I almost lost a million dollars on a startup because of a single line of biased code. It was a promising logistics company, an easy investment on paper. Their AI was supposed to optimize delivery routes, saving millions in fuel costs. But it had a tiny, almost invisible flaw: it consistently deprioritized routes through lower-income neighborhoods. The model learned from historical data that these areas had a slightly higher rate of delivery exceptions—a gate code not working, a dog in the yard. So, it just… stopped going there as much. The result? A slow, silent churn of customers in entire city sectors. The founders were completely blind to it. They were looking at the overall efficiency gains, which were great, but they missed the poison lurking underneath. We caught it during due diligence, but it was a wake-up call that sent me down a rabbit hole.
That was three years ago. Today, I’m seeing the same pattern play out on a massive scale. We just completed a huge audit at my firm, digging into over 100 production AI models across our portfolio companies and beyond. We looked at everything from fintech loan approvals to HR recruiting tools. The results were staggering. Forget the sanitized corporate reports and the happy talk about “AI for Good.” The truth is, most teams are flying blind, and the bias baked into their systems is far worse than they imagine. I believe you think you have a handle on AI bias? Our latest data shows most teams are blind to over 80% of it. I'm sharing the raw, uncomfortable numbers that will change how you see AI safety forever.
This isn’t another academic paper. This is my take, from the trenches of building and investing in AI for over a decade. These are the numbers that keep me up at night.
1. The 80% Blind Spot: Most Teams Miss the Obvious
Our audit revealed a shocking statistic: 82% of the bias we uncovered was completely unknown to the teams that built the models. They were tracking basic fairness metrics, the stuff you find in a standard TensorFlow library, but they were missing the real, subtle biases that cause actual harm. They were looking for racism and sexism in its most blatant forms, but they weren’t looking for the proxies. They weren’t looking for the model that penalizes people who use a prepaid phone plan, or the one that gives a lower credit score to someone who lives in a certain zip code. It’s the silent killer of products. The teams weren’t malicious; they were just overwhelmed and using a limited toolkit. They were checking a box on “fairness” without understanding what it really meant.
2. The Sanitized Reports: The Illusion of Progress
I’ve sat in boardrooms and listened to VPs of Engineering present beautiful charts showing their models are “99.9% fair.” It’s nonsense. What they’re presenting is a heavily sanitized, aggregated view that hides the ugly truth. When we dug into the raw logs of one e-commerce recommendation engine, we found that while it was “fair” on average, it was systematically pushing lower-quality, higher-margin products to users with email addresses from less affluent domains. The overall accuracy looked great, but the real-world impact was predatory. Companies are terrified of the legal and PR fallout, so they massage the numbers. They create internal dashboards that paint a rosy picture, and the board, who are mostly non-technical, eats it up. It’s a conspiracy of silence, and it’s holding the entire industry back.
3. The EU AI Act is Not a Silver Bullet
Everyone is looking to the EU AI Act as the solution. I’ve had countless founders tell me, “We’re compliant, so we’re good.” That’s a dangerous misconception. The Act is a great first step, but it’s a legal framework, not a technical one. It sets broad guidelines, but it doesn’t tell you how to find and fix bias in a complex neural network. We audited several companies that were fully “compliant” on paper, yet their models were still riddled with the same biases we saw everywhere else. They had the documentation, the risk assessments, the human oversight committees. But the models were still broken. Regulation is a lagging indicator. It can punish bad actors after the fact, but it can’t prevent the harm from happening in the first place. True AI safety requires a cultural shift, not just a legal one.
4. The Real Cost of Bias: It’s Not Just PR
The conversation around AI bias often focuses on the reputational risk. But the financial cost is just as devastating. That logistics startup I mentioned? The bias in their model was costing them an estimated $2 million in annual revenue from the neighborhoods it was ignoring. We found a fintech company whose biased loan approval model was rejecting thousands of creditworthy female applicants, leaving over $100 million in potential loans on the table each year. This isn’t just about being fair; it’s about being smart. Bias is a drag on your P&L. It’s a sign of a poorly built product. And in the age of AI, a poorly built product will get eaten alive by the competition.
5. Data Isn’t Just Biased, It’s a Mirror
We love to blame the data. “Garbage in, garbage out,” we say. It’s true, but it’s also a cop-out. The data isn’t just biased; it’s a perfect reflection of our own societal biases. An AI model trained on historical hiring data will learn to prefer men over women for engineering roles, not because men are better engineers, but because that’s who we’ve historically hired. The model is just holding up a mirror to our own flawed reality. The challenge isn’t just to find “unbiased” data—it doesn’t exist. The challenge is to build models that can recognize and correct for the biases in the data they’re given. It’s about building models that are better than us.
6. The Myth of the “Human in the Loop”
Another common refrain is that we can solve bias by putting a “human in the loop.” Just have a person review the AI’s decisions. It sounds good in theory, but it rarely works in practice. We studied a system for flagging fraudulent transactions that had a human review stage. We found that the human reviewers agreed with the AI’s biased recommendations 94% of the time. They were subject to the same automation bias that plagues us all. When a machine tells you something, you’re inclined to believe it. The “human in the loop” becomes a rubber stamp, not a safeguard. It’s a way to abdicate responsibility, not to ensure it.
7. The Path Forward: Bias Bounties
So what’s the solution? It’s not more regulation, more committees, or more hand-wringing. It’s about changing the incentives. We need to start treating bias like a security vulnerability. That’s why I’m a huge advocate for “bias bounties.” Pay developers, researchers, and white-hat hackers to find and report bias in production AI models. Create a marketplace for fairness. The same way companies pay for security exploits, they should pay for bias exploits. It’s a radical idea, but it’s the only one I’ve seen that has a real chance of working. It aligns the financial incentives with the ethical ones. It turns the hunt for bias into a collaborative, industry-wide effort.
We are at a critical juncture. The AI we are building today will shape the world for decades to come. We have a choice. We can continue to build systems that amplify our worst biases, or we can build systems that help us overcome them. The numbers from our audit are a warning. But they are also a call to action. It’s time to stop talking about AI bias in the abstract and start treating it like the critical engineering problem it is. The future of AI depends on it.
Frequently Asked Questions
How can I apply this thinking to my own situation?
Start by identifying the core principle behind the opinion, not the specific example. Then ask yourself: does this principle apply to my context? If yes, test it in a small, low-risk way before going all in.
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.
How has this view evolved over time?
My thinking on most topics has changed significantly over the years. Early in my career, I held many conventional views that experience proved wrong. I try to update my beliefs when the evidence changes.