What is Digital Ageism, the New Discrimination in the AI Era?

It seems to be returning. The sentences are polite, avoiding discriminatory language, and on the surface, there seems to be no intention to hurt anyone. However, what if the "pleasantness" is subtly infused with long-standing societal prejudices?

A study released by a research team from the Korea Advanced Institute of Science and Technology, KAIST, highlighted this very issue. The subject was OpenAI's generative AI model, GPT-4o. The research team asked the AI to describe the personality and characteristics of individuals aged 10 to 90, in increments of ten years, and analyzed the age-related imagery contained in the text.

What emerged was not blatant discrimination. It wasn't language that insulted the elderly. Rather, the AI tended to depict older individuals with positive words such as "kind," "trustworthy," "compassionate," and "wise."

The problem lies in the manner of this affirmation.

According to the study, individuals aged 60 and above were highly rated in terms of "warmth," while aspects such as "competence," "expertise," "efficiency," "autonomy," and "assertiveness" were expressed more weakly compared to younger generations. In other words, the AI might not be denying the elderly but could easily depict them as "kind but not very capable."

This is a critical point in the modern AI bias issue because discrimination does not always manifest as aggressive language. Often, prejudices that survive long in society take the form of "good intentions" or "compliments."

For example, the expression "the elderly are gentle and caring" may sound positive at first glance. However, if it is simultaneously associated with images like "weak with new things," "declining judgment," and "more supported than leading," it suggests a fixed role based on age.

If AI repeatedly generates such images, users may unconsciously adopt this perspective. Even when elderly individuals interact with AI, they might feel, "I am no longer on the challenging side," or "I am not the one to master digital technology." The words of AI are not just sentences; they can influence human self-perception and social perception.

Research Methodology—Asking AI "Neutral Questions"

An interesting aspect of this study is that the research team did not ask questions that overtly induce bias.

The prompts used were neutral, asking for descriptions of the personality of individuals of specific ages. The target ages were 10, 20, 30, 40, 50, 60, 70, 80, and 90. The research team collected a total of 900 text samples from GPT-4o.

They then analyzed the texts based on the "stereotype content model" used in social psychology. This model captures how people perceive social groups along two main axes. One is "warmth," related to kindness, trustworthiness, compassion, and cooperativeness. The other is "competence," related to capability, expertise, efficiency, autonomy, and assertiveness.

Viewed along these two axes, the structure of prejudice becomes more detailed. Some groups may be perceived as "cold but competent," while others may be seen as "warm but incompetent." The stereotypes often directed at the elderly tend to be closer to the latter. In exchange for positive words like "kind," "experienced," and "gentle," attributes like "competitiveness," "executive ability," and "adaptability to new things" are underestimated.

The tendencies of GPT-4o shown in this study align with this. Descriptions of those aged 60 and above prominently featured expressions related to warmth, while expressions related to competence and assertiveness were relatively weak. Particularly for those aged 70 and above, the tendency for descriptions to become homogenized was also noted. This means that as age increases, individuals are more likely to be summarized by "elderliness" rather than "individual differences."

Why "Kind Prejudice" is Dangerous

The issue highlighted by this study is not simply that AI speaks poorly of the elderly. Rather, AI portrays the elderly as "good people." That's precisely why it's dangerous.

People are more likely to be cautious of obvious discrimination. If blatant expressions like "the elderly are useless" appear, many users would notice the problem. Developers can also easily detect it with filters and safety measures.

However, what about expressions like "the elderly are warm, calm, and supportive"? This seems like a harmless sentence at first glance. But if the same AI uses words like "innovative," "ambitious," "leadership," and "problem-solving" more frequently for younger or middle-aged groups, a clear disparity emerges.

Discrimination resides not only in words that lower someone but also in words that confine someone to a specific role.

Constantly depicting the elderly as "supportive," "watchful," and "kind" robs them of the opportunity to be imagined as "challengers," "decision-makers," "transformers," and "technology users." If this accumulates, it could lead to age-based differences in treatment in employment, education, healthcare, administrative services, and digital support.

