A CJ Express Group Director on How AI Learns to Discriminate

A director at one of Thailand's busiest retail data operations says the scariest thing about AI bias isn't a rogue algorithm. It's a spreadsheet. On a recent episode of Students Incorporated, Jarun, who leads the technology innovation group at CJ Express Group, walked student host Premi through a hypothetical he clearly hadn't invented on the spot: a lending model trained on years of loan decisions that happened to reject a disproportionate share of Black applicants. Feed that history back into a machine-learning system, he said, and it doesn't correct the pattern. It learns it.

The Loan Model No One Asked to Discriminate

"If your data has, say, for example, like a race or nationality, the data in the past might tell you that... majority of the people who got loan rejected are the people who may be African-American," Jarun explained. Train a model on that history without correcting for it, he said, and "once the model sees that the person is African-American," it inherits the bias baked into every prior decision.

It's not a hypothetical. In 2021, the nonprofit newsroom The Markup analyzed federal mortgage data nationwide and found lenders 80 percent more likely to deny Black applicants than white applicants with similar financial profiles, 70 percent more likely to deny Native American applicants, and 50 percent more likely to deny Asian and Pacific Islander applicants. Lenders cited "credit history" for 33 percent of Black applicants' denials versus 21 percent of white ones. The investigation was public enough that the Justice Department and the Consumer Financial Protection Bureau cited it directly when announcing a joint crackdown on discriminatory lending later that year.

The pattern hasn't gone away just because regulators noticed it. A November 2025 academic review of financial-algorithm research found systematic disadvantages for minority groups persisting across newer models, including one study in which female applicants received credit scores six to eight points lower than male applicants with comparable finances. Other research has found racial and gender bias surviving even after developers strip out race and gender as direct inputs, because variables like ZIP code and spending patterns serve as stand-ins for the categories the model was supposedly blind to. Regulators are racing to catch up: the EU's AI Act classifies credit scoring and loan underwriting as "high-risk" systems, a classification that becomes legally enforceable on August 2, 2026, meaning developers will have to prove their models don't reproduce the pattern Jarun described.

Whose Call Is It When the Model Says 80 Percent?

Jarun's second example was medical, raising the same question in a different setting. Picture a model that scans a photo of a skin lesion and estimates an 80 percent chance it's cancerous. "Would you just let the model tell you that... this is, you know, 80 percent chance of being having a cancer," he asked, "or would you have it with the judgment of the people?"

That question is being worked out right now by regulators. The FDA still requires full medical-device oversight for any AI function that analyzes medical images to generate a diagnosis, per guidance the agency reaffirmed in January 2026, even as it loosened rules for lower-stakes clinical decision-support software in the same update. The agency's position threads a needle: AI should reduce a physician's workload by flagging likely cases, not replace the sign-off, and clinicians stay legally liable if they lean on a model's output over their own judgment. In practice, hospitals are formalizing the dual-review structure Jarun's question implies: a doctor examines the image independently, decides whether to accept the AI's read, and documents the reasoning either way.

Jarun's framing of both examples landed on the same point: the ethical risk in AI usually isn't a broken model doing something it wasn't designed to do. It's a working model faithfully reproducing a pattern nobody consciously chose to encode.

The Retailer Running These Lessons in Real Time

Jarun isn't a philosopher talking about AI from a distance. CJ Express Group, where he directs the technology innovation group, has spent the past few years turning into one of Thailand's fastest-scaling retail chains. Since 2022, when it brought in supply-chain software from RELEX Solutions to manage inventory across roughly 900 stores and two distribution centers, CJ Express has nearly doubled that footprint to more than 1,700 stores and seven distribution centers by 2025, according to a RELEX case study on the partnership. The company reported 33.97 percent net sales growth in 2025 and an estimated $956 million in annual revenue as of April 2026. Every store generates the transaction, inventory, and customer data that Jarun's team turns into forecasting and personalization models, the same category of system where the bias risks he described actually live.

His own path to that job ran through unglamorous data work that rarely makes it into investor materials. He told Premi he studied electrical engineering, worked on a neural-network senior project before he fully understood what neural networks were for, and later moved to the United States to help build identity-theft and tax-fraud detection models for federal tax returns, work he said saved the IRS "a number of millions of dollars." He returned to Thailand in 2016 to help build one of the country's earlier corporate data science teams, back when that capability was rare enough that his group became known locally as the place that knew how to build it.

Movies Oversold the Robot

Before he got to bias, Jarun named a more basic misconception he runs into constantly: that AI is a general-purpose problem solver that "could do everything for you, could solve everything," a belief he traced to how the technology gets portrayed on screen. The actual work looks nothing like that.

You have to define the problem for AI to solve.

Every model, he said, needs a specific problem and domain-specific data to learn from; no version of AI arrives pre-loaded with judgment about a business or a body of law it's never seen. That gap between the movie version and the engineering reality is also where the bias problem lives. A model that "just works" without anyone examining what it learned from is exactly the kind of system that can absorb a lending desk's history or a hospital's blind spots unnoticed until it's operating at scale.

What He Tells Students Headed Into This Field

Asked what advice he'd give young people aiming at AI and machine learning careers, Jarun didn't lead with a programming language. He led with calculus, statistics, and programming as the baseline, then pivoted to communication and an understanding of "the society" and "the psychology" of the people a system will affect, because, as he put it, the technology "changed the way people live," both individually and collectively.

The data backs up why he'd frame it that way rather than as a pure coding pitch. The World Economic Forum's Future of Jobs research projects AI will help create roughly 170 million new jobs globally by 2030 while displacing about 92 million, with AI-skilled workers commanding a 56 percent wage premium, but nearly 40 percent of employers say the skills gap, not the technology, is their biggest barrier to using AI well. Thailand's own shortage is stark: the country needs roughly 1.08 million high-skilled professionals across its ten targeted industries over the next five years, government workforce plans note, but produces only about 1,500 AI professionals a year against that need. The national strategy responding to that gap is tiered the same way Jarun's advice was: a small core of engineers who can build advanced systems, a larger layer who can apply AI inside real industries, and mass AI literacy for everyone else.

That's the version of an AI career Jarun is actually living day to day: less a wall of code, more a running argument with data about whose history it's allowed to repeat.

← Back to all posts