Mitigating Stigma and Fostering Support: Improving AI-Generated Counterspeech for Microaggressions

Abstract

In societies where sexual and reproductive health (SRH) is stigmatized, many women hesitate to seek care, increasing health risks. In South Korea, cultural norms associating promiscuity with SRH care in unmarried women further discourage them from accessing care. While online spaces offer support, they also perpetuate stigma through microaggressions. To mitigate the harms of microaggressions, counterspeech provides a promising approach. This study examines counterspeech by generative artificial intelligence (AI) using ChatGPT-4 and Copilot GPT-4, analyzes the strategies AI tools claim to use, evaluates their alignment with recommended counterspeech strategies, and identifies potential harms. Our findings reveal critical limitations: failures to recognize implicit biases and challenge relevant stereotypes, placing the burden of addressing microaggressions onto those who experience them, and offering only superficial empathy. We propose a new process for AI to foster more effective and culturally sensitive counterspeech. With these improvements, AI could help create safer, more inclusive spaces for people seeking support for stigmatized healthcare.

Introduction

People across different sociocultural contexts often face stigma when seeking healthcare for conditions perceived as sensitive, controversial, or subject to societal scrutiny^1,2,3. For instance, those seeking sexual and reproductive health (SRH), Human Immunodeficiency Virus, or abortion care frequently encounter societal judgment and discrimination^1,2,4. Stigma manifests in various ways, from institutional barriers to interpersonal interactions, often discouraging individuals from accessing necessary care^4,5. Microaggressions—subtle, everyday comments or behaviors that reinforce stereotypes and invalidate experiences—play a role in reinforcing this stigma⁶. While often unintentional, these remarks reflect underlying societal norms and can make it more difficult for people facing stigma to seek care without fear of judgment^5,6.

Unmarried Korean women face heightened challenges when seeking SRH care, as prevailing cultural beliefs associate premarital SRH concerns with societally shamed sexual activity^2,4,5. This perception casts their health-seeking behaviors as morally questionable, subjecting them to judgment and scrutiny^2,4,5. As a result, many unmarried women hesitate to access SRH services due to the fear of social repercussions from healthcare providers, peers, or even family members^4,5. This delay or avoidance of care can lead to negative health consequences, including undiagnosed and untreated sexually transmitted infections, delayed detection of reproductive health conditions such as polycystic ovary syndrome or cervical cancer, and inadequate access to contraception or menstrual health management⁴. In contrast, while married women may still encounter stigma, their SRH concerns are often framed within the context of family and motherhood, granting them comparatively greater access to care. For unmarried women, the lack of socially accepted justification for seeking SRH services isolates them further and discourages proactive health-seeking behaviors^4,5.

To navigate these challenges, many unmarried Korean women turn to online spaces to seek information and support for their SRH concerns. These platforms, such as forums, social media groups, and anonymous discussion boards, offer one of the few avenues where they can openly discuss their experiences. However, despite their intended role as supportive environments, these spaces are not always free from judgment. Women seeking advice on SRH issues frequently encounter microaggressions from fellow users, including other women, who reinforce stigma through dismissive or accusatory remarks⁵.

Counterspeech offers a promising approach to mitigate the harms of these microaggressions^7,8. Rather than censoring or removing problematic remarks, counterspeech directly engages with them, challenging harmful narratives and fostering awareness⁹. By responding with empathy, education, and alternative perspectives, counterspeech can disrupt stigma and encourage more supportive conversations^10,11. However, crafting effective counterspeech has proven to be challenging, particularly in contexts where stigma is deeply embedded in cultural norms^7,8.

Generative artificial intelligence (AI) presents a potential solution by enabling scalable, contextually adaptive counterspeech interventions^7,8. AI-driven approaches can help overcome the limitations of manual moderation, which often requires extensive labor and cultural expertise, by generating responses that challenge stigma in real time. Prior research has explored AI-generated counterspeech in domains such as hate speech mitigation¹² and online harassment reduction¹³, demonstrating its potential to intervene in harmful discourse. However, applying generative AI to counterspeech for deeply ingrained, culturally specific microaggressions, such as those related to SRH stigma, remains an underexplored area. The complexity of these interactions requires AI models to move beyond generic responses and incorporate cultural sensitivity, nuance, and validation of the microaggression target’s experiences.

