Understanding Potential Sources of Harm throughout the Machine Learning Life Cycle
Published on:
Article Reading:
Understanding Potential Sources of Harm throughout the Machine Learning Life Cycle — Harini Suresh & John Guttag
When ‘Fair’ Isn’t Enough: Seeing Harm in Machine Learning
Machine learning has become so normal in our daily lives that it almost feels invisible. Whether it’s Spotify recommending songs or Google Photos tagging faces, we rarely stop to ask how these systems make decisions—or who might get hurt along the way. Reading Harini Suresh and John Guttag’s “Understanding Potential Sources of Harm throughout the Machine Learning Life Cycle” made me realize that bias isn’t just about “bad data.” It’s about the human choices built into every stage of the process.
What stood out most to me was how harm can creep in even when everything seems technically correct. For example, Suresh and Guttag talk about representation bias, when a dataset underrepresents parts of a population. That part wasn’t new to me—it’s the classic “algorithm doesn’t recognize darker skin tones” problem. But what really surprised me was the idea of deployment bias, when a model is used differently than how it was designed. A tool meant to predict risk might end up deciding sentencing lengths in court. That’s not a coding flaw, it’s a human one. It made me think about how often we hand off decisions to systems without thinking about their limits.
Another part that hit me was measurement bias. The authors describe how proxies—like using “arrests” to represent “criminal behavior”, can distort reality. That’s something I’d never really questioned before. It reminded me of how social media platforms try to measure “toxicity” or “hate speech” but often mistake certain dialects or slang as offensive. The issue isn’t just the data; it’s the assumption that we can measure something as complex as morality or safety with a single label. Once you start noticing it, you realize how many systems around us do the same, turning complex human experiences into neat numbers.
These ideas connect closely to things I’ve seen outside the classroom. For instance, recommendation systems like TikTok’s “For You” page or YouTube’s suggestions learn from what users click on, then push similar content back to them. It feels personal and accurate, but it’s actually a feedback loop. The model assumes everyone uses the app in the same way, what Suresh and Guttag would call aggregation bias. It’s designed for efficiency, not reflection. The result is that our digital worlds become narrower the more we interact with them.
All of this made me think about accountability. If an algorithm causes harm, who’s responsible? The developers who coded it? The company that deployed it? Or the users who blindly trust it? Suresh and Guttag show that the problem doesn’t live in a single step, it’s spread across an entire lifecycle. That means no one can claim full innocence, but it also means everyone has some power to intervene. Maybe the more honest question isn’t “who’s to blame,” but “who could have asked better questions along the way?”
What I appreciated most about this case study was how it framed bias as a process, not an outcome. It pushed me to stop thinking of fairness as something you can “fix” with more data or better accuracy. Fairness, in this view, is about understanding context—why the data looks the way it does, who benefits from the model, and who gets left behind. Reading it made me more aware of the quiet decisions that shape the technology I use daily. Behind every clean interface and confident prediction is a messy chain of human choices, and recognizing that is the first step toward building systems that actually deserve our trust.