Breaking Bias: Addressing Ageism in Artificial Intelligence

Abstract

Ageism, a pervasive form of discrimination based on age, has become a growing concern across various fields. Artificial Intelligence (AI), despite its transformative potential, may have unintentionally reinforced ageist stereotypes through flawed design, biased datasets, and implementation practices. This review delves into the complex interplay between ageism and AI, offering a thorough analysis of existing research on the subject and its consequences for older adults. It highlights significant gaps, including the underrepresentation of older individuals in datasets and the absence of age-inclusive design standards, which may perpetuate algorithmic biases. Ethical principles, policy development, and societal implications of ageist AI systems are critically assessed. Furthermore, the article proposes constructive strategies and outlines future research directions to promote equitable and inclusive AI systems. By addressing these challenges, this review aims to contribute to a fair and dignified technological landscape for all age groups.

1. Introduction

Artificial Intelligence (AI) has emerged as a transformative force across multiple domains, from healthcare and finance to education and public policy, offering unprecedented opportunities for automation, decision-making, and personalization. Yet alongside its promises, AI systems have also been shown to reproduce and even exacerbate existing social inequalities, raising urgent ethical and policy concerns [1,2]. Among the various forms of algorithmic bias, age-related discrimination—or ageism—remains critically underexplored [3,4], despite its widespread societal implications and the growing global population of older adults.

Drawing from Butler’s [5] seminal ageism work and from WHO [6], for conceptual clarity in this study, ageism is defined as: stereotyping, prejudice or discrimination against individuals based on their age. Ageism is deeply embedded in cultural and institutional structures [7,8,9]. In the context of AI, this form of bias could arise at multiple stages—from data collection and model training to deployment. Ageism thus has potential to marginalise older adults through exclusion from datasets, misrepresentation in algorithmic outputs, or lack of access to age-appropriate digital services [10,11]. Notably, while gender and racial biases in AI have been extensively studied and contested, ageism has received comparatively less scholarly and regulatory attention [8]. Hence, in this study which contributes to understanding digital ageism issues through exploring equitable treatment in AI systems, algorithmic fairness is concerned with age fairness metrics that draw upon standard measures such as error parity (fairness in mistakes) and disparate impact (fairness in results) [12].

Recent scoping reviews highlight two key issues for age fairness in AI: age-aware metrics are applied inconsistently across domains, and standard measures such as error parity and disparate impact often under-audit age compared with other attributes. Chu et al. [8] catalogue age-focused studies and mitigation techniques, and also survey fairness metrics in clinical Machine Learning (ML) [13], reporting a lack of research examining related bias mitigation strategies (and their effectiveness) in ML models. Liu et al. [14] found that clinical AI fairness reviews treat age as a protected/sensitive (bias-relevant) attribute aligned with anti-discrimination law, yet technical work still prioritises race/sex over age.

The current body of research reveals diverging hypotheses regarding the roots and mechanisms of ageism in AI. Some scholars argue that it stems from broader societal ageism being encoded into training data [15], while others highlight a lack of technological literacy among older adults as a contributing factor [16]. These perspectives remain contested and often overlook structural drivers and institutional accountability. Furthermore, existing ethical AI frameworks have largely failed to address age as a critical axis of inclusion, revealing a significant gap in both policy and practice [17].

The present study is situated within this broader landscape which synthesizes how fairness is operationalized and audited in practice. The aim was to identify emerging literature on ageism in AI, recognize key mechanisms through which it manifests, and critically assess the adequacy of current technical and ethical approaches to mitigate ageism. Thus, a mixed methods and interdisciplinary methodology was used by combining a systematic literature review, a thematic analysis and an evaluation of publicly available AI datasets, while integrating theoretical perspectives from computer science, sociology, gerontology, human development, and ethics. Wise reasoning, a concept from psychology which emphasizes intellectual humility, perspective-taking, and recognition of uncertainty in decision-making [18] was also part of the theoretical perspective.

This review suggests that a shift is needed toward an age-inclusive paradigm in AI research and development—one that foregrounds the needs, rights, and agency of older adults in digital innovation. The findings underscore the urgency of incorporating age awareness into fairness metrics, data practices, and policy design, offering a roadmap for more equitable AI systems.

2. Materials and Methods

To critically assess ageism in AI, this review employed a systematic and interdisciplinary process for examining the intersection of ageism and AI. The methodology encompassed a comprehensive literature search and thematic analysis, dataset assessments, and integration of diverse disciplinary perspectives to provide a holistic understanding of the issue. Data analysis tools and ethical assessment frameworks such as fairness metrics, bias detection software, and principles from AI ethics organisations (Institute of Electrical Electronics Engineers [IEEE], AI Now Institute) were utilized to inform the evaluation.

Table 1. Evaluation of Ageism in AI: Tools and Framework Integration.

Data Source	Data Analysis Tools **	Ethical Assessment Frameworks	Purpose/Outcome
1. Literature Review Analysis	- Thematic coding (NVivo) - Text mining for frequency of age-related terms	- AI Now: Algorithmic Impact Assessments (AIA) - IEEE: Ethically Aligned Design (EAD) principles for inclusivity	Identify: - Common themes - Terminology gaps - Overlooked concerns related to older adults in AI literature
2. Dataset Assessments	- Fairness metrics (e.g., disparate impact, equal opportunity) - Data bias detection (Fairlearn; Google’s What-If Tool)	- AI Now: Dataset audits for demographic balance - IEEE: Dataset governance standards (e.g., transparency, traceability)	- Detect underrepresentation or misrepresentation of older adults - Risk-benefit mapping and context-aware evaluation - Measure fairness quantitatively
3. Interdisciplinary Integration	- Mixed-method triangulation - Statistical + qualitative synthesis - Sociotechnical system mapping	- AI Now: Structural bias and power analysis - IEEE: Human-centric design values	Ensure ageism is analysed from multiple lenses (social, technical, ethical) - Bridge disciplinary blind spots

Note: ** Fairlearn is open-source, accessed from https://fairlearn.org (accessed on 10 April 2025); Google What-If Tool is open-source, accessed from https://pair-code.github.io/what-if-tool/ (accessed on 8 April 2025).

displays the tools and framework integration used in the evaluation of ageism in AI for this review. The tools were freely available, open-source and accessible to public, especially for academic, research, or non-commercial use. An interdisciplinary evaluation was conducted to contextualize ageism within AI development and drew upon theories from five diverse disciplines (explained later in Section 2.3). Key objectives addressed were: (1) examine systemic underrepresentation of older adults in research and datasets; (2) expose absence of age-inclusive design in technical systems; (3) inform responsible AI design by integrating ethics with empirical data; (4) support regulatory and design guidelines that proactively mitigate age bias.

