When the Default Is Not Neutral: Four Inclusion Biases Hidden in Plain Sight
Alice opens her laptop on a Monday morning and decides to sign up for a new professional networking platform she has heard good things about. She is a software engineer based in Nairobi, originally from Ethiopia, and she is excited to connect with peers across the industry.
The first screen asks for her full name. She types her single legal name into the “First Name” field and leaves “Last Name” blank. The form throws an error: Last Name is required. She sighs, types her name again into the second field, and moves on. The platform will address her as “Alice Alice” for as long as she uses it.
The next screen asks her to upload a profile photo for verification. She takes a clear, well-lit photo and uploads it. The system tells her it cannot detect a face in the image. She tries again with more light. Same result. A colleague sitting next to her, who is white, tries the same feature on a whim and it works instantly.
Further into the setup, the platform asks for her date of birth using a dropdown that only recognizes the Gregorian calendar, with no acknowledgment that other calendars exist. Then it asks her to select her gender from two options. She closes the laptop.
Alice has not encountered four separate bugs. She has encountered four separate assumptions, each one baked into a system by people who never had to think about them, because those assumptions described their own lives perfectly. This blog post is about those assumptions, why they persist, and what the computing community can do to dismantle them.
Introduction
Software does not discriminate the way people sometimes do. It does not look at a face, hear an accent, or check a passport. And yet, software can systematically disadvantage certain users in ways that are just as consequential as any deliberate act of exclusion. The diƯerence is that these failures tend to be invisible to the people they do not aƯect.
What follows is an examination of four structural patterns through which computing systems encode cultural defaults and then treat departures from those defaults as errors, anomalies, or lower-quality signals. These are not failures of intent. They are failures of imagination, cases where systems built by a relatively homogeneous group of practitioners were designed around the experiences of that group and deployed globally without revisiting their foundational assumptions. Each issue has known, actionable remedies. The necessary first step is making them visible.
Inclusion Bias #1: The “Default User” Problem and Western Name Structures
The form Alice encountered in our opening story is not unusual. It is, in fact, one of the most common data collection patterns in the world: a mandatory first name field, an optional middle name field, and a mandatory last name field. This structure reflects the naming conventions of a specific cultural tradition and treats that tradition as a universal human standard.
It is not universal. As McKenzie documented in his widely cited enumeration of naming assumptions that programmers routinely get wrong, practitioners carry dozens of mistaken beliefs about names, among them that people have exactly one canonical full name, that names are written in ASCII, that names fit comfortably into a first-name and last-name database schema, and that people who break these assumptions are “weird outliers” rather than the norm across large portions of the world [1]. Across Indonesia, Ethiopia, Iceland, many parts of India, and dozens of other countries and cultures, people legally and culturally hold a single name. In other traditions, family names precede given names, making the “first name” and “last name” labels actively misleading rather than merely inadequate. In some Spanish-speaking cultures, individuals carry two family names, neither of which fits neatly into a single “Last Name” box. In many East Asian contexts, names that appear short in Roman script carry enormous cultural meaning that systems routinely truncate or reject due to character limits [1].
The consequences of getting this wrong are not just cosmetic. When a payroll system duplicates a person’s single name across two fields, legal documents can be generated incorrectly. When an airport check-in system cannot match a passport name to a booking because of how names were stored, a traveler can miss a flight. When a hospital records system garbles a patient’s name, it can create ambiguity in medical records with serious safety implications.
The fix is well understood. Name fields should not impose structural assumptions. A single free-text name field with suƯicient character length covers the full range of human naming conventions. Where structured data is genuinely needed for legal or logistical reasons, systems should be designed with cultural consultants and tested against a representative global dataset before deployment. Mandatory “last name” fields with no provision for single-name users should be treated as what they are: a bug, not a design choice.
Inclusion Bias #2: Skin Tone Bias in Computer Vision Systems
The facial recognition failure Alice encountered is one of the most extensively documented inclusion failures in modern computing. Research by Buolamwini and Gebru published in 2018 quantified the problem with striking precision: commercial gender classification systems had error rates of less than 1% for lighter-skinned males but up to 34.7% for darker-skinned females [2]. The researchers
evaluated three commercial classifiers and found that all of them performed worst on darker female subjects, with the gap between best-case and worst-case performance spanning more than 34 percentage points [2].
The root cause is not mysterious. These systems are trained on datasets that overrepresent lighter-skinned faces. As Yucer et al. document in a comprehensive survey of racial bias in face recognition, the benchmark datasets most commonly used to train and evaluate these models were overwhelmingly composed of lighter-skinned subjects, with some exceeding 86% lighter-skinned representation, so the models learn features that are more salient in those images and perform poorly when the contrast and feature distributions diƯer [3]. Because the people building and testing these systems often belong to the overrepresented group, the disparity goes unnoticed until someone from outside that group encounters it and, more often than not, concludes that they themselves are the problem.
This matters enormously because facial recognition is no longer a novelty. It is used in hiring platforms to screen video interviews, in law enforcement to identify suspects, in healthcare to monitor patient distress, and in financial services to verify identity. A system that reliably fails on darker-skinned users is not a neutral tool with an edge case problem. It is a system that actively excludes a portion of the global population from services that others take for granted.
The path forward requires diversifying training datasets deliberately and auditing model performance across demographic subgroups before deployment. It also requires organizations to ask a harder question: whether deploying a facial recognition system at all is appropriate in contexts where the error rates across groups are not yet equitable enough to justify the decision.