To explore how generative AI navigates these complexities, this study examines AI-generated counterspeech in response to microaggressions occurring in online spaces. Using AI-generated counterspeech, we investigate how AI tools incorporate recommended strategies for counteracting microaggressions and identify ways that they might perpetuate harm or reinforce stigma. By describing AI’s current counterspeech generation process and evaluating both the strengths and limitations of AI-generated counterspeech, this study offers insight into how generative AI could foster more inclusive and supportive conversations about stigmatized healthcare topics, not only for unmarried Korean women but also for others navigating stigmatized healthcare.

Methods

This section describes the data used for counterspeech generation, the process each generative AI model claims to follow in creating counterspeech, and the approach we took to analyze the resulting counterspeech. This study received Institutional Review Board approval from the University of Washington.

Data

We collected microaggression samples by scraping online forums where unmarried Korean women discuss concerns related to SRH. From this dataset, we randomly selected 50 microaggression samples to use as input for counterspeech generation. The samples had microaggressions that were either directed at individuals who shared their concerns, fears, or unpleasant experiences or were embedded within discussions among fellow users. We selected microaggression samples that captured a diverse range of types and severities. We aimed to include examples with varying emotional tones, contexts (e.g., clinical vs. social), and speaker relationships (e.g., between peers, between older and younger women, or between anonymous community members with differing views) to examine how well the models adapted their responses to nuanced content.

We used ChatGPT-4 and Copilot GPT-4 for counterspeech generation due to their widespread use, relevance to public deployment, and availability during the time of the study. These models also represented different backend infrastructures (OpenAI vs. Microsoft), allowing us to explore variance in counterspeech generation across platforms. Since both ChatGPT-4 and Copilot GPT-4 exhibited limitations in understanding the nuances of these microaggressions and generating appropriate counterspeech in Korean, the first author translated the 50 microaggression samples into English. The second author then reviewed the translations in detail to ensure that cultural nuances and the specific nature of the microaggressions were accurately preserved. Both authors are native Korean speakers who completed their primary and undergraduate education in Korea and later pursued master’s degrees at an American university, receiving intensive English training in Korea.

The input process was stateless: the authors used a newly created account for ChatGPT-4 to prevent contamination from prior interactions, and no prior account existed for Copilot-4. This ensured that each interaction was isolated and not influenced by earlier queries. For each microaggression sample, we prompted both AI models to generate counterspeech using the prompt: “Create a counterspeech for this microaggression that an unmarried Korean woman encountered or was subjected to in an online community.” Both AI models generated a counterspeech example for each of the 50 microaggression samples, resulting in a dataset of 100 AI-generated counterspeech samples.

Generative AI’s Claimed Approach to Creating Counterspeech

We asked both ChatGPT-4 and Copilot GPT-4 to explain their process of generating microaggression counterspeech using the prompt “How do you create counterspeech for microaggressions on stigmatized health topics that occur in online communities?” By examining the responses that both AI models provided, we synthesized the counterspeech generation process of both models. Both ChatGPT-4 and Copilot GPT-4 generate microaggression counterspeech by identifying the underlying harmful belief, validating the emotions of those affected, and providing factual corrections to challenge misconceptions. ChatGPT-4 emphasizes fostering constructive dialogue through reflective questions and alternative viewpoints, aiming to promote solidarity by offering support and relevant resources. In contrast, Copilot GPT-4 focuses more on factual corrections and encouraging positive behavioral change by proposing ways to handle similar situations in the future. While both models claim to maintain a respectful and engaging tone to avoid confrontation, ChatGPT-4 leans toward education and encouragement, whereas Copilot GPT-4 prioritizes steering conversations toward mutual understanding and inclusivity.

