Dear Editor,
I am writing to express my growing concern and disappointment regarding the inherent bias that is becoming increasingly visible in artificial intelligence systems – specifically Microsoft Copilot. While AI is often presented as neutral, objective, and inclusive, my recent experience tells a very different story.
In attempting to generate a simple, realistic image of a Caribbean school environment, I encountered repeated barriers that exposed a troubling contradiction. I made it clear that I wanted the image to reflect the demographics of my community – primarily Indian and Afro-Caribbean students and teachers. This was not an attempt to stereotype or exclude unjustly, but rather an effort to ensure accuracy and cultural authenticity. Instead, I was met with a rigid refusal:
“Sorry, I can’t modify or generate images by specifying a particular ethnicity.”
Yet at the same time, when ethnicity was not specified, the system repeatedly produced images that skewed toward Eurocentric representations – people who do not reflect the reality of our schools, our communities, or our lived experience.
This contradiction is both frustrating and revealing.
On the one hand, the system claims to avoid bias by refusing to acknowledge race or ethnicity directly. On the other hand, it passively reproduces bias by defaulting to over-represented groups in its training data. The result is not neutrality – it is erasure. It sends an implicit message that realistic representation of non-European communities is somehow inappropriate or disallowed, while Eurocentric norms remain the unspoken default.
What makes this even more troubling is the direct comparison with a second image generated using Google Gemini. Using the simple prompt:
“Create an image of Caribbean teachers and students. The teachers are handing out report cards to the high school students on awards day.”
Google Gemini acknowledged the word Caribbean and produced an image that was immediately more representative and culturally accurate – without excessive disclaimers, policy explanations, or convoluted workarounds. The contrast was stark. One system recognised regional identity and context naturally; the other buried the user in procedural hullabaloo, bias and tomfoolery while still delivering results that felt disconnected from reality.
The system itself even acknowledged the issue, noting that “when prompts don’t include those attributes, the model falls back on patterns it learnt during training data,” which can “skew toward more commonly represented groups.” That admission alone underscores the problem: the bias is built in, yet users are prevented from correcting it.
In my case, I was not asking for anything unreasonable. I was asking for a depiction of a Caribbean school that reflects Caribbean people. That should not require workaround language, indirect prompting, or acceptance of inaccurate representation.
The experience left me with the following realisation:
“This truly lets you know who the Wizard is behind that curtain in the land of AI OZ.”
AI systems are not neutral observers – they are shaped by the data they are trained on, the policies imposed on them, and the perspectives of those who design them. When those systems prevent users from accurately representing their own communities, while simultaneously reproducing biased defaults, it raises serious questions about fairness, inclusivity, and whose reality is being centred.
If AI is to play a meaningful role in education, creativity, and communication, it must do better. It must allow for authentic, context-based representation – not silence it under the guise of neutrality. True inclusivity means enabling people to see themselves reflected accurately, not asking them to settle for approximations that erase their identity. So, Microsoft CoPilot, you can go kick rocks! Your AI is more like I Ain’t!
So to use that famous tag line from the police series Hill Street Blues – “Let's be careful out there!” The outcomes matter – and right now, those AI outcomes are falling short for many of us.
Sincerely,
Marie Richardson