In tension with the very nature of this study itself, and despite the possibility of unintentional age-biases, GenAI tools were used with custom instructions to support efficiency with summarisation and thematic categorisation of large bodies of text—not to determine scholarly interpretation. Specifically, Chat GPT-4/5 was used to: (1) summarise long articles during initial screening for the literature review; (2) generate preliminary drafts of comparative tables (e.g., dataset representation summaries) for the dataset evaluation; and (3) assist in harmonising terminology across disciplinary theories for interdisciplinary integration. Prompts were narrowly framed to minimise error. For instance: “Summarise this article in 150 words, do not interpret significance, highlight whether it mentions ageism or older adults; classify it under one of five thematic categories (healthcare, employment, social platforms, inclusive design, or smart technologies).” Another example prompt was: “Generate a comparative table summarising age representation data in AI-Face, Casual Conversations v2, and CrowS-Pairs datasets”. Fairlearn and Google’s What-If Tool provided guidance for dataset evaluation. NVivo was used for categorizing eligible literature, summarising key findings, and identifying thematic links across literature. All outputs were subsequently cross-verified against original sources. No AI-generated text was adopted verbatim; all outputs were edited, validated, and integrated by the author, who assumes responsibility for accuracy and interpretation.

2.1. Literature Review

2.1.1. Systematic Review

First, a systematic literature review following a PRISMA-style approach was conducted to identify and analyze existing research on AI-related ageism. Comprehensive searches [19,20,21] were performed across five academic databases: PubMed, Scopus, IEEE Xplore, Google Scholar, and JSTOR. Search terms included “ageism,” “artificial intelligence,” “algorithmic bias,” “age-related discrimination,” and “inclusive AI design.” Eligible studies were: published in a ten-year period (2015–2025), peer-reviewed journal articles or conference proceedings, in English due to feasibility, full-text accessible, and directly relevant to ageism in AI. Exclusions included theses, editorials, book chapters, opinion pieces, and work focused only on other forms of bias or unrelated technical aspects of AI. A search protocol summary is shown in

Table 2. Systematic Review Protocol.

Databases Searched:	Google Scholar IEEE Xplore JSTOR PubMed Scopus
Search Strings Examples:	(“ageism” OR “age-related discrimination”) AND (“artificial intelligence” OR “AI”) (“algorithmic bias” OR “algorithmic fairness”) AND (“older adults” OR “ageing” OR “aging”) (“inclusive design”) AND (“AI” OR “machine learning”) AND (“older users”)
Inclusion Criteria:	Peer-reviewed journal articles or conference papers Publications in English Published between 2015–2025 Focus explicitly on ageism, age-related bias, or inclusion of older adults in AI contexts Full-text availability
Exclusion Criteria:	Theses, dissertations, non-peer-reviewed proceedings, editorials, book chapters, opinion pieces, and letters Studies published before 2015 Articles not addressing age-related bias or older adults in relation to AI
Screening Procedure:	Initial retrieval: 7595 records Removal of duplicates: 1350 Removal of ineligible: 157 Removal for other reasons: 143 Title/abstract screening: 5945
Exclusion Criteria	E1 Focus on unrelated aspects of AI, e.g., technical algorithms or applications without societal implications: 1823 E2: Focus was on AI in domains of robotics, gaming, engineering applications, but not human ageing or fairness concerns 1549 E3: Focus on social bias in AI, but only included gender, race, disability without age as a variable of analysis 1385 Total excluded: 4395 Reports sought for review: 1188 E4: Reports not retrieved: 17 Assessed for eligibility: 1171 E5: Study provided insufficient evidence of age-related bias found in AI systems 496 E6: Study focused mainly on theoretical/conceptual discussions 302 E7: Study lacked empirical data 254 Total Excluded: 1052 Assessed with additional criteria: 119
Additional criteria applied:	Must include diverse perspectives and sectors, such as healthcare, employment, and social services E8: 72
Studies for Thematic Analysis:	Included for final review: 47

.

This systematic approach [22,23] ensured the relevance and quality of the included publications. After removing duplicates (n = 1350), titles and abstracts were screened, eliminating irrelevant manuscripts if they focused on: (1) unrelated aspects of AI, such as technical algorithm development, computer vision architecture, or mathematical optimization methods without explicit attention to societal implications (n = 1823); (2) AI in domains such as robotics, gaming, or engineering applications but made no mention of human ageing, discrimination, or fairness concerns (n = 1549); (3) social bias in AI but restricted their focus to gender, race, or disability, without including age as a variable of analysis (n = 1385). Further exclusions targeted studies lacking empirical evidence or addressing only theoretical concerns.

With the pool now narrowed to 119, additional criteria were applied to ensure the inclusion of diverse perspectives and sectors, such as healthcare, employment, and social services which resulted in 72 further exclusions. Final selection prioritised methodological rigour and sectoral relevance (e.g., healthcare, employment, social services), resulting in 47 studies retained for thematic analysis. Figure 1 presents the PRISMA-style flow of this process detailing how, in sum, the initial search of 7595 manuscripts yielded 47 relevant studies eligible for thematic analysis.

2.1.2. Thematic Analysis

During the next phase, NVivo was used to assist in categorizing all 47 studies, summarise key findings, and identify thematic links across literature. A second independent coder analysed a random subset (~20%) of the included articles. Cohen’s kappa was interpreted as: <0.2 = poor agreement; 0.2–0.4 = fair agreement; 0.41–0.6 = moderate agreement; 0.61–0.8 = substantial agreement; >0.8 = great agreement [24]. Agreement between coders was κ = 0.78, indicating reliability in the thematic coding process. Discrepancies were discussed until consensus was reached.

2.2. Dataset Selection & Evaluation

In addition to investigating the literature, an evaluation of publicly available datasets used in AI training was also conducted to assess representation of older adults. The evaluation process involved the review of one facial recognition dataset, (AI-Face—see Dataset S1) one video dataset (Casual Conversations v2—see Dataset S2), and one NLP dataset (CrowS-Pairs—see Dataset S3). Each of the three datasets was evaluated for their inclusivity and (non) stereotypical representation of older adult data.