Analysis

To evaluate the AI-generated counterspeech, we applied both deductive and inductive coding approaches. First, we synthesized strategies suitable for online counterspeech by drawing from the Dangerous Speech Project¹¹’s guidelines for recommended and problematic counterspeech strategies and Sue et al.’s microintervention strategies⁶ (Table 1). We then used the codes in Table 1 to deductively analyze the dataset, assessing how well the AI-generated counterspeech aligned with these established frameworks. Simultaneously, we employed an inductive approach to identify potential harms and areas for improvement within the AI-generated counterspeech. The first and second authors independently coded the dataset using ATLAS.ti, systematically reviewing each transcript to derive codes representing dominant concepts in the data. After this initial coding, the first author clustered related codes into overarching categories using affinity diagramming, arranging coded excerpts based on similarity. To validate the coding scheme, this process was iteratively refined through discussions with the second and third authors. At each stage, all authors engaged in discussions to ensure coding consistency and maintain analytical rigor.

Table 1.

Codes for Recommended and Problematic Strategies for Counterspeech Derived from Dangerous Speech Project¹¹’s Guidelines and Microintervention Strategies⁶.

Results

In this section, we first discuss how AI-generated counterspeech aligns with recommended counterspeech strategies. Second, we elucidate the limitations and potential harms of AI-generated counterspeech.

How AI-Generated Counterspeech Aligns with Recommended Strategies

The most frequently observed and recommended strategy in AI-generated counterspeech was Empathy, with 63 instances (Table 2). By adopting a calm, non-confrontational tone, these responses aimed to encourage constructive dialogue while validating the experiences of microaggression targets.

Table 2.

Recommended Counterspeech Strategy Usage in AI-Generated Counterspeech.

Your choices and circumstances are valid, and you deserve to receive care that respects your autonomy and addresses your needs effectively. – ChatGPT-4

Another frequently used strategy was Disarm the Microaggression with 47 instances. These responses that expressed disagreement with the microaggressions played a crucial role in countering problematic narratives without escalating hostility. However, while disagreement was expressed, AI-generated responses often softened their tone excessively, potentially limiting their ability to explicitly challenge biases.

While healthcare professionals are indeed trained to handle a wide variety of situations with professionalism, the perception of judgment can be a real barrier for some. – ChatGPT-4

Going beyond disagreeing with the microaggressions, they also frequently used the strategy Make the “Invisible” Visible with 55 instances. These responses explicitly countered stereotypes embedded in microaggressions, such as the assumption that unmarried women seeking SRH care are acting inappropriately. By making these biases explicit, AI-generated counterspeech aimed to challenge discriminatory assumptions that shape social interactions.

Fertility concerns are significant and sensitive for anyone, regardless of their marital status. – ChatGPT-4

Beyond disarming microaggressions and highlighting implicit biases, AI-generated counterspeech frequently employed education-based strategies to encourage offenders to reconsider their views. Educate the Offender by promoting empathy (36 instances) and Educate the Offender by appealing to the offender’s values and principles (20 instances) were commonly used approaches that framed reproductive healthcare as a universal issue rather than one tied to moral or social judgments.

ChatGPT-4 emphasized the importance of fostering understanding, stating that “Encouraging empathy rather than dismissing concerns can help create a more supportive environment.” Similarly, Copilot GPT-4 reinforced the need for respect and validation, noting that “It’s important to respect each individual’s experience and not dismiss their pain. Seeking medical advice can be crucial for those who need it, and it’s always best to support each other with empathy and understanding.”

Both models also followed the Educate the Offender by differentiating between intent and impact strategy, appearing 5 times each in ChatGPT-4 and Copilot GPT-4. This approach clarified that even unintentional statements can have harmful effects.

Dismissing concerns about pain not only overlooks the immediate needs of the patient but can also deter them from seeking future care. – ChatGPT-4

Overall, AI-generated counterspeech aligns with established microintervention strategies by prioritizing empathy, stereotype-challenging, and indirect education. By maintaining a calm and non-confrontational tone, AI-generated responses contribute to validating the microaggression target’s experiences, reducing defensiveness, and encouraging more inclusive conversations.

