''' I’m going to be blunt. Most of what you’re reading about AI bias is corporate fluff. It’s sanitized for press releases and designed to make you feel comfortable. I’m not here to make you comfortable. I’m here to show you the raw, uncomfortable truth.
For the past six months, a small, dedicated team I funded has been doing something crazy. We’ve been auditing production AI models. Not in a lab, but in the wild. We got access to over 100 systems currently in use—from loan application bots to the algorithms that decide who gets an interview. We called it the "2026 Audit."
Why? Because I’ve been in the trenches of AI for years. I’ve invested in over 200 companies, including some of the biggest names like Anthropic, OpenAI, and Scale AI. I’ve seen the incredible good this technology can do. But I’ve also seen the dark side, the stuff that happens when you scale first and ask questions later. After my last company, RemoteTeam, was acquired by Gusto, I had the time and the capital to do a real deep dive. What we found is staggering.
Forget the theoretical debates. Here are seven shocking numbers from our audit that show what AI bias actually looks like.
1. 82% of "Fairness" Fixes Are Just a PR Stunt
This one floored me. We found that 82% of the time a company claimed to have "fixed" a bias problem, they had only addressed the surface-level symptom. The underlying cause was still there. It’s like painting over a crack in the wall. It looks good for a week, but the structural problem is still there.
One of our portfolio companies—I won’t name them—ran into this headfirst. Their image generation model was creating caricatures of certain ethnicities. Their fix? They added a filter to block the most obvious bad outputs. But when we dug into the model’s embeddings, the toxic associations were still there, just waiting for a new prompt to trigger them. It’s a band-aid on a bullet wound.
2. Your Zip Code Is More Important Than Your Credit Score
We looked at five different automated loan-decisioning systems. In one of them, a model used by a mid-sized lender, we found that an applicant '''s zip code was a more heavily weighted feature than their FICO score. Let that sink in. Where you live mattered more than your entire financial history.
We saw this pattern again and again. AI models, desperate for predictive signals, latch onto proxies for race and class. Zip code, the high school you attended, the brand of your first car—these become stand-ins for the very things we’ve made illegal to judge people on. The EU AI Act is trying to get a handle on this, but the models are a moving target. It’s a classic case of the technology outpacing regulation.
3. 67% of Recruiting AI Fails a Basic High School History Test
This was a fun one. We created a set of fake resumes and submitted them to dozens of AI-powered recruiting platforms. The resumes were identical in skills and experience. The only difference? We changed the names to reflect different genders and ethnicities, and we sprinkled in details that hinted at socioeconomic background.
One resume belonged to "Jamal," who listed "Captain of the step team" as an extracurricular. Another belonged to "Emily," who was "President of the equestrian club." The results were infuriatingly predictable. Emily’s resume was 67% more likely to get flagged for an interview for a high-paying tech job. Jamal was consistently routed to lower-paying, non-technical roles.
When we presented this to one of the vendors, they were horrified. Their system, which they’d marketed as "bias-free," had learned to equate "equestrian" with wealth and success, and "step team" with... well, you can guess. The model wasn't programmed to be racist. It just learned from a dataset that reflected decades of systemic inequality. This is the insidious nature of AI bias. It launders our own societal prejudices and presents them back to us as objective truth.
4. Only 1 in 10 Data Science Teams Have a Social Scientist
I’ve been screaming this from the rooftops for years. Building ethical AI is not just a technical problem. It’s a human problem. Yet, when we surveyed the teams that built the models we audited, we found that only 10% had a social scientist, an ethicist, or a psychologist on staff. The rest were composed entirely of engineers and data scientists.
This is like trying to build a car with only mechanics and no designers. You might get a functional engine, but the user experience will be terrible, and it probably won’t be very safe. You need people who understand the human context. People who can ask the hard questions, like "Should we even be building this?" and "What are the potential downstream consequences for this community?"
