The Dummy That Never Quite Fits Like a Person

Opening Scene

A tailor’s dress form is a genuinely useful stand-in for a real customer — it holds a garment’s shape for fitting and adjustment work between actual appointments, and most of the time it’s close enough to be practically valuable. But it’s not the customer. It doesn’t shift its posture, doesn’t breathe, doesn’t sit down. A garment that fits the dummy perfectly can still surprise everyone once the real person actually puts it on, because the dummy was always an approximation, never a perfect substitute.

That’s the honest nature of a proxy feature — a genuinely useful stand-in for something that can’t be measured directly, that still carries real, sometimes important limits.

In Plain English

A proxy feature stands in for something a model actually cares about but can’t directly measure — using zip code as a rough proxy for economic context, or session duration as a proxy for genuine user engagement. Proxies are often the only practical option, and they can be genuinely useful, but they’re never identical to the real thing they’re standing in for, and that gap can matter — sometimes a lot, sometimes systematically, in ways worth understanding before leaning on a proxy too heavily.

The Old Way

Before “proxy feature” had a formal name, people relied on stand-ins constantly, with the same honest limitations:

Using a report card as a proxy for genuine understanding — useful and correlated, but not a perfect measure of what a student actually knows.
Using attendance as a proxy for genuine engagement at work — informative, but someone can be present and disengaged, or genuinely engaged while occasionally absent.
Using a thermometer reading as a proxy for how a room actually feels — correlated with comfort, but not identical to it, since humidity and airflow matter too.

In each case, the proxy carried real, useful signal, while still being honestly distinct from the thing it stood in for.

What’s Changing (and Why AI Is the Reason)

AI systems increasingly rely on a growing web of proxy features for concepts that are genuinely hard to measure directly — creditworthiness, risk, quality — which raises the stakes of understanding exactly where a specific proxy diverges from the real concept it represents.
Tooling can now help quantify how strongly a proxy actually correlates with the real underlying concept, rather than leaving that judgment purely to intuition or convention.
Regulatory and fairness scrutiny of proxy features has increased meaningfully, particularly where a proxy might indirectly encode something it legally or ethically shouldn’t — this series touches this theme briefly here, and this content library’s dedicated series on responsible AI and governance covers it in far more depth.

The Metaphor, Fully Extended

Tailor Shop	Proxy Feature Concept
The dress form standing in for the real customer	A proxy feature standing in for a hard-to-measure concept
A garment that fits the dummy well, most of the time	A proxy that correlates reasonably well with the real target
The dummy’s lack of posture, breathing, or movement	The systematic ways a proxy diverges from the real thing
A surprise once the real person tries the garment on	A model’s real-world failure traced back to proxy limitations
Using the dummy for practical convenience, not as a perfect substitute	Using a proxy feature deliberately, aware of its limits
Occasionally checking the fit against the real customer directly	Periodically validating a proxy against the real concept it represents

For Beginners: What to Actually Do

When using a proxy feature, be explicit — at least to yourself — about what it’s actually standing in for, and how well it’s known to correlate with that real concept.
Look for opportunities to validate a proxy against more direct measurements, even occasionally, rather than trusting it indefinitely without checking.
Be alert to cases where a convenient proxy might be encoding something else entirely, especially anything sensitive or protected.

For Practitioners and Leaders: The Deeper Layer

Document every proxy feature’s known limitations clearly, especially in higher-stakes models — this protects both model quality and the organization’s ability to explain decisions later.
Periodically re-validate proxies against ground truth where possible; the strength of a proxy’s correlation with the real concept can itself drift over time.
Treat proxy features as a genuine fairness and compliance consideration, not just a technical convenience — a proxy correlated with a protected characteristic can create real legal and ethical exposure even without anyone intending it.

Quick Recap

A proxy feature stands in for a concept that can’t be measured directly, and is genuinely useful while never being identical to the real thing.
This mirrors familiar honest stand-ins — report cards, attendance, thermometer readings — correlated with, but distinct from, what they represent.
AI systems’ growing reliance on proxies raises the stakes of understanding exactly where and how they diverge from the real concept.
Proxy features deserve explicit documentation and periodic validation, especially in higher-stakes or regulated contexts.

Where This Fits in the Series

Article 12 covered extracting meaningful signal from timing; this article covered the honest limits of features that stand in for something they can’t directly measure. Article 14 looks at how AI is now helping suggest the first cut on many of these feature engineering decisions.