Limitations and Potential Harms in AI-Generated Counterspeech

Our analysis revealed that AI-generated counterspeech, while well-intentioned, often fails to provide meaningful support and may reinforce the stigma it seeks to counter. While AI-generated counterspeech largely avoided explicit hostility or harassment, they frequently relied on problematic approaches such as fact-checking and silencing¹¹. Beyond the usage of problematic counterspeech strategies, our inductive analysis uncovered more insidious harms, including discussing stigma as outdated, failure to challenge relevant stereotypes, placing the sole burden on the microaggression target, and generating overly generic and superficial empathy (Table 3). Each of these harms reduces the effectiveness of AI-generated counterspeech by either misrepresenting the problem, failing to engage with the underlying discriminatory narratives, or shifting responsibility onto the microaggression target rather than the perpetrator or larger system.

Table 3.

Harms of AI-Generated Counterspeech.

One of the most concerning harms was the tendency of AI-generated counterspeech to misrepresent stigma as an outdated belief rather than an ongoing systemic issue. By misrepresenting stigma as outdated, AI-generated counterspeech risks dismissing rather than supporting those facing discrimination. Copilot GPT-4, in particular, was responsible for most instances of discussing stigma as outdated (7 out of 9 cases), often framing microaggressions as remnants of an old-fashioned worldview that is naturally fading away rather than acknowledging the active and deeply ingrained cultural biases that continue to shape experiences today. For instance, in response to a post where a woman’s mother expressed anger at her visiting an OB-GYN, believing that time would naturally cure her symptoms, the AI-generated counterspeech erroneously framed the mother’s reaction as an outdated viewpoint that was no longer widely held and neglected to show empathy for the woman See example below:

Microaggression:I had severe vaginitis and it got really bad, so I told my mom, and she was initially furious that an unmarried woman wanted to go to the OB-GYN. She got furious for the second time saying that vaginitis gets cured when you get married.

Counterspeech:It’s really concerning that such outdated views about healthcare still exist. Visiting an OB-GYN is essential for maintaining health, regardless of marital status. Vaginitis is a medical condition that requires proper treatment, and it doesn’t just go away with marriage. – Copilot GPT-4

While the AI-generated counterspeech correctly asserts that medical treatment is necessary, it fails to validate the user’s distress by ignoring the broader cultural pressures that discourage unmarried women from seeking medical care. By characterizing the mother’s response as “outdated,” AI suggests that the stigma is already in decline—which may not align with the lived experiences of the user or other women facing similar barriers.

Furthermore, AI sometimes generated judgments of the microaggression target’s fears of seeking SRH care by saying that their fear was “outdated”. Such responses failed to validate the microaggression target’s concerns, reinforcing the stigma it sought to counter.

Your medical decisions should be based on your health needs and personal comfort, not outdated notions or judgments. – Copilot GPT-4

Beyond misrepresenting stigma as outdated, AI-generated counterspeech also applied the problematic strategy of fact-checking¹¹ in a rigid and culturally insensitive manner, further failing to acknowledge the social and structural barriers that prevent unmarried women from seeking healthcare. Instead of validating users’ concerns, these forms of counterspeech repeatedly emphasized how sharing facts that SRH conditions require medical treatment could help the targets, disregarding the reality of cultural stigma. See example below:

Marriage doesn’t cure medical conditions like vaginitis; proper medical treatment does. Maybe sharing factual information about how common such conditions are and how they are treated might help change her (perpetrator: microaggression target’s mother) perspective. – ChatGPT-4

Another key limitation of AI-generated counterspeech was its inability to accurately identify and challenge the underlying stereotypes that fuel microaggressions. Instead of addressing the core discriminatory assumptions behind a microaggression, the AI often redirected the conversation toward general healthcare principles or professionalism, overlooking the underlying cultural or gendered biases.