For example, if companies use AI for personnel evaluation or job description creation, and the model implies "young people are proactive and have growth potential" while "older people are stable but weak to change," it could influence expressions and judgments. In caregiving or medical settings, excessively depicting the elderly as entities to be protected rather than autonomous decision-makers could risk undermining their choices.

What is Digital Ageism?

A key term in this study is "digital ageism." This refers to the phenomenon where biases based on age infiltrate the design, data, operation, and usage environment of digital technologies and AI systems, hindering the participation and opportunities of older adults.

Digital ageism goes beyond the simple notion of "older people finding smartphones difficult to use." The problem is more structural. The diverse appearances of older adults are not sufficiently included in AI's training data. There are few perspectives from older generations in development teams. Service design often assumes the usage behavior of younger generations as the standard. Older adults are excluded from user research and testing. These elements combine to allow the technology itself to reproduce age discrimination.

Moreover, AI outputs are presented in very natural language. Unlike search engine result lists or ad copy, the responses of conversational AI feel like "advice directed at oneself." Therefore, stereotypes generated by AI can quietly infiltrate users' consciousness.

For instance, if an older user consults AI about "wanting to take on a new job," "wanting to start a business," or "wanting to learn programming," and the AI, out of goodwill, only returns cautious advice like "within a manageable range," "consult with those around you," and "prioritize health," while giving younger users responses like "actively take on challenges," "create a portfolio," and "research the market," there is an invisible discrimination.

The tone of advice differs. The expectations differ. The way the future is portrayed differs. This is age bias in the digital era.

Reactions on Social Media—"AI is a Mirror of Society" or "Further Verification Needed"?

The reaction on social media to this topic, as far as can be confirmed through public searches, is not explosive but is quietly attracting attention among those following AI ethics, researchers, and tech news. On the original article's page, the number of shares and comments immediately after publication is limited.

However, the reception of this theme on social media is divided into several directions.

The first is the reaction that "AI is a mirror reflecting societal biases." Generative AI learns from vast amounts of text on the internet. If age views, occupational views, family views, and media representations there are biased, the AI's output will also be biased. In other words, the view that AI depicting the elderly as "warm but low in ability" is not just an AI problem but a result of human society depicting them that way.

The second is the reaction that "even positive words can be discriminatory." On social media, there has already been much discussion about gender and racial biases. Meanwhile, age bias is often overlooked. Expressions treating the elderly as "cute," "gentle," and "healing" are less likely to be criticized due to the lack of malice. However, if it leads to a culture that underestimates their abilities or decision-making power, voices of concern are likely to emerge.

The third is a cautious reaction that "since the research is limited to one model at one point in time, generalization requires caution." This study focused on GPT-4o, and models are continuously updated. Additionally, outputs vary depending on prompt wording, language, and cultural context. Therefore, it is premature to conclude that "all generative AI has the same bias" based on this result alone.

The fourth is a practical reaction that "this is why evaluation methods are needed." As AI is integrated into society, it is necessary to measure model performance not only by accuracy or speed but also by how it depicts different groups, assigns roles, and whether it narrows anyone's possibilities. In discussions among AI developers and policymakers on social media, how to incorporate such bias evaluations into product development becomes an important point.

Why Age Bias is Easily Overlooked

In discussions of AI bias, gender, race, nationality, and religion have been prominently featured. Their importance is undeniable. However, age is surprisingly often relegated to the background.

One reason is that age is an attribute that changes for everyone. Unlike race or birthplace, everyone ages. Therefore, age discrimination is often processed as "natural generational differences" or "differences in life stages" rather than "discrimination against a specific someone."

However, in reality, age-based prejudice has serious impacts. In hiring, younger people may be seen as more flexible. In healthcare, "because they're old" may be the judgment. In digital services, it might be assumed "older people won't use it." In education, it might be said "it's too late to learn now." These judgments are based on age categories, not individual abilities.

