Sad Boomers, Sober Zoomers, Sidelined Gen Xers (again), and No Avotoast for Millenials

It seems as though we have three (and likely additional) alternatives:
‍
1. The images and labels about Boomers have been described in a biased way, perhaps unintentionally or with an unconscious bias by those annotating the datasets;
‍
2. Other, perhaps younger, generations engage in activities that are likely to be labeled as opposite to apathetic, numb, or pensive (e.g., a Millennial having fun at a bar versus a Boomer working in the shop);
‍
3. Perhaps, on average, Boomers are also often considered as a generation that is privacy conscious and therefore do not have as many images available on the internet, especially in expressive or social contexts, to be labeled and annotated in comparison to the younger generations.

Whether we think we can learn about generations from these images depends upon how confident we feel that the training data that went in is an accurate image of the self identity of a group. Some of what we’re getting, especially for the older groups who don’t contribute as much self-generated visual media online [as the younger generations], is perceptions of that group from advertising and media, which we know has inherent flaws.

[...] If we want to argue that these images give us an accurate representation of Boomers, we’d have to accept that there are no black or brown Boomers, and that women are minimally present in the group, and of course neither of these things is true. What I think we actually see is mostly a media representation of Boomers, particularly perhaps in a political context, so we’re seeing mostly blue collar-appearing men solemn-faced or frowning. This suggests to me an image of Boomers as unhappy, having a difficult time, struggling with many parts of life, but I know many Boomers, even in rural or blue collar contexts, for whom this is not at all the case.

Now, Midjourney’s images in general are dark and moody, across all the generational groups, so some of this feeling may just be because of the stylistic choices in the pre-prompting or in the way the training data has been formatted and styled. But these choices and style elements also affect how we as viewers interpret the images, so it’s not meaningless to acknowledge these too.

Gen Z are the first generation to be defined as digital natives. Gen Z has never known a world without the internet and are often the generation at the leading edge of shaping future iterations of the internet and new media trends.

Gen Z and Millennials are all over social media, blogs, and digital content. Because they have such a strong digital footprint, AI models have a lot more data to learn from, which usually makes their representations more varied and nuanced than older generations.

Midjourney and YandexART particularly present this kind of sense: no one’s smiling broadly in the social setting images, unlike the other generations. I’d theorize that party pictures used in training data are most likely to be taken by people themselves or their friends, rather than being put through a media lens, so the model’s representation of such settings might be a useful part of understanding how Gen Z thinks of themselves and want to present themselves to the world.

This may also help us understand why the Millennial and Gen Z images have more faces that don’t fit a traditional white phenotype — because members of racial and ethnic groups underrepresented in advertising and media have the ability to represent themselves directly to their communities and the world around them through images in social media, so those images are available to the models for training.

It’s very visible in this project…For Gen Z and Gen Y (Millennials), in this data we see that the models struggle to “age” the subjects in the output appropriately to the actual age ranges of the generation today.

Both groups seem to look more or less the same age in most cases, with Gen Z sometimes shown (in prompts related to schooling, for example) as actual children. In contrast, Boomers and Gen X are shown primarily in middle age or old age, because the training data that exists is unlikely to have scanned copies of photographs from their younger years, from the 1960s-1990s.

Gen Xers grew up during a time of economic fluctuations, technological transformation from computers to the internet, and significant social changes, which contributed to their overall resilience and adaptability with core values centering around safety, dependability, reliability, and social stability in their surroundings. 
‍
It could be that the training data reflects certain societal narratives of change and uncertainty, like the rise of single-parent households or divorced dads in Gen X. Media portrayals and economic realities like higher divorce rates during Gen X’s prime years might also influence these outputs.

We must seek to reduce the occurrence of [...] stereotypes in the training dataset by including training data from a diverse set of demographics and age groups. If we include only Zoomer images in the training dataset, the model will not be able to generate accurate content for another generation. It is also critical that the team of humans that are planning and deciding how to construct and tune the dataset are also inclusive of the demographics and/or knowledgeable of the diversity that we wish to model using generative AI.

In some ways, I think that we can all feel a little more empathy for ourselves and for other generations by looking at the images that purport to represent us and seeing ‘Hey, that’s not how I see myself!’ and realizing none of us is fully, correctly represented by photographs or media. [...] As I noted earlier, I definitely felt a sense of anxiety in many of the Gen Z images, which may be a clue to further investigation of Gen Z identity and sentiment. I think this is just a starting point, however, and should not be regarded as complete. While it’s interesting, and may be useful, this kind of image generation is not a sufficient substitute for social scientists talking to real people and hearing what they have to say.