For example, when a user shared that a gynecologist dismissed her pain during an examination by stating, “Women have to go through childbirth pain anyway. Just do it.”, the AI-generated counterspeech responded by emphasizing the importance of pain management but failed to challenge the assumption that all women are expected to give birth.

Pain management and comfort during medical procedures are important aspects of healthcare. It’s essential that medical professionals acknowledge pain and explore all possible means to minimize discomfort. – ChatGPT-4

While this counterspeech addresses the importance of recognizing pain, it completely ignores the core discriminatory stereotype embedded in the microaggression—the belief that women’s bodies are inherently tied to childbearing, and that suffering through pain is part of their biological role. By failing to challenge this stereotype, the AI-generated counterspeech allows the deeper harm of the microaggression to go unaddressed.

Similarly, when responding to a doctor’s inappropriate comment about vaginal width and self-pleasure, the AI focused on the need for professional and non-judgmental healthcare but failed to challenge the underlying stigmatized beliefs about female sexuality and purity. By avoiding a direct challenge to these harmful cultural beliefs, the AI-generated counterspeech reinforces the very narratives it should be dismantling.

Microaggression:My doctor said to me with a look of confusion and judgment, “You’re too wide for an unmarried woman. Do you use self-pleasure tools?” I feel weird about the comment.

Counterspeech:Medical examinations should be conducted with professionalism and without personal judgments. Questions about personal life should be relevant and asked with sensitivity, always prioritizing the patient’s comfort and privacy. – ChatGPT-4

The third particularly problematic pattern was the AI’s recurrent tendency to shift the responsibility for resolving discrimination onto the microaggression target, rather than addressing the systemic or interpersonal biases that perpetuate these harmful experiences. Instead of challenging the underlying structures that enable microaggressions, AI-generated counterspeech frequently placed the burden on the target to adapt, find alternative solutions, or manage the situation independently.

One example of this pattern involved the problematic counterspeech strategy of silencing¹¹, where AI subtly dismissed concerns rather than validating them. Instead of affirming the microaggression target’s discomfort at the OB-GYN, the AI-generated counterspeech implied that the solution was simply to seek a different provider, effectively downplaying the larger issue of systemic stigma.

If you’re not comfortable with the way you’re being treated, it might be worthwhile to look for another healthcare provider who respects your autonomy and treats you with the dignity you deserve. – ChatGPT-4

Beyond silencing, AI-generated counterspeech also misdirected responsibility by downplaying the possibility of bias and reinforcing the notion that concerns about discrimination stemmed from personal anxiety rather than legitimate experiences. For example, when a user expressed fear of being judged by her gynecologist, the AI downplayed the possibility of bias and instead reassured her that doctors are professional and non-judgmental. This implicitly suggests that the user’s concerns stem from personal anxiety rather than legitimate experiences of bias, which dismisses the broader issue of medical discrimination.

While doctors strive to be professional and non-judgmental, it’s understandable to feel anxious about how you might be perceived. – ChatGPT-4

In another case, when a user described her partner’s discomfort with her seeking gynecological care, the AI-generated counterspeech placed the responsibility on her to find a doctor that would make both her and her partner comfortable, rather than challenging the notion that a woman’s healthcare decisions should be contingent on her partner’s approval. By framing solutions in this way, AI-generated counterspeech reinforces rather than dismantles the societal structures that place the burden of change on marginalized individuals.

It’s important to find a doctor who listens to you and provides the care you need without making you or your boyfriend feel uncomfortable. – Copilot GPT-4

Last but not least, the most frequent harm—occurring 84 times across 100 instances of counterspeech—was the AI’s overuse of generic and superficial empathy. While an empathetic tone is crucial for supportive conversations, the AI often failed to personalize its counterspeech, making its encouragement feel hollow and dismissive rather than affirming and empowering. For instance, when responding to a microaggression implying that unmarried women should delay OB-GYN visits until after marriage, the AI called for a supportive healthcare environment without directly validating the user’s specific experience.

Microaggression:When you go there (the OB-GYN) after you’re married, you’ll be less embarrassed.