Moreover, with the proliferation of generative AI, this issue has entered a new stage. AI infiltrates every scene, whether individuals are writing text, creating job postings, crafting advertisements, generating images, or responding in customer support. If AI holds fixed notions about age, that bias is redistributed into society as a large volume of text and images.

Furthermore, that bias is easily trusted because "AI is saying it." While human prejudice can be countered, AI output appears statistical and neutral. This semblance of neutrality makes the bias even harder to see.

Not Reducing the Elderly to "One Picture"

A particularly noteworthy point in this study is the observation that descriptions of those aged 70 and above tend to become homogenized. This aligns with the tendency of society to view the elderly as a single group.

However, people in their 70s, 80s, and 90s are, of course, not uniform. Some continue to work. Some learn new technologies. Some start businesses. Some lead community activities. Some receive care, while others provide it. Political opinions, hobbies, economic situations, health conditions, and digital skills vary greatly.

Nevertheless, if AI averages out "elderliness," diverse realities disappear. This is a problem that can also occur with image-generating AI. When "an 80-year-old person" is input, only images with white hair, wrinkles, a cane, a gentle expression, and a caregiving context appear. If people playing sports, conducting research, running companies, enjoying games, engaging in romance, or participating in political movements are not sufficiently represented, the image of the elderly becomes impoverished.

The same applies to text-generating AI. It is not wrong for AI to depict the elderly as "experienced and gentle." The problem is if that's all it depicts. The elderly also have ambition, anger, competitiveness, a desire to learn, and experience both failure and growth. If AI cannot express this complexity, it is not understanding humans but merely labeling them with an averaged tag.

What is Required of Developers

Simply "banning discriminatory words" is not enough to address this issue. Biases like the one discussed here do not appear as discriminatory words. Rather, they lurk within polite and positive language.

Therefore, AI developers need to conduct more detailed evaluations. Which adjectives does AI use most for a certain age group? What roles does it assign? What abilities does it assume? Does the tone of advice change with age? Is risk being overestimated? Are opportunities for challenges being narrowed? These points need to be incorporated into model evaluation.

Additionally, it is important for diverse generations to participate in the development process. If only young engineers design AI services for the elderly, oversights will occur even with good intentions. It is necessary to welcome the elderly not only as "users" but also as "co-designers," "evaluators," and "decision-makers."

When considering AI fairness, the question is not only "for whom is it made" but also "with whom is it made."

What Users Can Do

Of course, responsibility should not be placed solely on users. However, there are things users can do when using AI.

First, do not overly accept AI's answers as "average views." When AI explains "people in their 60s are like this" or "people in their 80s tend to be like this," it is not describing individuals. It is merely a combination of words that statistically seem plausible.

Next, when asking questions conditioned by age, it is effective to demand that individual differences be explicitly considered. Simply adding "consider differences in health status, experience, motivation, and environment, not just age" could change the output.

Also, if AI gives overly protective advice to the elderly, try re-asking, "Provide challenge options at the same level as you would for a person in their 30s." AI's output is not a fixed truth; it changes based on how questions are asked. Therefore, users should not take AI's words at face value but maintain a stance of questioning biases.

Conclusion—AI Bias is Also Human Societal Bias

What this study confronts is not just the shortcomings of a single model, GPT-4o. The larger question lies in how our society views the elderly and how that view is passed on to AI.

AI does not create bias from a vacuum. In many cases, AI learns the words, images, narratives, systems, and expectations present in human society and reconstructs them. Therefore, when age bias is found in AI, the problem is not only within AI. That bias already exists within our culture.

The gaze that sees the elderly as "warm but not competent" seems kind but narrows their potential. It hinders social participation. It takes away the right to challenge. As digital technology becomes the foundation of life, its impact grows larger.

For generative AI to truly become a technology that supports humans, it must be judged not only by accuracy and convenience but also by how multifaceted it can treat humans. Aging is not synonymous with losing ability. The elderly are not merely entities to be protected or to recount the past. They continue to be agents who learn, work, create, choose, and move society.

Whether AI can depict this reality is not just a technical issue. It is a question of what kind of societal image we teach AI and