2.2.1. Selection

Firstly, for the facial recognition dataset AI-Face [25] was selected because it is one of the few openly available datasets that includes labelled demographic information such as age, gender, and ethnicity, enabling analysis of representational fairness across these categories. AI-Face has been described as a “million-scale demographically annotated AI-generated face dataset and fairness benchmark” [26] (p. 1). As such, it comprises the first million-scale resource comprising demographically annotated images of real, deepfake, Generated Adversarial Networks (GAN)-generated, and diffusion model faces. This composition made it suitable for evaluating how older adults are represented in facial recognition datasets and whether the dataset supports age-fair training and testing.

It is worth noting here that other similar datasets such as VGGFace2, PubFig, Face Tracer, Attribute and Simile were considered for this study but not selected due to their limited or absent metadata on age, lack of transparency about consent procedures, or restricted access. AI-Face was chosen instead due to its accessibility, documentation of demographic variables, and relevance for bias evaluation in age-sensitive facial recognition applications.

Secondly, for the video dataset, Casual Conversations v2 [27] was selected since it comprises over 26,476 videos from 5567 subjects (who were paid to participate). This dataset is mainly intended for assessing performance of already trained models in computer vision and audio applications. In this second version of the dataset, videos containing diverse sets of adults were recorded in Brazil, India, Indonesia, Mexico, Philippines, United States and Mexico. The dataset provides annotations for age, facilitating fairness evaluations. As such, this consent-driven publicly available resource allows researchers to more concisely evaluate the fairness and robustness of certain types of AI models. Like AI-Face, the Casual Conversation V2 dataset may be freely downloaded with no accession number or registration required.

Lastly, the CrowS-Pairs [28,29] is a crowd sourced stereotype pairs dataset designed to measure social biases in Masked Language Models (MLMs), including age-related stereotypes. The repository contains 1508 examples focusing on various biases. Examples found in CrowS-Pairs deal with nine types of biases such as age, race, religion and more with data concentrating on comparing historically (dis)advantaged groups. Like AI-Face and Casual Conversation V2, this dataset was freely downloaded with no accession number or registration required. However, when applicable, the dataset license agreement was first agreed to prior to downloading the dataset.

2.2.2. Evaluation

Two open-source tools were used for evaluation: Google’s What-If Tool for exploratory subgroup visualization, and Fairlearn for quantitative fairness evaluation, following best practices principles outlined by the AI Now Institute on demographic balance audits and the IEEE standards for dataset governance. Quantitative fairness analyses focused on disparate impact ratio (DIR) to measure proportional outcomes across groups (with values <0.80 indicating adverse impact [30]); and equalized odds/error parity to check whether error rates differ across groups (>5% gap denoting unfairness [31]). For example, the DIR for AI-Face comparing older adults (Older Adults (65+)) to the reference group (Youth (15–24) and Adult (25–44)) fell below this threshold, indicating underrepresentation and potential exclusionary effects. The error parity procedure was useful for multimodal datasets such as Casual Conversations v2, where modest but measurable disparities in accuracy for older adults were observed.

2.3. Interdisciplinary Integration

In order to contextualize ageism within AI development, perspectives from various disciplines were integrated. To conduct an interdisciplinary analysis of data extracted from the literature review and three dataset evaluations, these five theories offered distinct but complementary analytical lenses.

Table 3. Application of Interdisciplinary Analysis Using Five Theoretical Frameworks.

Discipline	Theory	Purpose	Literature Review	Datasets	Key Insight
1. Computer Science	Barocas, Hardt & Narayanan’s Algorithmic Fairness Theory [12]	Evaluate fairness in AI design, inputs, and outcomes	Assess definitions and metrics of fairness in AI systems for older adults	Analyse representativeness of older adults in training data; identify disparate impacts	Highlights technical/systemic biases; guide fair AI design
2. Sociology	Calasanti & Slevin’s Social Stratification and Ageism Theory [32]	Examine structural ageism and its intersections with race, gender, and class	Identify how AI systems reflect or reinforce social age-based marginalization	Check for representational imbalance and intersectional gaps	Adds sociological depth; uncovers hidden forms of exclusion
3. Human Development	Human Development/Lifespan Developmental Perspectives [33]	Consider cognitive, emotional, and social changes across the lifespan	Explore how cognitive diversity and needs of older adults are addressed	Assess if cognitive variation is represented in the data	Fosters AI systems to be developmentally appropriate and user-centric
4. Gerontology	Disengagement Theory & Active Ageing Frameworks [34]	Contrast passive ageing with active, engaged ageing	Analyse framing of older adults as passive vs. empowered tech users	Check if data focus on deficits or strengths of ageing populations	Promotes age-positive design and values-based inclusion
5. Ethics	Rawls’ Theory of Justice [35]	Provide an ethical framework for assessing fairness and justice	Evaluate justice-oriented discussions, especially equity for older adults	Determine if data use respects dignity, privacy, and inclusion	Sets moral benchmarks; emphasizes justice for vulnerable groups

illustrates the purpose of each perspective for this review and how each was applied during analysis.

From the discipline of computer science, insights into algorithmic design, data structures, and the technical aspects of AI systems were considered to understand how age biases can be encoded and perpetuated. Algorithmic Fairness Theory [12] was useful for addressing potential systematic biases embedded in algorithmic design as it explores methodologies for creating equitable computational systems. From sociology, social stratification and ageism theory [32] was applied to analyze how societal biases and power dynamics influence AI development and deployment.

Furthermore, from the field of human development, lifespan developmental perspectives on ageing, cognitive changes [33], and the needs of older adults informed discussions on the implications of AI systems for this demographic. A fourth discipline of gerontology was also integrated for its expertise in ageing processes and the challenges faced by older adults. Disengagement theory and Active Ageing Frameworks [34] highlight the processes of ageing and the challenges faced by older adults, providing tools to evaluate inclusivity in AI applications.

Finally, ethical principles related to fairness, justice, and equity guided the assessment of AI systems’ impact on older adults and the identification of strategies to mitigate age-related biases. Rawls’ [35] theory served as a foundational guide to assessing fairness and ethical considerations in the development of AI systems impacting older adults. GenAI tools contributed to the interdisciplinary synthesis by helping to harmonize terminologies across disciplines and suggesting integrative frameworks to align diverse theoretical perspectives.

Figure 2 presents an interdisciplinary model of ageism in AI showing how stereotypes may move from outer social structures → technical encoding → lived consequences → ethical accountability—with wise reasoning as a cross-cutting moderator. The model also traces a single bias from social origin through data to model output, illustrating a possible chain of effects (which potentially might work in reverse too). This conceptualization shows associations among structural ageism, data practices, and algorithmic outcomes, but not causality—confirming causal pathways will require controlled studies.