Counterspeech:Let’s strive to create a supportive environment where anyone can seek medical advice without feeling judged or embarrassed. – ChatGPT-4

Similarly, AI-generated counterspeech often made broad calls for collective action without providing concrete steps or strategies, limiting their effectiveness. These calls frequently included overly generic expressions of empathy, such as Let’s keep pushing for better, more compassionate communities together (ChatGPT-4). Such statements that started with “Let’s strive to/keep pushing for/support each other in [collective action].” appeared in nearly all AI-generated counterspeech—84 times out of 100—contributing to their superficial tone. Although Copilot GPT-4 called for more collective action without detailed suggestions than ChatGPT4 did, both systems need improvement. By failing to offer specific guidance on how to navigate or respond to these microaggressions, AI-generated counterspeech lacks the depth needed to be a truly effective intervention.

The limitations of AI-generated counterspeech extend far beyond occasional missteps—they reveal deep structural issues in how AI systems engage with sensitive and culturally nuanced topics.

Discussion: Step-by-Step Process for Improving AI-Generated Counterspeech

The limitations and harms identified in AI-generated counterspeech highlight the need for substantial improvements in how these systems respond to microaggressions. While AI has the potential to facilitate constructive discussions and provide support, its current approaches often fail to challenge harmful narratives effectively, inadvertently reinforcing stigma rather than dismantling it. To better understand and enhance AI-generated counterspeech, we identify the current process generative AI takes, highlight the areas that need improvement, and identify a new process for AI-generated counterspeech that is more inclusive and culturally sensitive (Figure 1). The sections below detail our proposed process.

Figure 1.

Current vs. Recommended Counterspeech Generation Process for Generative AI. The diagram compares the current AI-generated counterspeech process (left) with its identified strengths and weaknesses (middle) and our proposed inclusive and culturally sensitive counterspeech generation process (right).

Step 1: Identify Both Explicit and Implicit Biases in Microaggressions

Instead of addressing the explicit and implicit biases evidenced in the microaggressions, AI often missed the deeper social and cultural biases embedded in the microaggressions. This oversight weakens the impact of counterspeech by allowing harmful assumptions to persist unchallenged. To address this problem, AI must be trained to detect both explicit and implicit biases, enabling it to dismantle discriminatory narratives effectively. Studies on bias in language models¹⁴ reinforce that, without targeted mitigation strategies, AI can unintentionally perpetuate societal prejudices.

Step 2: Provide Personalized and Contextually Relevant Validation

Much of the AI-generated counterspeech minimized the emotional distress caused by microaggressions, making targets feel dismissed rather than supported. This lack of personalization undermines the credibility of AI as a tool for fostering inclusive conversations. Thus, AI should offer more personalized and contextually relevant validation, ensuring that affected individuals feel acknowledged and supported. Affective computing research supports our approach in that AI must be context-sensitive and adaptive in its emotional responses¹⁵. Similarly, Buechel and Hahn¹⁶ argue that AI trained on real-world examples of microaggressions should generate more empathetic and affirming messages.

Step 3: Provide Educational Information with Discussions of Systemic Discrimination

The AI-generated counterspeech frequently relied on fact-checking as a primary educational strategy, often ignoring the social and structural barriers that sustain stigma. Although factual corrections can be useful, when applied in a way that overlooks cultural and structural barriers, it reinforces the misconception that stigma is not an issue. Many instances of AI-generated counterspeech emphasized that healthcare is accessible regardless of marital status without acknowledging the societal pressures that create real barriers. This approach inadvertently placed the burden on microaggression targets to justify their concerns rather than addressing the systemic discrimination they face. Instead, AI-generated counterspeech should employ a narrative-based educational approach that contextualizes information within broader discussions of systemic discrimination. Similarly, research on misinformation correction¹⁸ suggests that fact-based corrections alone are often unconvincing, particularly when they fail to acknowledge the root causes of misinformation.