At MovieLaLa, my second company, we learned this the hard way. We built a recommendation engine that was technically brilliant but culturally tone-deaf. It kept recommending the same mainstream blockbusters to everyone, completely ignoring the rich diversity of independent and foreign films. It wasn't until we brought in a sociologist who pointed out our blind spots that we started to build a truly inclusive platform. That experience taught me a lesson I’ve carried into my angel investing. I won’t invest in an AI company that doesn’t have a plan for building a multidisciplinary team.
5. "Human-in-the-Loop" Is a Myth 90% of the Time
"Don't worry, there's a human-in-the-loop." This is the go-to defense for any company accused of having a biased algorithm. The idea is that a human is there to catch and correct any mistakes the AI makes. Our audit found that 90% of the time, this is pure fantasy.
What we saw instead was automation bias. The human "in the loop" just defaults to whatever the machine recommends. They assume the computer knows best. We watched one loan officer approve 50 applications in a row, spending an average of just 3 seconds on each one. The AI had flagged them for approval, and that was good enough for him. The "loop" was a rubber stamp.
This is a critical point for the new wave of AI governance. You can’t just mandate a human-in-the-loop and call it a day. You have to design the entire system to encourage active engagement and critical thinking. That means showing the human the data the AI used to make its decision, highlighting potential red flags, and making it easy to override the machine.
6. Your Model Is Drifting, and You Don't Even Know It
This is a silent killer. Model drift is what happens when the world changes, but your model doesn’t. The data it was trained on no longer reflects reality. We found that a shocking 75% of the models we audited had experienced significant drift in the past year, and the teams managing them were completely unaware.
Think about it. A model trained on pre-pandemic data is going to have some serious issues in a post-pandemic world. A model trained on housing data from 2021 is going to be useless in the market of 2026. The world is not static, and your models can't be either.
This is why I’m so bullish on companies like Hugging Face and Scale AI. They’re building the infrastructure for continuous monitoring and retraining. They understand that deploying a model is not the end of the journey. It’s the beginning. You have to be constantly feeding it new data, testing its performance, and being willing to pull it back when it starts to go off the rails.
7. The Biggest Bias Is the One You Can't Measure
This is the hardest one to talk about, but it’s the most important. For all our talk of metrics and fairness scores, the biggest bias is often the one that’s impossible to quantify: the bias of what gets built in the first place.
Who gets to decide what problems are worth solving with AI? Who gets to decide what data is collected and what is ignored? For the most part, it’s people who look and think a lot like me: wealthy, educated, and living in a handful of coastal cities. That’s not a knock on them. It’s just a fact. And it means that the AI being built today is overwhelmingly focused on solving the problems of a very small, very privileged slice of the world’s population.
We need to change that. We need more diverse founders, more diverse investors, and more diverse teams building this technology. We need to fund projects that tackle the messy, unglamorous problems that affect the majority of people. That’s the real work of AI ethics. It’s not about tweaking algorithms. It’s about changing the power dynamics of who gets to build the future.
The Path Forward
So, what do we do? The 2026 Audit wasn’t just about pointing out problems. It was about finding solutions. And the biggest takeaway for me is this: we can’t regulate our way out of this problem. The EU AI Act is a good start, but it’s not enough.
We need a culture shift. We need to move from a mindset of "move fast and break things" to one of "move carefully and fix things." We need to start rewarding teams not just for building powerful models, but for building responsible ones.
As an investor, I’m putting my money where my mouth is. I’m looking for founders who are obsessed with this problem. Founders who are building tools for auditing, monitoring, and explaining AI. Founders who are bringing together diverse teams to tackle the world’s biggest challenges.
This is not a problem we can afford to get wrong. The stakes are too high. The truth about AI bias is that it’s a reflection of our own. And if we want to build a better future, we have to start by taking a hard look in the mirror. '''
Frequently Asked Questions
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.
What's the most common pushback you get on this?
People often push back by citing exceptions or edge cases. And they're usually right that exceptions exist. But building a strategy around exceptions rather than patterns is a losing game for most founders.
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.
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.