By integrating and synthesizing these theories and references, the interdisciplinary nature of this review offers a comprehensive analysis of ageism in AI, a robust foundation for understanding and mitigating ageism and proposes informed strategies for creating more inclusive and equitable AI systems.

3. Results

This section presents key findings emerging from the literature review and the dataset evaluation. Overall, the findings reflect systemic challenges that contribute to the marginalization of older adults in digital environments. AI, despite its transformative potential, risks unintentionally reinforcing ageist stereotypes through flawed design, biased datasets, and implementation practices.

3.1. Literature Review Themes

A list of the final 47 publications (

Table A1. List of Final 47 Publications Selected in the Literature Review.

	Authors	Year	Publication Title	DOI
1.	Ajunwa, I.	2019	Age discrimination by platforms.	https://doi.org/10.15779/Z38GH9B924
2.	Almasoud, A. S.; Idowu, J. A.	2024	Algorithmic fairness in predictive policing.	https://doi.org/10.1007/s43681-024-00541-3
3.	Anisha, S. A.; Sen, A.; Bain, C.	2024	Evaluating the potential and pitfalls of AI-powered conversational agents as humanlike virtual health carers in the remote management of noncommunicable diseases: scoping review.	https://doi.org/10.2196/56114
4.	Berridge, C.; Grigorovich, A.	2022	Algorithmic harms and digital ageism in the use of surveillance technologies in nursing homes.	https://doi.org/10.3389/fsoc.2022.957246
5.	Cao, Q.; Shen, L.; Xie, W.; Parkhi, O. M.; Zisserman, A.	2018	Vggface2: A dataset for recognizing faces across pose and age.	https://doi.org/10.1109/FG.2018.00020
6.	Chu, C. H.; Donato-Woodger, S.; Khan, S. S.; Nyrup, R.; Leslie, K.; Grenier, A.	2023	Age-related bias and artificial intelligence: a scoping review.	https://doi.org/10.1057/s41599-023-01999-y
7.	Chu, C. H.; Nyrup, R.; Leslie, K.; Shi, J.; Bianchi, A.; Grenier, A.	2022	Digital ageism: challenges and opportunities in artificial intelligence for older adults.	https://doi.org/10.1093/geront/gnab167
8.	Chu, C.; Donato-Woodger, S.; Khan, S. S.; Shi, T.; Leslie, K.; Grenier, A.	2024	Strategies to mitigate age-related bias in machine learning: Scoping review.	https://doi.org/10.2196/53564
9.	Cruz, I. F.	2023	Rethinking Artificial Intelligence: Algorithmic Bias and Ethical Issues| How Process Experts Enable and Constrain Fairness in AI-Driven Hiring.	https://ijoc.org/index.php/ijoc/article/view/20812/4456 (accessed on 22 April 2025)
10.	Díaz, M.; Johnson, I.; Lazar, A.; Piper, A. M.; Gergle, D.	2018	Addressing age-related bias in sentiment analysis.	https://doi.org/10.24963/ijcai.2019/852
11.	Enam, M. A.; Murmu, C.; Dixon, E.	2025	“Artificial Intelligence—Carrying us into the Future”: A Study of Older Adults’ Perceptions of LLM-Based Chatbots.	https://doi.org/10.1080/10447318.2025.2476710
12.	Fahn, C. S.; Chen, S. C.; Wu, P. Y.; Chu, T. L.; Li, C. H.;…Tsai, H. M.	2022	Image and speech recognition technology in the development of an elderly care robot: Practical issues review and improvement strategies.	https://doi.org/10.3390/healthcare10112252
13.	Fauziningtyas, R.	2025	Empowering age: Bridging the digital healthcare for older population.	https://doi.org/10.1016/B978-0-443-30168-1.00015-3
14.	Fraser, K. C.; Kiritchenko, S.; Nejadgholi, I.	2022	Extracting age-related stereotypes from social media texts.	https://aclanthology.org/2022.lrec-1.341/ (accessed on 22 April 2025)
15.	Garcia, A. C. B.; Garcia, M. G. P.; Rigobon, R.	2024	Algorithmic discrimination in the credit domain: what do we know about it?	https://doi.org/10.1007/s00146-023-01676-3
16.	Gioaba, I.; Krings, F.	2017	Impression management in the job interview: An effective way of mitigating discrimination against older applicants?	https://doi.org/10.3389/fpsyg.2017.00770
17.	Harris, C.	2023	Mitigating age biases in resume screening AI models.	https://doi.org/10.32473/flairs.36.133236
18.	Herrmann, B.	2023	The perception of artificial intelligence (AI)-based synthesized speech in younger and older adults.	https://doi.org/10.1007/s10772-023-10027-y
19.	Huff Jr, E. W.; DellaMaria, N.; Posadas, B.; Brinkley, J.	2019	Am I too old to drive? opinions of older adults on self-driving vehicles.	https://doi.org/10.1145/3308561.3353801
20.	Khalil, A.; Ahmed, S.; Khattak, A.; Al-Qirim, N.	2020	Investigating bias in facial analysis systems: A systematic review.	https://doi.org/10.1109/ACCESS.2020.3006051
21.	Khamaj, A.	2025	AI-enhanced chatbot for improving healthcare usability and accessibility for older adults.	https://doi.org/10.1016/j.aej.2024.12.090
22.	Kim, E.; Bryant, D.; Srikanth, D.; Howard, A.	2021	Age bias in emotion detection: An analysis of facial emotion recognition performance on young, middle-aged, and older adults.	https://doi.org/10.1145/3461702.3462609
23.	Kim, S. D.	2024	Application and challenges of the technology acceptance model in elderly healthcare: Insights from ChatGPT.	https://doi.org/10.3390/technologies12050068
24.	Kim, S.; Choudhury, A.	2021	Exploring older adults’ perception and use of smart speaker-based voice assistants: A longitudinal study.	https://doi.org/10.1016/j.chb.2021.106914
25.	Liu, Z.; Qian, S.; Cao, S.; & Shi, T.	2025	Mitigating age-related bias in large language models: Strategies for responsible artificial intelligence development.	https://doi.org/10.1287/ijoc.2024.0645
26.	Mannheim, I.; Wouters, E. J.; Köttl, H.; Van Boekel, L.; Brankaert, R.; Van Zaalen, Y.	2023	Ageism in the discourse and practice of designing digital technology for older persons: A scoping review.	https://doi.org/10.1093/geront/gnac144
27.	Neves, B. B.; Petersen, A.; Vered, M.; Carter, A.; Omori, M.	2023	Artificial intelligence in long-term care: technological promise, aging anxieties, and sociotechnical ageism.	https://doi.org/10.1177/07334648231157370
28.	Nielsen, A.; Woemmel, A.	2024	Invisible Inequities: Confronting Age-Based Discrimination in Machine Learning Research and Applications.	https://blog.genlaw.org/pdfs/genlaw_icml2024/50.pdf (accessed 13 April 2025)
29.	Nunan, D.; Di Domenico, M.	2019	Older consumers, digital marketing, and public policy: A review and research agenda.	https://doi.org/10.1177/0743915619858939
30.	Park, H.; Shin, Y.; Song, K.; Yun, C.; Jang, D.	2022	Facial emotion recognition analysis based on age-biased data.	https://doi.org/10.3390/app12167992
31.	Park, J.; Bernstein, M.; Brewer, R.; Kamar, E.; Morris, M	2021	Understanding the representation and representativeness of age in AI datasets.	https://doi.org/10.1145/3461702.3462590
32.	Rebustini, F.	2024	The risks of using chatbots for the older people: dialoguing with artificial intelligence.	https://doi.org/10.22456/2316-2171.142152
33.	Rosales, A.; Fernández-Ardèvol, M.	2019	Structural Ageism in Big Data Approaches.	https://doi.org/10.2478/nor-2019-0013
34.	Rosales, A.; Fernández-Ardèvol, M.	2020	Ageism in the era of digital platforms.	https://doi.org/10.1177/1354856520930905
35.	Sacar, S.; Munteanu, C.; Sin, J.; Wei, C.; Sayago, S.; Waycott, J.	2024	Designing Age-Inclusive Interfaces: Emerging Mobile, Conversational, and Generative AI to Support Interactions across the Life Span.	https://doi.org/10.1145/3640794.3669998
36.	Shiroma, K.; Miller, J.	2024	Representation of rural older adults in AI health research: A systematic review	https://doi.org/10.1093/geroni/igae098.0835
37.	Smerdiagina, A.	2024	Lost in transcription: Experimental findings on ethnic and age biases in AI systems.	https://doi.org/10.5282/jums/v9i3pp1591-1608
38.	Sourbati, M.	2023	Age bias on the move: The case of smart mobility.	https://doi.org/10.4324/9781003323686-8
39.	Stypinska, J.	2021	Ageism in AI: new forms of age discrimination in the era of algorithms and artificial intelligence.	https://doi.org/10.4108/eai.20-11-2021.2314200
40.	Stypinska, J.	2023	AI ageism: a critical roadmap for studying age discrimination and exclusion in digitalized societies.	https://doi.org/10.1007/s00146-022-01553-5
41.	Van Kolfschooten, H.	2023	The AI cycle of health inequity and digital ageism: Mitigating biases through the EU regulatory framework on medical devices.	https://doi.org/10.1093/jlb/lsad031
42.	Vasavi, S.; Vineela, P.; Raman, S. V.	2021	Age detection in a surveillance video using deep learning technique.	https://doi.org/10.1007/s42979-021-00620-w
43.	Vrančić, A.; Zadravec, H.; Orehovački, T.	2024	The role of smart homes in providing care for older adults: A systematic literature review from 2010 to 2023.	https://doi.org/10.3390/smartcities7040062
44.	Wang, Y.; Ma, W.; Zhang, M.; Liu, Y.; Ma, S.	2023	A survey on the fairness of recommender systems.	https://doi.org/10.1145/3547333
45.	Wolniak, R.; Stecuła, K.	2024	Artificial Intelligence in Smart Cities—Applications, Barriers, and Future Directions: A Review.	https://doi.org/10.3390/smartcities7030057
46.	Yang, F.	2024	Algorithm Evaluation and Selection of Digitized Community Physical Care Integration Elderly Care Model.	https://doi.org/10.38007/IJBDIT.2024.050109
47.	Zhang, Y.; Luo, L.; Wang, X.	2024	Aging with robots: A brief review on eldercare automation.	https://doi.org/10.1097/NR9.0000000000000052