Inclusion Bias #3: Calendar and Date Format Assumptions
The date of birth dropdown Alice encountered assumed that all users reckon time by the Gregorian calendar. This is a strikingly parochial assumption for a global platform to make. The Ethiopian calendar, which Alice uses, is approximately seven to eight years behind the Gregorian calendar depending on the time of year, runs thirteen months rather than twelve, and is the basis of daily life for over 100 million people. The Hebrew calendar, the Islamic Hijri calendar, the Persian calendar, the Chinese lunisolar calendar, and several others are similarly in active everyday use by hundreds of millions of people worldwide, and the W3C’s own guidance on working with time and time zones explicitly acknowledges that systems based on the Gregorian calendar must support conversion to and from calendars including the Ethiopic, Islamic, and Chinese solar-lunar systems [4].
When a system stores dates in a Gregorian-only format without providing culturally appropriate input options, it forces users to perform mental conversions at every interaction, increases the risk of data entry errors, and signals clearly that the system was not designed with them in mind. For date-critical applications like healthcare or financial services, a birthday recorded incorrectly because a user had to convert from one calendar system to another is not a trivial inconvenience.
Beyond calendar systems entirely, the format in which dates are displayed creates its own inclusion failures. The W3C’s Internationalization Working Group has explicitly warned that all-numeric date formats such as “04/05/2024” are inherently ambiguous, since the MM/DD/YYYY convention common in the United States conflicts with the DD/MM/YYYY convention used in most of the rest of the world, and recommends using unambiguous representations — such as spelling out the month name — wherever misreading could cause harm [5]. This is not a theoretical problem: it has caused real errors in contracts, medical dosing schedules, and travel bookings.
Addressing this requires treating date handling the way internationalized software should treat all locale-sensitive data: by separating the stored representation from the displayed representation, supporting locale-aware input methods, and never assuming that a particular calendar or format is self-evidently correct.
Inclusion Bias #4: Binary Gender Fields and the Limits of Categorical Thinking
The final barrier Alice hit was a gender selector with two options. This is a design pattern that has persisted in computing systems for decades even though its inadequacy has been widely understood and discussed for just as long. Stumpf et al., in a comprehensive review of gender inclusive HCI research and design, trace how the assumption of a gender binary is embedded across software systems from video games to mobile applications, creating what they describe as “gendered interfaces” that marginalize users who do not fit the binary model [6]. The review documents how these assumptions are rarely the product of deliberate exclusion but instead reflect the blind spots of design teams that did not pause to consider whether their model of gender matched the diversity of their users [6].
The population of people who do not fit a binary gender model includes nonbinary individuals, intersex individuals, people whose legal gender documentation is in transition, and people in cultural contexts where gender has historically been understood in ways that do not map onto a Western binary framework. This is not a negligible edge case. It is a substantial and growing portion of the global user base, particularly among younger demographics.
The practical solutions depend on what the data is actually needed for. In many cases, gender is collected out of habit or for marketing segmentation purposes that do not justify the harm caused by forcing a choice. In those cases, the correct answer is to stop collecting it. Where gender data serves a genuine purpose, a free-text field, or a structured set of options that includes nonbinary and self
describe choices, is far more inclusive than a binary selector. Where legal or medical contexts require specific recorded values, systems can still oƯer inclusive input while noting the applicable legal categories separately.
Conclusion
Alice’s Monday morning is a composite, but it is not an exaggeration. Millions of people encounter versions of these four biases every day, on forms, platforms, and services that were built without them in mind. What makes these failures particularly diƯicult to address is that they are invisible to the people they do not aƯect. A developer who has a Western two-part name, a lighter skin tone, a Gregorian frame of reference, and a gender that maps cleanly onto a binary selector will build and test a system their whole career without ever noticing any of these problems, because the system will never fail them.
This is precisely why inclusion cannot be treated as a feature to add at the end of a product cycle. It has to be built into the questions teams ask at the beginning: who are we assuming our users are, and what happens to everyone who falls outside that assumption? The answers to those questions, asked honestly and early, are what separate systems that work for the world from systems that work only for the people who built them.
References
[1] P. McKenzie, “Falsehoods programmers believe about names,” Kalzumeus Software, Jun. 16, 2010. [Online]. Available: https://www.kalzumeus.com/2010/06/17/falsehoods-programmers believe-about-names/
[2] J. Buolamwini and T. Gebru, “Gender shades: Intersectional accuracy disparities in commercial gender classification,” in Proc. 1st Conf. Fairness, Accountability and Transparency (FAT*), vol. 81, PMLR, pp. 77–91, 2018. [Online]. Available: https://proceedings.mlr.press/v81/buolamwini18a
[3] S. Yucer, F. Tektas, N. Al Moubayed, and T. P. Breckon, “Racial bias within face recognition: A survey,” arXiv preprint, arXiv:2305.00817, May 2023, doi: 10.48550/arXiv.2305.00817.
[4] A. Phillips, Ed., “Working with time and timezones,” W3C Group Draft Note, World Wide Web Consortium, Jul. 2025. [Online]. Available: https://www.w3.org/TR/timezone/
[5] W3C Internationalization Working Group, “Date formats,” W3C, 2003. [Online]. Available: https://www.w3.org/International/questions/qa-date-format
[6] S. Stumpf, A. Peters, S. Bardzell, M. Burnett, D. Busse, J. Cauchard, and E. Churchill, “Gender inclusive HCI research and design: A conceptual review,” Foundations and Trends in Human– Computer Interaction, vol. 13, no. 1, pp. 1–69, Mar. 2020, doi: 10.1561/1100000056.