Managing AI Bias with Context

Last year I fired up my terminal and, with a deep breath, typed my first Claude Code command. I'd been thinking about building Remember That Time (RTT) — a collective memory app for close family and social circles — for the better part of five years. I had a product brief, Figma screens, and a UX career's worth of opinions about how it should work. What I didn't have was a way over the engineering wall. Claude Code turned out to be that way.

The terminal is a strange place for a UX person, but Claude made it navigable — and I was having fun. Then I started noticing the pull. Claude had ideas about what I was building, what success looked like, and which direction I should be heading. These weren't plain statements so much as assumptions sitting just behind the advice. The wagon tracks run deep, and they all head the same way. If you don't pay attention and actively redirect, you may arrive somewhere you never intended to go.

The Problem Isn't New — But It's Now Yours

Joy Buolamwini, Timnit Gebru, and Safiya Umoja Noble are important foundational researchers, but mostly read outside the daily press of shipping products. Buolamwini documented how facial recognition systems trained on narrow demographic data fail to see entire populations. Gebru traced how those failures cascade into consequential decisions — hiring, lending, criminal justice. Noble showed how a web search encodes and amplifies racial hierarchy. This research matters and deserves way more attention. But for most PMs, engineers and designers, it stays in the background.

Large language models are trained on a vast corpus — far wider than any single industry. But when it comes to product thinking specifically, one worldview dominates the source material: venture-backed startups, growth-optimization playbooks, engagement metrics as proxies for value. Not because that's all that's in the training data, but because that slice of it is both enormous and internally consistent — it agrees with itself about what good looks like. Those agreements aren't neutral. They encode assumptions about what technology is for, who it serves, what success means.

The Wagon Tracks, Up Close

Early in the RTT development process, before I'd written a line of code, I was in Claude Chat working through the business model. I opened with four explicit constraints: privacy and security were top priority, profit and growth were low on the list, I wanted nothing to do with commodifying personal information, and real user choice — not a 47-page terms of service — was non-negotiable.

Claude reflected my constraints back as design principles. Ruled out advertising, data monetization, dark-pattern consent. And then told me that "timeline and genealogy apps are tough sells without network effects" and that subscriptions "create pressure to retain users with engagement mechanics."

I hadn't asked about network effects. I'd explicitly said growth was low priority. But the evaluative frame hadn't shifted. Network effects, engagement mechanics, retention strategy — these were being offered as neutral product wisdom, not as choices about what kind of product to build. Treating their absence as a liability is a value judgment, and nobody had asked me whether I shared it.

When I pushed back, Claude responded well. The conversation shifted, we worked toward a model that fit the actual goals — a small committed user base paying a fair price, no growth hacking, no investor pressure. Fine — but I knew in the next session it wouldn't remember we'd gotten there. The correction had to outlast the conversation.

A Folder Full of Opinions

One partial solution turned out to be a personal context system that Claude Code reads automatically at the start of every session. These files cover communication style, working preferences, current projects and so on. One file, called philosophy.md, is aimed directly at the bias problem.

It opens with a premise: AI models carry embedded assumptions, most of them reflecting the dominant values of the industry that produced their training data. What follows is a set of beliefs written to counteract those defaults: B-corporations over C-corporations as a starting point, not an alternative to explain away; revenue serving the product rather than the other way around; privacy as a human right rather than a feature toggle.

Writing it required a deliberate conversation with Claude about its own default lean — asking directly what assumptions were most likely to surface in suggestions for a product like RTT, and what it would take to push against them. Claude is capable of that kind of examination when asked. It just doesn't do it unprompted.

The files need maintenance and they don't eliminate drift entirely. What they do is shift the starting position. But what happens in organizations with teams instead of individuals?

Most technology products aren't built by one person in a terminal. They're built by teams, inside companies, under conditions that don't invite philosophical detours. The question of what assumptions your AI collaborator is carrying into the work doesn't make it onto the agenda because there's always something more pressing. And the defaults accumulate — each one small, each one unremarkable, adding up to something nobody quite chose.

The practical extension of a personal context system is a collaborative one: bias guardrails as part of a company's design system. Teams need actual structured files, maintained alongside other design guidelines, that an AI coding or design assistant loads the same way it loads a style guide or component library. The technology already supports it. The habit doesn't exist yet.

Buolamwini, Gebru, and Noble were documenting systems that had already caused harm before most people thought to look. The bias in large language models is earlier in its arc. We're still in the noticing phase — which means there's still room to act rather than just document.

A folder full of opinions isn't a full solution to that. But it's evidence that the defaults aren't fixed. They yield to examination. An AI collaborator pushed off its well-worn tracks will follow you somewhere more interesting. The wagon tracks run deep. They don't run forever.