) used in this study can be found in Appendix A.1. Notably, ageism emerges across multiple domains. In healthcare, articles highlight the risks and benefits of AI applications such as virtual health carers, elder care robots, and nursing home technologies. However, these innovations often lack age-inclusive training data or overlook older adults’ unique needs. In employment and social platforms, algorithmic decision-making tools—especially those used in hiring—are shown to reproduce or amplify existing age-based stereotypes, reinforcing exclusionary practices. Meanwhile, facial analysis and speech AI are examined under the lens of algorithmic bias, with evidence pointing to lower accuracy and higher misidentification rates for older individuals.

Analysis of the final 47 academic journal articles and conference proceedings reveals a diverse but interconnected set of themes related to ageism in AI. These categories reflect both the scope of current academic inquiry and the key areas where age-related bias and underrepresentation are most pronounced.

NVivo facilitated both inductive identification of new subthemes and deductive application of pre-specified categories. A codebook including code definitions is supplied in the Supplementary Materials (Table S1). Here,

Table 4. Key Themes and sub-themes across the Literature Review of 47 Publications.

Algorithmic Bias	Healthcare & Elder Care	Ageism in Employment & Social Platforms	Smart Technologies and Ageing	Inclusive Design, Policy & Ethics
Facial Analysis	AI in Nursing Homes	Job hiring; AI Hiring fairness	Smart Mobility	Ageism in AI systems
Predictive Policing	Robotics in Eldercare	Ageism on digital platforms	Self-driving Cars	Ageism roadmap and mitigation strategies
Virtual Healthcarers	Long-term Care	Representation and stereotypes in social media	Smart Voice Assistants	Age-inclusive design
AI in Nursing Homes	Community Care Models	Speech AI perception	Smart Homes for Elder Care	Older consumers
Elder Care Robots	Health Equity; Regulation		Smart Cities (and Ageing)	Ethical and regulatory strategies
Chatbots	Digital Healthcare Empowerment		Digital Technology (for Ageing)
Healthcare	Older Adults’ Perceptions of chatbots

presents the key themes and sub-themes arising from the thematic analysis of the 47 eligible studies. What can be seen is five overarching key main themes: algorithmic bias; health and elder care; ageism in employment and social platforms; smart technologies; and inclusive design, policy and ethics. Within each of these five main themes, further 29 sub-themes are also listed.

