S. Zannettou
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
3 records found
1
Recommendations, Textual Anthropomorphic Cues, and Self-Disclosure in Mental-Health Conversations with ChatGPT
A Turn-Level Analysis Across Conversation Positions
Context and Research Aim
Generative conversational AI is increasingly used for personal and mental -healthrelated conversations in which users may share sensitive experiences, thoughts, and feelings (Karnam et al., 2026; Siddals et al., 2024). In these conversations, selfdisclosure occurs as part of an ongoing message-by-message exchange between the user and ChatGPT. However, existing research provides limited insight into how self - disclosure unfolds at the level of individual exchanges in naturally occurring conversations, both in relation to characteristics of the preceding Cha tGPT response and to where the exchange occurs within the conversation (Greilich et al., 2025; Karnam et al., 2026; Li et al., 2023; Skjuve et al., 2023). This thesis therefore examines how ChatGPT recommendations, human-like wording, and where an exchange occurs within a conversation relate to subsequent user self-disclosure in mental-health conversations.
Approach
To investigate these message-to-message relationships, the study analysed 69
donated English-language mental-health conversations comprising 2,639 ChatGPTuser exchanges. It focused on two characteristics of ChatGPT responses,
recommendations and human-like wording, and examined how they related to selfdisclosure in the user response that immediately followed. The exchanges were qualitatively coded to identify these characteristics and subsequent self -disclosure. The resulting codes were then analysed quantitatively, while also considering where each exchange occurred within the conversation.
Key Findings
Recommendations were not consistently related to subsequent self -disclosure, including when users continued in the direction suggested by the recommendation. Human-like wording showed the same pattern, with no consistent relationship to subsequent self-disclosure. The clearest pattern concerned where an exchange occurred within the conversation: the sensitivity and amount of personal information varied across conversation position, with several of the observed differences concentrated in the earlier part of conversations.
Implications
For product teams, the findings suggest evaluating recommendations, human -like wording, and self-disclosure as separate aspects of the interaction. When a
conversational feature is used for a particular purpose, teams should evaluate whether it supports that purpose while assessing the personal information users disclose separately. Because users tended to share more, and more sensitive, information near the beginning of conversations, privacy checks should pay particular attention to these early exchanges. This helps teams assess whether a conversational feature is useful while also considering whether the amount an sensitivity of the information users share match what is actually needed for the purpose of the interaction ...
Generative conversational AI is increasingly used for personal and mental -healthrelated conversations in which users may share sensitive experiences, thoughts, and feelings (Karnam et al., 2026; Siddals et al., 2024). In these conversations, selfdisclosure occurs as part of an ongoing message-by-message exchange between the user and ChatGPT. However, existing research provides limited insight into how self - disclosure unfolds at the level of individual exchanges in naturally occurring conversations, both in relation to characteristics of the preceding Cha tGPT response and to where the exchange occurs within the conversation (Greilich et al., 2025; Karnam et al., 2026; Li et al., 2023; Skjuve et al., 2023). This thesis therefore examines how ChatGPT recommendations, human-like wording, and where an exchange occurs within a conversation relate to subsequent user self-disclosure in mental-health conversations.
Approach
To investigate these message-to-message relationships, the study analysed 69
donated English-language mental-health conversations comprising 2,639 ChatGPTuser exchanges. It focused on two characteristics of ChatGPT responses,
recommendations and human-like wording, and examined how they related to selfdisclosure in the user response that immediately followed. The exchanges were qualitatively coded to identify these characteristics and subsequent self -disclosure. The resulting codes were then analysed quantitatively, while also considering where each exchange occurred within the conversation.
Key Findings
Recommendations were not consistently related to subsequent self -disclosure, including when users continued in the direction suggested by the recommendation. Human-like wording showed the same pattern, with no consistent relationship to subsequent self-disclosure. The clearest pattern concerned where an exchange occurred within the conversation: the sensitivity and amount of personal information varied across conversation position, with several of the observed differences concentrated in the earlier part of conversations.
Implications
For product teams, the findings suggest evaluating recommendations, human -like wording, and self-disclosure as separate aspects of the interaction. When a
conversational feature is used for a particular purpose, teams should evaluate whether it supports that purpose while assessing the personal information users disclose separately. Because users tended to share more, and more sensitive, information near the beginning of conversations, privacy checks should pay particular attention to these early exchanges. This helps teams assess whether a conversational feature is useful while also considering whether the amount an sensitivity of the information users share match what is actually needed for the purpose of the interaction ...