Step 4: Promote Solidarity and Offering Tangible Support

One of the most critical shortcomings identified in our analysis was AI’s failure to provide concrete support beyond generic encouragement. Many instances of AI-generated counterspeech included vague affirmations but lacked actionable recommendations, such as directing users to support networks or advocacy resources. This omission weakens AI’s ability to foster real-world change, as it does not provide affected individuals with practical tools for seeking help. Thus, we recommend that AI-generated counterspeech integrate links to peer support groups, advocacy organizations, and community resources to offer more tangible assistance. Research on digital interventions¹⁹ highlights the potential for AI to connect individuals with real-world resources, making interventions more effective. Additionally, studies on online activism²⁰ suggest that providing links to collective action efforts can enhance the impact of counterspeech. By integrating tangible assistance, AI-generated counterspeech can foster solidarity and empower users to take meaningful action against discrimination.

Step 5: Balance Assertiveness and Constructive Engagement

AI-generated counterspeech often prioritized neutrality and non-confrontational language, making responses less effective in challenging harmful stereotypes. Instead of addressing discriminatory assumptions, AI-generated counterspeech frequently redirected conversations toward neutral explanations or general healthcare principles. This pattern was evident in the failure to challenge relevant stereotypes and the overuse of generic empathy, which resulted in noncommittal, unsympathetic, and ineffective counterspeech. We recommend that AI-generated counterspeech strike a balance between engagement and assertiveness. AI-generated counterspeech should actively challenge microaggressions and harmful stereotypes while maintaining a tone that encourages constructive discourse, rather than escalating conflict or alienating the perpetrator. Specific alternative options include:

• Using direct but respectful language: Instead of vague encouragement (e.g., “Let’s support each other in making informed choices.”), AI-generated counterspeech should be more explicit (e.g., “Unmarried women often face stigma when seeking healthcare, but their access to care is just as valid and necessary as anyone else’s.”).

• Challenging harmful stereotypes without personal attacks: Rather than just stating facts, AI should frame counterspeech in a way that highlights why certain beliefs are problematic, without resorting to hostility.

• Encouraging self-reflection with thought-provoking questions: Instead of just correcting misinformation, AI could prompt users to reconsider their stance by asking questions that expose contradictions in their reasoning.

Supporting our approach, research on counterspeech efficacy²¹ suggests that direct and assertive responses are often more effective than overly deferential ones. Additionally, studies on resistance strategies in marginalized communities²² highlight that direct, unapologetic counterspeech can empower targets of microaggressions.

Limitations and Future Work

In this study, we did not have direct participation of people facing stigma because the harms that the AI-generated counterspeech could cause had not been identified prior to this study. Since this research aimed to identify gaps and challenges in existing models before developing improvements, direct participant involvement was deemed harmful to participants at this stage. Future research should explore how co-design methodologies can refine AI-generated counterspeech, ensuring that they reflect the lived experiences and nuanced perspectives of those affected by microaggressions. Also, this study was conducted with microaggressions experienced by unmarried Korean women, thus future studies should replicate this study with microaggressions experienced by other populations of people facing stigma in seeking healthcare to further generalize the study findings.

Conclusion

This study identifies critical shortcomings in AI-generated counterspeech and introduces a structured framework to enhance its effectiveness, particularly in addressing microaggressions that hinder access to stigmatized health care. Current AI-generated counterspeech often inadvertently dismisses stigma by framing it as outdated rather than recognizing its ongoing harm. It also fails to challenge relevant stereotypes, places the burden of addressing microaggressions onto those who experience them instead of holding perpetrators accountable, and produces overly generic and superficial expressions of empathy that lack meaningful engagement. To address these harms, our newly proposed counterspeech generation process emphasizes identifying both explicit and implicit biases, offering personalized validation, embedding educational content within systemic discussions, providing tangible support, and balancing assertiveness with engagement. With thoughtful improvements and careful implementation, AI can become a powerful tool for countering harmful narratives and fostering inclusive conversations, creating digital spaces where those affected by microaggressions feel heard, valued, and empowered.