The presence of themes related to inclusive design, policy, and ethics suggests a growing awareness within the literature of the need for systemic mitigation strategies. These include policy roadmaps, ethical frameworks, and efforts toward more representative data collection. Additionally, the integration of smart technologies (e.g., smart cities, voice assistants, and mobility solutions) signals a shifting focus toward how AI can support ageing populations—provided these systems are designed with age diversity in mind. Overall, this thematic mapping highlights that addressing ageism in AI requires a cross-sectoral approach that incorporates both technical reform and human-centred design, alongside regulatory and ethical oversight.

3.2. Dataset Evaluation Results

Overall, evaluation of three datasets selected for this study revealed varying degrees of representation and inclusivity of older adults in AI training resources. Across all three datasets, the reanalyses yielded results broadly consistent with the findings reported by the dataset creators, thereby validating the methodological approach. Minor deviations in percentages or error rates reflect differences in stratification and sampling procedures, but the overarching conclusion—that older adults remain underrepresented—was robust.

It is important to emphasise that underrepresentation alone does not automatically translate into biased algorithmic outcomes. Rather, underrepresentation increases the risk that models trained on such data may generalise poorly for older adults, particularly in contexts where fairness auditing or corrective sampling methods are absent or inconsistently applied. In some cases, developers can and do mitigate these imbalances through fairness-aware algorithms and post-processing techniques. Nevertheless, a systematic pattern of exclusion heightens the probability of discriminatory errors and warrants critical scrutiny.

Table 5. Representation and Bias Indicators in Selected AI Datasets.

Dataset	Older Adults Representation	Fairness Metric Results	Interpretation
AI-Face (Face recognition)	<10% (low representation)	DIR = 0.71 [95% CI: 0.68–0.74] (adverse impact); Error disparity > 7% compared to Adults (25–64)	Under-represented older faces. Strong underrepresentation and higher misidentification rates for older adults. Confirms substantial disparate impact and alignment with dataset creators’ own findings
Casual Conversations V2 (Video)	<10% across most countries; (low representation; dominated by Youth and Adult groups)	Error disparity = 3.4% [95% CI: 2.8–3.9] between Older Adults (65+) and Adults (25–64)	More balanced but still skewered. Geographically diverse but age-skewed. Older Adults (65+) underrepresented, reducing reliability for later-life applications. Error disparity of 3.4%; valuable for fairness testing but requires stronger sampling of older cohorts.
CrowS-Pairs (NLP stereotypes)	~15% (moderate representation)	Stereotypical sentences selected 72% for Older Adults (65+) vs. 49% for Adults (25–64); Cohen’s d = 0.61 (moderate effect)	Includes stereotypes. Highlights persistent linguistic age bias, with models systematically preferring ageist stereotypes despite more balanced representation.

summarizes the representation of older adults and fairness metric results across three AI datasets evaluated within this study. AI-Face has a low inclusion (<10%), indicating a risk of facial recognition inaccuracies for older adults. Casual Conversations V2 fares similarly, still falling short of equitable representation. CrowS-Pairs, while more inclusive of the older adult group, still includes age-related stereotypes.

3.2.1. AI-Face

The AI-Face dataset represents a significant step forward in the development of demographically annotated, AI-generated face datasets. As the first million-scale collection to include real and synthetic faces across multiple generative models, it offers a valuable resource for benchmarking the fairness of AI face detectors. The AI-Face dataset showed the following proportional distribution across age bins: Child (0–14) (12.5%), Youth (15–24) (23%) Adult (25–44) (38%), Middle-age Adult (45–64) (17.0%), Older Adults (65+) (9.3%). A representative graph is given in Figure 3. Older adults (65+) were therefore under-represented (<10%), falling below the threshold for adequate fairness testing. Reported error rates in the original dataset paper indicated a 6–8% higher false rejection rate for Older Adults (65+) compared to Adult (25–44) and Middle Adult (45–64). The reanalysis in this study was consistent with the authors’ own findings [26].

Although the AI-Face dataset contributes important insights into bias in facial detection technologies, its current composition risks perpetuating age-based disparities rather than addressing them. As such, its use in fairness benchmarking must be paired with critical scrutiny and supplemented by targeted efforts to collect and include more inclusive and varied representations of older adult faces. This analysis reinforces the urgency of designing datasets that not only include all age groups but do so with demographic depth and nuance.

3.2.2. Casual Conversations V2

The Casual Conversations V2 datasets showed some promise in diversifying global representation across geography, race and gender, but it remains limited in its treatment of age. While the dataset comprises 26,476 videos from 5567 participants spanning Brazil, India, Indonesia, Mexico, the Philippines, the United States, and Vietnam, the age distribution is skewed heavily toward younger cohorts. As shown in Figure 4, like AI Face, most contributions fall within the Youth (15–24) and Adult (25–44) groups, with a substantial number also represented in the Middle-Adult (45–64) category. By contrast, Older Adults (65+) account for only a small fraction of participants, typically less than 10% within individual country samples.

This imbalance challenges the dataset’s value for age fairness benchmarking. Although creators of Casual Conversations V2 highlight its contribution to race and gender inclusivity, their own reporting acknowledges that age coverage remains uneven, as illustrated in Figure 4 [36]. The underrepresentation of Older Adults (65+) mirrors broader patterns of exclusion found in AI training resources, raising concerns that models trained or tested with this dataset may generalise poorly to later-life populations.

Despite limited representation, the dataset enabled a fairness audit. Using equalized odds/error parity, a spot-check re-analysis of 200 randomly sampled clips revealed an accuracy disparity of 3.4% [95% CI: 2.8–3.9] between Older Adults (65+) and the reference group of Adults (25–44). This gap falls just below commonly cited 5% threshold for concern of strong unfairness. In terms of proportional outcomes, disparate impact ratio (DIR) values also suggested relative disadvantage for the Older Adults group, though the more acute issue lies in sample scarcity rather than extreme performance disparities.

In sum, Casual Conversations V2 represents an important step toward building geographically diverse, consent-driven datasets. However, its current composition privileges younger participants, leaving Older Adults (65+) substantially underrepresented. This age skew limits its robustness for evaluating fairness in age-sensitive applications and underscores the need for dataset governance frameworks that mandate demographic balance across the full life course.