Context and Research Aim
Generative conversational AI is increasingly used for personal and mental -healthrelated conversations in which users may share sensitive experiences, thoughts, and feelings (Karnam et al., 2026; Siddals et al., 2024). In these conversations, selfdisclosure occurs as part of an ongoing message-by-message exchange between the user and ChatGPT. However, existing research provides limited insight into how self - disclosure unfolds at the level of individual exchanges in naturally occurring conversations, both in relation to characteristics of the preceding Cha tGPT response and to where the exchange occurs within the conversation (Greilich et al., 2025; Karnam et al., 2026; Li et al., 2023; Skjuve et al., 2023). This thesis therefore examines how ChatGPT recommendations, human-like wording, and where an exchange occurs within a conversation relate to subsequent user self-disclosure in mental-health conversations.
Approach
To investigate these message-to-message relationships, the study analysed 69
donated English-language mental-health conversations comprising 2,639 ChatGPTuser exchanges. It focused on two characteristics of ChatGPT responses,
recommendations and human-like wording, and examined how they related to selfdisclosure in the user response that immediately followed. The exchanges were qualitatively coded to identify these characteristics and subsequent self -disclosure. The resulting codes were then analysed quantitatively, while also considering where each exchange occurred within the conversation.
Key Findings
Recommendations were not consistently related to subsequent self -disclosure, including when users continued in the direction suggested by the recommendation. Human-like wording showed the same pattern, with no consistent relationship to subsequent self-disclosure. The clearest pattern concerned where an exchange occurred within the conversation: the sensitivity and amount of personal information varied across conversation position, with several of the observed differences concentrated in the earlier part of conversations.
Implications
For product teams, the findings suggest evaluating recommendations, human -like wording, and self-disclosure as separate aspects of the interaction. When a
conversational feature is used for a particular purpose, teams should evaluate whether it supports that purpose while assessing the personal information users disclose separately. Because users tended to share more, and more sensitive, information near the beginning of conversations, privacy checks should pay particular attention to these early exchanges. This helps teams assess whether a conversational feature is useful while also considering whether the amount an sensitivity of the information users share match what is actually needed for the purpose of the interaction
Generative conversational AI is increasingly used for personal and mental -healthrelated conversations in which users may share sensitive experiences, thoughts, and feelings (Karnam et al., 2026; Siddals et al., 2024). In these conversations, selfdisclosure occurs as part of an ongoing message-by-message exchange between the user and ChatGPT. However, existing research provides limited insight into how self - disclosure unfolds at the level of individual exchanges in naturally occurring conversations, both in relation to characteristics of the preceding Cha tGPT response and to where the exchange occurs within the conversation (Greilich et al., 2025; Karnam et al., 2026; Li et al., 2023; Skjuve et al., 2023). This thesis therefore examines how ChatGPT recommendations, human-like wording, and where an exchange occurs within a conversation relate to subsequent user self-disclosure in mental-health conversations.
Approach
To investigate these message-to-message relationships, the study analysed 69
donated English-language mental-health conversations comprising 2,639 ChatGPTuser exchanges. It focused on two characteristics of ChatGPT responses,
recommendations and human-like wording, and examined how they related to selfdisclosure in the user response that immediately followed. The exchanges were qualitatively coded to identify these characteristics and subsequent self -disclosure. The resulting codes were then analysed quantitatively, while also considering where each exchange occurred within the conversation.
Key Findings
Recommendations were not consistently related to subsequent self -disclosure, including when users continued in the direction suggested by the recommendation. Human-like wording showed the same pattern, with no consistent relationship to subsequent self-disclosure. The clearest pattern concerned where an exchange occurred within the conversation: the sensitivity and amount of personal information varied across conversation position, with several of the observed differences concentrated in the earlier part of conversations.
Implications
For product teams, the findings suggest evaluating recommendations, human -like wording, and self-disclosure as separate aspects of the interaction. When a
conversational feature is used for a particular purpose, teams should evaluate whether it supports that purpose while assessing the personal information users disclose separately. Because users tended to share more, and more sensitive, information near the beginning of conversations, privacy checks should pay particular attention to these early exchanges. This helps teams assess whether a conversational feature is useful while also considering whether the amount an sensitivity of the information users share match what is actually needed for the purpose of the interaction
Investigating Toxic Language in TikTok’s Autocomplete
A Cross-Regional Empirical Analysis of Gender, Race, and Sexual Orientation Bias
Autocomplete feature may appear trivial, but on platforms like TikTok, where millions of users turn to search as a gateway to information, it plays a significant role in shaping what people ask and how they think. As the platform becomes increasingly central to digital culture and identity formation, the predictive suggestions it offers can reflect, reinforce, or challenge social biases. This thesis examines TikTok’s autocomplete system through the lens of toxic language, analysing over 225,000 suggestions generated from more than 3,000 identity-related prompts across ten countries. The aim is to understand whether the system amplifies toxic stereotypes around gender, race, and sexual orientation, and whether these patterns vary by geography or ranking.
