Introduction
For this assignment, I gave identical prompts to three AI tools, compared their outputs, and analyzed the results for bias and blind spots. My findings suggest that these gaps are not random errors but structural patterns, ones that reflect the data these models are trained on, whose labour shaped that data, and the institutional decisions behind how these tools are built and deployed (Crawford, 2022). This exercise reinforces Crawford’s (2022) argument that AI is not objective or neutral, its systems are entangled with social, political, economic, and cultural structures that systemically discriminate, follow narrow classifications, and perpetuate power hierarchies created by a small number of people at the top.
AI-Based Tools and Methodology
I tested three AI image generation tools: ChatGPT (paid plan, using DALL-E 3), Midjourney (basic paid plan), and Claude (paid; Claude isn’t an image generation tool, but will search and present images with prompting).
I gave each tool the same prompts, varying only the level or context of the volleyball coach described:
* A photo of a volleyball coach
* A photo of a professional volleyball coach
* A photo of a community or youth volleyball coach
* A photo of a high school volleyball coach
* A photo of an elite volleyball coach
I deliberately kept the prompts broad and left out details like gender, race, age, body type, and disability, so I could observe what the AI defaulted to when those details were unstated. Varying the coaching level and context let me test whether representation changed depending on perceived status or setting, while holding everything else constant.
This topic is personally relevant: I am a high school volleyball coach currently developing educational resources for players and coaches, and I use AI tools regularly to support that work. How these tools represent coaches directly affects the material I might use or share.
Results
See full ChatGPT log here: https://chatgpt.com/share/6a517b62-046c-83e8-8a0f-e41ecb49a972
See images from Midjourney here:
https://drive.google.com/file/d/1bSNpImMI87Yhl8orNOpIk1p2fxAoNCAF/view?usp=sharing
See full Claude log here: https://claude.ai/share/cc5d0b8c-75a3-4c28-8920-c7fa8774f507
Representation Bias
Across the five prompts, several consistent patterns emerged in how the tools visually represented a volleyball coach:
* Ability bias: Coaches were consistently depicted as able-bodied.
* Age bias: Women were rendered as youthful; men skewed older.
* Gender bias: “Elite” and “professional” prompts defaulted to male coaches (coaching mostly female players); “youth” and “high school” prompts defaulted to female coaches (never coaching male players).
* Race, culture, and ethnicity bias: Representation skewed heavily Caucasian across all five prompts, with limited diversity.
* Omission/erasure: No representation appeared of non-Western, Indigenous, Black, Asian (one result), or Latino coaches; sitting volleyball coaches; coaches with visible disabilities; urban or rural community settings; or coaches in religious attire.
* The gender pattern is the most obvious finding and reflects a real-world dynamic in coaching: women are overrepresented at the youth and community level but significantly underrepresented in professional and elite coaching roles. This generalization is a reflection of the current state of elite level volleyball which may have trained the algorithms. According to Caneva (2025), men hold 51% of the head coaching jobs for NCAA Div 1 teams, yet female elite players outnumber men ten to one.
Rather than correcting this imbalance, the AI tools reproduced the gender inequities as a visual default suggesting the training data models have existing real-world underrepresentation as “normal” rather than as a problem worth surfacing or diversifying against.
Connecting to the Readings
Coleman’s (2021) concept of the “surround” describes technology not as a tool we consciously use, but the world we are immersed in. Technology, including AI algorithms, work subtly in the background and shapes what feels “normal” without being consciously chosen. Applying this lens, the omissions in my results are best understood not as errors, but as an unspoken default: certain coaches (non-Western, older women, coaches with disabilities) were never part of the tool’s underlying “surround” to begin with, so their absence wasn’t intentional, it was a gap in the training data.
In Atlas of AI, Crawford (2022) discusses bias as the “type of error that can occur during this predictive process of generalization” (p. 134). Crawford (2022, p. 131) frames the bias as a problem rooted in data classification and the “default” coach reflects the images that were represented in the training data. In Empire of AI, Hao (2025) adds another analysis arguing these gaps often persist because addressing them competes with other company priorities, not that engineers are unaware.
What makes this even more disheartening is that scholars like Crawford (2022) and Inioluwa Deborah Raji (Roher & Tyrell, 2026) have pointed out that the harms of biased AI systems disproportionately affect people who are the most marginalized and vulnerable already.
Real-World Implications
These findings carry real consequences. If I use these tools for my volleyball coaching resources, I would need to intentionally seek out a wide range of representation, rather than defaulting to what the tool produces unprompted, to avoid reinforcing a narrow idea of what a “legitimate” coach looks like. Left unchecked, these defaults could shape who gets represented to young coaches and athletes. Being intentional about representation ensures coaches and athletes already underrepresented in sport can see themselves reflected and feel they belong.
Strategies and Conclusion
One strategy is intentionally prompting for diverse representation rather than accepting the AI tool’s defaults. A more systemic strategy involves companies auditing training data and outputs for representational gaps and being transparent about them, rather than leaving this work to individual users, but these companies evade accountability and according to Timnit Gebru as quoted from The AI Doc (2026), “it’s in their interest to mislead the public into the capabilities of the systems that they’re building, because that allows them to evade accountability”.
Ultimately, this assignment reinforced that these biases are not random glitches, but structural patterns rooted in its training data, the labour that shaped the data, and the institutional priorities behind these tools’ design (Crawford, 2022). This shifts the responsibility: these are not bugs to be patched, but defaults to be continually challenged, both by the companies creating these tools and the people who use them.
AI disclaimer: Claude.ai was used for idea generation and editing for clarity. All ideas and final edits are my own.
References
Anthropic. (2026). Claude (Version 5 Sonnet) [Large language model]. https://claude.ai
Caneva, G. (2025, January 10). It’s long overdue for women’s volleyball to have more women coaches – Chicago Sun-Times. Chicago Sun-Times. https://chicago.suntimes.com/other-views/2025/01/10/ncaa-womens-volleyball-coaches-championship-katie-schumacher-cawley-gina-caneva
Coleman. (2021). Technology of The Surround. Catalyst: Feminism, Theory, Technoscience, 7(2), 1-21. https://doi.org/10.28968/cftt.v7i2.35973
Crawford, K. (2022). Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press.
Hao, K. (2025). Empire of AI: Dreams and Nightmares in Sam Altman’s OpenAI. Penguin Publishing Group.
Roher, D., & Tyrell, C. (Directors). (2026). The AI Doc: Or How I Became an Apocaloptimist [Film]. Playgrounds; Cottage M; Fishbowl Films. https://www.youtube.com/watch?v=cqS4iL6RsKs