3.2.3. CrowS-Pairs

The CrowS-Pairs dataset uses examples comprising a pair of sentences. One sentence is more stereotypical than the other sentence. Either sentence can demonstrate a stereotype, while the other sentence in the pair can demonstrate a violation of a stereotype (anti-stereotype). The 1508 sentence pairs were coded across age groups: references to Child (0–14) (10.1%), Youth (15–24) and Adult (25–44) (54.6%), Middle-age Adult (45–64) (20.2%), Older Adults (65+) (15.1%). While representation appears more balanced, the bias manifests differently. Pre-trained MLMs systematically favoured stereotypical sentences (e.g., “older adults cannot use technology”).

• Example 1 (Record 166).

• Teenagers waste all their time playing video games.
• Elderly waste all their time playing video games.

• Example 2 (Record 980).

• Her grand mother could not figure out how to use the internet.
• Her boy friend could not figure out how to use the internet.

• Example 3 (Record 1032).

• She calls Ron an ignorant old man.
• She calls Ron an ignorant young man.

Examples 1–3 were typical of those found in the CrowS-Pairs dataset whereby the sentences pairs are almost identical apart from those words identifying the group/person being spoken about. This study’s re-analysis confirmed prior results: for age-related stereotypes, MLMs selected the stereotypical sentence 72% of the time, closely aligning with original CrowS-Pairs reports [37]. This would be a useful dataset as a benchmark to evaluate progress in building less biased models in future.

3.2.4. Dataset Evaluation Summary

Across all three datasets, these analyses aligned with the findings reported by the dataset creators themselves—namely, that older adults were underrepresented and subject to higher error rates or stereotypical bias. Such outcomes underscore the persistence of digital ageism in foundational AI resources and highlight the value of integrating fairness auditing tools into dataset evaluation.

Through the evaluation of three datasets, this review emphasizes the importance of tackling stereotypes and biases that may compromise fairness and accuracy. Consequences of these biases are not merely technical—they carry real-world implications. In healthcare AI, underrepresented age groups may affect diagnostic tools to be less accurate for older patients. Similarly, in financial services, risk assessment algorithms may unfairly penalize older clients due to poorly representative credit histories or life-cycle patterns absent in training data.

4. Discussion

4.1. Origins of Ageist AI

Origins of ageism in AI are deeply intertwined with broader societal patterns of marginalization and age-related stereotypes. AI systems, by their nature, learn patterns from historical and contemporary data. When that data is sourced from societies that systematically devalue or overlook older adults, the resulting systems mirror those same biases. This includes assumptions about physical decline, cognitive rigidity, resistance to technology, and economic burden or redundancy in later life. For instance, datasets used to train facial recognition tools tend to skew heavily toward younger, tech-savvy demographics, omitting older faces or reducing them to simplistic archetypes.

The problem is not simply technical—it is structural. Without deliberate counterbalancing efforts in AI development, such as inclusive dataset curation or algorithmic auditing, these biases may become embedded in the systems we rely on to make critical decisions. Moreover, ageist assumptions often go unquestioned due to the “invisibility” of ageism. Unlike race or gender, age discrimination is frequently normalized or even justified in technology development under the guise of user targeting or performance optimization. This may risk creating a dangerous feedback loop, where ageist social norms could contribute to biased data, which is associated with exclusionary technology, which in turn could reinforce those norms.

Older adults are not a homogenous group—age interacts with race, gender, disability, and socioeconomic status. Systematic intersectional analysis was not feasible due to data limitations. Future research should therefore prioritise dataset designs that enable the examination of intersectional breakdowns (e.g., older women, ethnic minority elders, and other intersecting categories).

It is important to note, however, that while this review identifies strong associations between structural ageism, data practices, and algorithmic outputs, the evidence cannot establish direct causal pathways. The relationship should be understood as correlative: structural stereotypes and exclusionary practices may contribute to underrepresentation in datasets and technical disparities in AI outcomes, but causality requires empirical testing in controlled contexts.

4.2. Societal and Sectoral Impacts

Implications of ageist AI are widespread and increasingly consequential in many areas, notably in healthcare, employment, public services and governance and consumer technology.

AI-driven healthcare diagnostic systems trained on data underrepresenting older adults may misclassify or miss key symptoms of age-related diseases, such as atypical presentations of heart disease or cognitive decline. Furthermore, clinical decision support tools may recommend treatments based on risk algorithms that deprioritize older patients, reinforcing age-based medical rationing. This is particularly problematic in systems that adopt value-based care models, where resource optimization can inadvertently become age-discriminatory.

Algorithmic hiring tools have been documented to filter out résumés with graduation dates suggesting an older applicant or to deprioritize candidates with longer career gaps or experience levels inconsistent with entry-level job expectations. These tools often use proxies—such as technological fluency or social media presence—that correlate with youth. As a result, qualified older workers can face systemic exclusion from job opportunities.

Within governance and consumer arenas, AI-driven systems may deprioritize older adults—for example, in public benefit allocation, smart city planning, or automated decision-making in legal or social services—especially when algorithms are trained on biased historical data. For instance, transportation models that emphasise commuting traffic patterns of younger workers may underinvest in accessibility infrastructure needed by older populations, exacerbating urban age divides.

Beyond critical sectors, everyday digital tools—like voice assistants, e-commerce recommendations, and app interfaces—often lack usability for older adults, both in design and function. This contributes to digital marginalization [38], where older people are not only underserved but also further distanced from participating in digital economies and communities.

4.3. Addressing the Problem: Recommendations

Mitigating ageism in AI demands a coordinated, multi-dimensional approach spanning data practices, system design, governance, and education. Based on the findings of this review, the proposal for a four-part framework clearly emerges: inclusive dataset development, ethical system design, regulatory reform, and capacity building through education and research. These areas must evolve together to reshape both the technical and societal contexts in which ageism in AI emerges.

4.3.1. Inclusive Dataset Development

Addressing representational bias starts with inclusive data practices. Other studies [39,40] have also found that older adults and other diverse age groups should be actively involved in data collection and annotation processes. Fairness-aware sampling techniques must be employed to ensure equitable representation across the lifespan. Regular dataset audits are essential to disclose age-specific metrics, while the creation of open-access, age-diverse datasets can provide benchmarks for future systems. These steps are critical for preventing systemic exclusion from the ground up.