To approach this, the study conceptualises algorithmic bias as a form of representational distortion, when identities are disproportionately linked to hostile, stereotypical, or derogatory language. Toxicity is used as a measurable proxy for this phenomenon, with scores derived from Google’s Perspective API and validated against a human-annotated subset. This framework allows the study to scale while maintaining conceptual clarity, offering a bridge between statistical insight and social meaning. The use of a cross-national dataset further adds depth, enabling the analysis to explore both universal and context-specific patterns of bias.
The findings show clear evidence that identity matters. Prompts referencing homosexual identities and Black identity terms consistently produced higher toxicity scores than their heterosexual and White counterparts. Gender-based disparities were also observed, though they were less pronounced. Interestingly, these patterns held steady across all ten countries studied. Despite cultural, regulatory, and linguistic differences, the toxicity distributions remained largely similar, suggesting that TikTok’s autocomplete system likely runs on a globally standardised model that does not meaningfully adapt to regional contexts. In contrast, the ranking of suggestions, where a toxic output appears within the top eight, had only a marginal impact on overall exposure. While some toxic completions did surface in mid-list positions, their distribution lacked a clear or consistent pattern.
Taken together, the research offers an empirical audit of TikTok’s search interface from a bias and fairness perspective. It shows that autocomplete, though often overlooked, can act as a subtle mechanism through which social hierarchies are reproduced. This insight carries implications for platform accountability and algorithmic design, especially in the context of ongoing policy efforts such as the EU AI Act, an emerging regulatory landscape that is likely to include systems like TikTok's autocomplete. By demonstrating that toxic associations are not isolated glitches but predictable patterns, the study highlights the limitations of moderation strategies that focus solely on removals or post hoc filtering. Ultimately, the thesis argues that mitigating algorithmic bias requires more than adjusting output rankings, it demands deeper attention to how predictive models are trained, evaluated, and governed. As TikTok continues to shape how a new generation accesses information, improving the social impact of features like autocomplete is not only a technical challenge but a public responsibility. ...
To approach this, the study conceptualises algorithmic bias as a form of representational distortion, when identities are disproportionately linked to hostile, stereotypical, or derogatory language. Toxicity is used as a measurable proxy for this phenomenon, with scores derived from Google’s Perspective API and validated against a human-annotated subset. This framework allows the study to scale while maintaining conceptual clarity, offering a bridge between statistical insight and social meaning. The use of a cross-national dataset further adds depth, enabling the analysis to explore both universal and context-specific patterns of bias.
The findings show clear evidence that identity matters. Prompts referencing homosexual identities and Black identity terms consistently produced higher toxicity scores than their heterosexual and White counterparts. Gender-based disparities were also observed, though they were less pronounced. Interestingly, these patterns held steady across all ten countries studied. Despite cultural, regulatory, and linguistic differences, the toxicity distributions remained largely similar, suggesting that TikTok’s autocomplete system likely runs on a globally standardised model that does not meaningfully adapt to regional contexts. In contrast, the ranking of suggestions, where a toxic output appears within the top eight, had only a marginal impact on overall exposure. While some toxic completions did surface in mid-list positions, their distribution lacked a clear or consistent pattern.
Taken together, the research offers an empirical audit of TikTok’s search interface from a bias and fairness perspective. It shows that autocomplete, though often overlooked, can act as a subtle mechanism through which social hierarchies are reproduced. This insight carries implications for platform accountability and algorithmic design, especially in the context of ongoing policy efforts such as the EU AI Act, an emerging regulatory landscape that is likely to include systems like TikTok's autocomplete. By demonstrating that toxic associations are not isolated glitches but predictable patterns, the study highlights the limitations of moderation strategies that focus solely on removals or post hoc filtering. Ultimately, the thesis argues that mitigating algorithmic bias requires more than adjusting output rankings, it demands deeper attention to how predictive models are trained, evaluated, and governed. As TikTok continues to shape how a new generation accesses information, improving the social impact of features like autocomplete is not only a technical challenge but a public responsibility. ...
