC has already shown it will pursue cases. In banking, SR 26-2 replaced SR 11-7 in April 2026, but it explicitly leaves generative and agentic AI out of scope, so model risk teams cannot rely on it for these systems.

What to do

• Run a fairness audit before launch, not after. Measure selection rates and error rates by protected group.
• Give a human the authority to override the model, and track how often they do.
• Document your data, your testing, your findings and your fixes. If you cannot show it, you cannot defend it.
• Train the people building and buying these systems to ask who is affected and how you would know.
• Treat bias as a design constraint, not a bug report. The cheapest fix is the one you make before anyone is harmed.

More from our platforms

These sister platforms cover the parts of this problem that sit outside governance.

• Argus (argus.threatclaw.ai) records every trace an AI application produces and scans it for prompt injection, jailbreaks and data leaks, including the attacks hidden inside retrieved documents and tool results rather than in what the user typed. Governance decides what an AI agent is allowed to do. Argus shows what it actually did.
• ThreatClaw (www.threatclaw.ai) tracks the threat side of the same systems: 22 live intelligence feeds, exploitation predicted before it is officially confirmed, threat actor profiles, and detection rules you can deploy straight away. A control is only as good as the threat it is sized against.