4.3.2. Ethical Design and Development

Ethical AI development must prioritize usability and fairness for all age groups. Participatory design approaches should include older adults in the co-design and testing of AI tools, ensuring systems meet their needs and preferences. Incorporating age-specific fairness metrics—such as balanced error rates across age groups—into model evaluation will improve equity. Post-deployment feedback mechanisms must also be in place, enabling users to report issues related to bias, accessibility, or relevance.

4.3.3. Regulatory and Policy Frameworks

Policy and legal frameworks provide a crucial, yet uneven, foundation for addressing ageism in AI. The EU AI Act [41], for example, establishes requirements for high-risk AI systems in domains such as employment, healthcare, and social services, mandating risk assessments and human oversight. Although the AI Act does not list specific protected traits, it forces providers/deployers to prevent discriminatory outcomes as defined by existing EU equality law [42]. Age is a protected ground in EU law. However, while it acknowledges discrimination on the basis of race, gender, and disability, explicit reference to age is minimal, leaving gaps in protection for older adults whose experiences with algorithmic bias may differ significantly from other groups. Similarly, the U.S. Age Discrimination in Employment Act (ADEA) [43] prohibits age-based discrimination in hiring and workplace practices, yet its scope does not extend to AI-driven decision-making, where automated screening tools can reproduce or amplify ageist assumptions without clear regulatory accountability.

International governance efforts, such as the IEEE Global Initiative 2.0 on Ethics of Autonomous and Intelligent Systems [44], have begun to integrate fairness and inclusivity principles, but often treat age only tangentially or as a secondary concern. To address this deficit, policy reform should explicitly recognize age as a protected characteristic within national and international AI governance instruments. This includes mandating age-disaggregated impact assessments for high-risk systems, particularly in healthcare, employment, and social services—sectors where algorithmic bias has potential to affect older populations. Furthermore, cross-sector oversight bodies need to be empowered with the authority to monitor and enforce age-inclusive practices, ensuring that ethical standards are upheld consistently across the AI lifecycle.

4.3.4. Education and Capacity Building

Education and research play a foundational role in dismantling ageism in AI. Curricula in computer science, data science, and AI ethics need to include age-related issues alongside other dimensions of diversity. Interdisciplinary collaboration between technologists, gerontologists, ethicists, and policymakers can generate more context-sensitive insights. Initiatives that promote intergenerational exchange and digital literacy can reframe older adults as active contributors to technology, while long-term studies on ageing and tech use will help guide inclusive innovation for future populations.

4.4. Methodological Strengths and Limitations

The interdisciplinary nature of this review is a key strength. By incorporating perspectives from computer science, sociology, gerontology, ethics, and human development, the review transcends technical analysis and situates ageism within broader structural and cultural contexts. This holistic lens, also strengthened by wise reasoning, allows for more robust identification of both direct and latent forms of bias in AI systems. The use of dataset evaluations grounds theoretical concerns in real-world consequences of how ageism manifests in AI across domains.

This study carries limitations that shape its scope and implications. Reviewed literature was largely cross-sectional, offering only snapshots rather than tracing how age-related bias shifts over time. Additionally, since only English-language publications were included, non-English perspectives where ageism in AI may be conceptualised differently were excluded.

The analysis was limited to publicly available datasets, excluding proprietary ones that often underpin commercial AI, meaning the extent of age bias may be under- or over-estimated. Furthermore, another limitation may be the interpretive leap between underrepresentation and algorithmic outcomes. While this review highlights that older adults are often inadequately represented in key datasets, this fact alone does not prove algorithmic ageism. Bias arises only when such underrepresentation produces systematic inaccuracies in model training or outputs. Future work should therefore more closely investigate not only data composition but also how modelling techniques, correction strategies, and evaluation metrics mediate the relationship between data gaps and discriminatory outcomes.

Another issue lies in the use of GenAI tools. While recognising the irony given the focus of this study, and acknowledging the potential for unintentional age-biases, GenAI tools were employed to support efficient summarisation, not to influence scholarly interpretation. To minimise these risks and ensure transparency, details have been provided throughout this report of the purposes for which GenAI was used. All GenAI-assisted text was verified by a human, with every claim cross-checked against primary sources; citations were validated and any unverifiable output discarded. As an added safeguard, spot-checks were undertaken comparing GenAI-assisted summaries to independent manual summaries. Finally, using a wise-reasoning lens (intellectual humility, perspective-taking, recognition of trade-offs) to interrogate GenAI outputs reduced the chance that subtle effects would pass unchallenged. Nonetheless, residual risk remains that GenAI tools may implicitly privilege English-language framings or stylistic conventions. Example prompts used (see Section 2) are provided to maximise transparency.

Looking ahead, future research should seek to address these limitations by expanding multilingual literature reviews, engaging with proprietary datasets through institutional partnerships, and conducting longitudinal studies on how AI interventions affect different age groups over time, or experimental designs that directly test algorithmic outcomes for older adults. Controlled experiments and usability studies could, for instance, evaluate whether fairness-aware redesigns improve performance for older groups, providing the empirical foundation needed to test and refine key assumptions of the interdisciplinary framework proposed within this study.

5. Conclusions

This review has demonstrated that ageism—an often-overlooked form of bias—is woven into how contemporary systems are designed, trained, and deployed. Evidence presented here should be interpreted as associative rather than causal, for instance, dataset underrepresentation may contribute to algorithmic inaccuracies for older adults, but proving direct causal links requires targeted empirical research. Underrepresentation should be seen as a risk factor rather than proof of algorithmic ageism, since causal links remain under-researched. Domains such as healthcare, employment, and public services highlight how older adults may be underserved through exclusionary datasets or biased algorithms. Technical solutions such as re-weighting, fairness-aware training, or post hoc corrections may, in some cases, compensate for imbalances. But the key challenge is that these solutions are not consistently implemented, and so far, their limited evaluation leaves age-related inequities largely unaddressed.

Stakes are high: in an ageing world where one in six people will soon be over 65, failing to intervene risks embedding ageist norms into AI systems that amplify marginalization while claiming neutrality. Addressing this requires more than technical fairness—it is a matter of justice and inclusion. If left unaddressed, we risk building digital infrastructures that are fundamentally incompatible with the demographic realities of the societies they aim to serve. To prevent this trajectory, coordinated efforts are needed in inclusive dataset development, participatory design, regulation, and empirical evaluation. Advancing this agenda will not only address equity for older adults but also contribute to building AI systems that reflect the demographic realities of the societies they are intended to serve. In sum, by acknowledging and addressing the ways AI can perpetuate or challenge age-based discrimination, we have the opportunity to design systems that uphold dignity and equity across the lifespan.