Mirror, mirror on the wall, does AI see us all?
The consequences of limited data on the accuracy of AI image generation
Indulge me for a minute; imagine you’re tasked with decorating a house but can only use red. Furniture, walls, art — everything. Sounds ludicrous, right? Yet, this is akin to how some AI image generators operate, trained on such homogeneous data that anything outside a narrow scope turns into a guessing game. It’s funny until it isn’t — when real people and cultures get ignored or misrepresented.
This article aims to shed some light on AI image generators, particularly when generating images of people from specific demographics. These limitations become apparent when uploading a photo to create a representative copy or providing direct prompts to some AI image-generator platforms.
The purpose of this article is not to name and shame specific AI image-generation platforms or language models but to highlight the importance of having diverse and inclusive training data in AI image-generation platforms and language models to avoid potential inaccuracies and biases. When AI systems are trained on a broader range of data representing different ethnicities, cultures, and backgrounds, their outputs can more accurately depict and represent various groups.
The Diversity Deficit
Upton Sinclair’s quote resonates: “It is difficult to get a man to understand something when his salary depends upon his not understanding it.” Sinclair’s quote could well describe the oversight in AI development, where the focus is on immediate gains rather than comprehensive accuracy.
A modern twist on an old saying might be. “It is difficult to get a machine to understand everyone when it’s only trained on a few.” Still, it pinpoints the crux of the issue. The datasets feeding AI often overlook vast swathes of racial and cultural diversity, leading to errors that range from the bizarre to the misleading. These are more than just digital missteps; they’re lost opportunities for AI to truly serve everyone equitably.
What could possibly go wrong?
I’ve been testing out some AI image-generators and found that there are significant limitations to their capabilities. For example, I tried using an image generator that claimed to create an accurate representation of me based on a prompt. However, the results were quite startling. When I asked for an image of a man in a suit standing in line to get coffee from a stall, based on my avatar, the images generated were far from accurate, by some distance.
Sometimes, I appear as the barista, or I’m not even standing in the queue, and at other times, my skin tone is completely wrong (check the hands, people). For those unfamiliar, some AI image-generators struggle with creating realistic hands. While this is understandable, they should at the very least be able to match the correct skin tone.
In another experiment on the same platform, using the exact same photograph, this time focusing on single malt whisky — a personal favourite — the results were again underwhelming and disappointing. Some might say the AI-generated images almost captured my likeness. However, they significantly missed the mark; my hair and hands were far from accurately depicted. These results are not one-offs but symptoms of a widespread issue: AI trained on inadequate, non-diverse data. Astonishingly, this remains a problem in 2024.
It gets better (I am joking; my tongue is firmly embedded in my cheek) when experimenting with custom language models specifically built to replicate or reinterpret uploaded photographs. I’ve encountered some laughable and disturbing results in equal measure. However, I am sharing just few of the experiments; believe me there were plenty more! From these examples, it’s clear these types of AI image-generation tools can be way off the mark; they are far from flawless.
In one of my last experiments, I uploaded photos of a male friend and myself, leading to some rather intriguing outcomes. The AI-generated images not only swapped our ethnicities but, strangely enough even changed my friend’s gender. This gender swap was not a one-off either as I had the same results when uploading a picture of myself and the resulting image presented me as a white female, with short brown hair (not shown). These surprising results were both amusing and a bit disheartening. More importantly, they highlight a significant problem: AI systems that has been fed a too-restricted diet of data.
The Tipping Point: A Poll Perspective
A recent poll from the AI newsletter Mindstream revealed a sobering split: When asked if AI developers should prioritise fixing racial bias in AI-generated images, 50% said, “Yes — It’s 2024!” while the other 50% believed the issue would “fix itself eventually.”
Other than the results, no further data could be obtained, such as the number of voters or their backgrounds, yet this split mirrors troubling apathy. The underlying message? For many, if the problem doesn’t directly impact them, it’s not a priority. But indifference today paves the way for ingrained biases tomorrow.
AI’s limited lens
As humorous as these errors can be, they underscore a severe need: broadening AI’s horizons. Just as you wouldn’t wear ski gear to the beach, AI shouldn’t apply one region’s data to the entire globe’s perspective. The current lack of diversity in AI training materials isn’t just a technical issue — it’s a cultural blind spot that affects how real people are perceived and treated by technology.
Expanding the Spectrum
So, how do we fix this? First, AI developers need to enrich their datasets with a more representative mix of humanity. This means not just adding numbers but enhancing the quality and variety of data sources. Vigilance is crucial to prevent these systems from perpetuating existing biases.
Conclusion and Call to Action
We must not be content with an AI landscape that resembles a monochrome painting when it should be a vibrant, multi-hued mosaic. The richness of global diversity must be mirrored in the data that teaches our machines. It’s not merely about diversification — it’s about enriching the data pool to reflect the total human mosaic. Demand more from tech developers: support initiatives that diversify AI datasets and ensure your voice is heard when creating technology that truly sees everyone.
Have you used AI image-generating tools? What has been your experience? Share your stories to help paint a fuller picture of the need for comprehensive data in AI development.
Thanks for reading all the way to the end!
If you found this article informative, give it a clap or two. 👏👏🏻👏🏼👏🏽👏🏾👏🏿
Once again, please share your thoughts and feedback in the comments below. This will help me improve and also inspire me to create more.
You can find me on Linkedin.
Until next time, keep experimenting and making this world a better place.
Sources: Mindstream, DALL-E (AI image-generated hero image), Various AI image generating platforms and language models. (The names of which have been excluded).