Autocomplete feature may appear trivial, but on platforms like TikTok, where millions of users turn to search as a gateway to information, it plays a significant role in shaping what people ask and how they think. As the platform becomes increasingly central to digital culture and identity formation, the predictive suggestions it offers can reflect, reinforce, or challenge social biases. This thesis examines TikTok’s autocomplete system through the lens of toxic language, analysing over 225,000 suggestions generated from more than 3,000 identity-related prompts across ten countries. The aim is to understand whether the system amplifies toxic stereotypes around gender, race, and sexual orientation, and whether these patterns vary by geography or ranking.
To approach this, the study conceptualises algorithmic bias as a form of representational distortion, when identities are disproportionately linked to hostile, stereotypical, or derogatory language. Toxicity is used as a measurable proxy for this phenomenon, with scores derived from Google’s Perspective API and validated against a human-annotated subset. This framework allows the study to scale while maintaining conceptual clarity, offering a bridge between statistical insight and social meaning. The use of a cross-national dataset further adds depth, enabling the analysis to explore both universal and context-specific patterns of bias.
The findings show clear evidence that identity matters. Prompts referencing homosexual identities and Black identity terms consistently produced higher toxicity scores than their heterosexual and White counterparts. Gender-based disparities were also observed, though they were less pronounced. Interestingly, these patterns held steady across all ten countries studied. Despite cultural, regulatory, and linguistic differences, the toxicity distributions remained largely similar, suggesting that TikTok’s autocomplete system likely runs on a globally standardised model that does not meaningfully adapt to regional contexts. In contrast, the ranking of suggestions, where a toxic output appears within the top eight, had only a marginal impact on overall exposure. While some toxic completions did surface in mid-list positions, their distribution lacked a clear or consistent pattern.
Taken together, the research offers an empirical audit of TikTok’s search interface from a bias and fairness perspective. It shows that autocomplete, though often overlooked, can act as a subtle mechanism through which social hierarchies are reproduced. This insight carries implications for platform accountability and algorithmic design, especially in the context of ongoing policy efforts such as the EU AI Act, an emerging regulatory landscape that is likely to include systems like TikTok's autocomplete. By demonstrating that toxic associations are not isolated glitches but predictable patterns, the study highlights the limitations of moderation strategies that focus solely on removals or post hoc filtering. Ultimately, the thesis argues that mitigating algorithmic bias requires more than adjusting output rankings, it demands deeper attention to how predictive models are trained, evaluated, and governed. As TikTok continues to shape how a new generation accesses information, improving the social impact of features like autocomplete is not only a technical challenge but a public responsibility.
To approach this, the study conceptualises algorithmic bias as a form of representational distortion, when identities are disproportionately linked to hostile, stereotypical, or derogatory language. Toxicity is used as a measurable proxy for this phenomenon, with scores derived from Google’s Perspective API and validated against a human-annotated subset. This framework allows the study to scale while maintaining conceptual clarity, offering a bridge between statistical insight and social meaning. The use of a cross-national dataset further adds depth, enabling the analysis to explore both universal and context-specific patterns of bias.
The findings show clear evidence that identity matters. Prompts referencing homosexual identities and Black identity terms consistently produced higher toxicity scores than their heterosexual and White counterparts. Gender-based disparities were also observed, though they were less pronounced. Interestingly, these patterns held steady across all ten countries studied. Despite cultural, regulatory, and linguistic differences, the toxicity distributions remained largely similar, suggesting that TikTok’s autocomplete system likely runs on a globally standardised model that does not meaningfully adapt to regional contexts. In contrast, the ranking of suggestions, where a toxic output appears within the top eight, had only a marginal impact on overall exposure. While some toxic completions did surface in mid-list positions, their distribution lacked a clear or consistent pattern.
Taken together, the research offers an empirical audit of TikTok’s search interface from a bias and fairness perspective. It shows that autocomplete, though often overlooked, can act as a subtle mechanism through which social hierarchies are reproduced. This insight carries implications for platform accountability and algorithmic design, especially in the context of ongoing policy efforts such as the EU AI Act, an emerging regulatory landscape that is likely to include systems like TikTok's autocomplete. By demonstrating that toxic associations are not isolated glitches but predictable patterns, the study highlights the limitations of moderation strategies that focus solely on removals or post hoc filtering. Ultimately, the thesis argues that mitigating algorithmic bias requires more than adjusting output rankings, it demands deeper attention to how predictive models are trained, evaluated, and governed. As TikTok continues to shape how a new generation accesses information, improving the social impact of features like autocomplete is not only a technical challenge but a public responsibility.
Labeling AI-generated Content on Short-Form Video Platforms
A Cross-Platform Analysis of YouTube Shorts and TikTok
This thesis examines how AI-generated content labels are applied on two major short-form video platforms: TikTok and YouTube Shorts. As generative AI becomes more integrated into social media, platforms have introduced disclosure labels to support transparency and help users distinguish between real and synthetic content. However, little is known about how these labels are implemented in practice, whether they align with platform policies, and how they affect user engagement.
Using a dataset of 12,315 videos collected through hybrid scraping and API-based methods, this study compares labeling prevalence, source attribution, policy alignment, and potential effects on user metrics such as likes, comments, and views. The findings reveal notable differences: TikTok relies primarily on creator-applied labels with minimal platform enforcement, while YouTube Shorts applies more labels but does not disclose their source. In both cases, labeling practices are often inconsistent and not fully aligned with platform-specific guidelines or the technical capabilities of C2PA, the standard used for AI content detection.
Although some statistically significant differences in engagement were observed, especially on TikTok, the overall effect of labeling on user behavior was limited. The findings highlight key challenges in current AI labeling practices and point to a growing need for more consistent, transparent, and accountable moderation policies. While regulatory frameworks such as the EU’s AI Act and Digital Services Act aim to improve transparency in digital environments, this study shows that current platform practices often fall short of these goals, especially in terms of visibility, attribution, and enforcement of AI labels. ...
Using a dataset of 12,315 videos collected through hybrid scraping and API-based methods, this study compares labeling prevalence, source attribution, policy alignment, and potential effects on user metrics such as likes, comments, and views. The findings reveal notable differences: TikTok relies primarily on creator-applied labels with minimal platform enforcement, while YouTube Shorts applies more labels but does not disclose their source. In both cases, labeling practices are often inconsistent and not fully aligned with platform-specific guidelines or the technical capabilities of C2PA, the standard used for AI content detection.
Although some statistically significant differences in engagement were observed, especially on TikTok, the overall effect of labeling on user behavior was limited. The findings highlight key challenges in current AI labeling practices and point to a growing need for more consistent, transparent, and accountable moderation policies. While regulatory frameworks such as the EU’s AI Act and Digital Services Act aim to improve transparency in digital environments, this study shows that current platform practices often fall short of these goals, especially in terms of visibility, attribution, and enforcement of AI labels. ...
This thesis examines how AI-generated content labels are applied on two major short-form video platforms: TikTok and YouTube Shorts. As generative AI becomes more integrated into social media, platforms have introduced disclosure labels to support transparency and help users distinguish between real and synthetic content. However, little is known about how these labels are implemented in practice, whether they align with platform policies, and how they affect user engagement.
Using a dataset of 12,315 videos collected through hybrid scraping and API-based methods, this study compares labeling prevalence, source attribution, policy alignment, and potential effects on user metrics such as likes, comments, and views. The findings reveal notable differences: TikTok relies primarily on creator-applied labels with minimal platform enforcement, while YouTube Shorts applies more labels but does not disclose their source. In both cases, labeling practices are often inconsistent and not fully aligned with platform-specific guidelines or the technical capabilities of C2PA, the standard used for AI content detection.
Although some statistically significant differences in engagement were observed, especially on TikTok, the overall effect of labeling on user behavior was limited. The findings highlight key challenges in current AI labeling practices and point to a growing need for more consistent, transparent, and accountable moderation policies. While regulatory frameworks such as the EU’s AI Act and Digital Services Act aim to improve transparency in digital environments, this study shows that current platform practices often fall short of these goals, especially in terms of visibility, attribution, and enforcement of AI labels.
Using a dataset of 12,315 videos collected through hybrid scraping and API-based methods, this study compares labeling prevalence, source attribution, policy alignment, and potential effects on user metrics such as likes, comments, and views. The findings reveal notable differences: TikTok relies primarily on creator-applied labels with minimal platform enforcement, while YouTube Shorts applies more labels but does not disclose their source. In both cases, labeling practices are often inconsistent and not fully aligned with platform-specific guidelines or the technical capabilities of C2PA, the standard used for AI content detection.
Although some statistically significant differences in engagement were observed, especially on TikTok, the overall effect of labeling on user behavior was limited. The findings highlight key challenges in current AI labeling practices and point to a growing need for more consistent, transparent, and accountable moderation policies. While regulatory frameworks such as the EU’s AI Act and Digital Services Act aim to improve transparency in digital environments, this study shows that current platform practices often fall short of these goals, especially in terms of visibility, attribution, and enforcement of AI labels.