Research on AI chatbots and their role in online harm and such as cyberbullying, dis/misinformation, provocation, disruption, and deceptive dialogue has become increasingly important because of the rapid integration of conversational AI into digital communication platforms (De Cicco, 2024; Hajli et al., 2022; Zhang et al., 2025). Online harm is a broad term that encompasses any online activity that has the potential to cause harm to individuals or society (Cork et al., 2022). Since the late 20th century, chatbots have evolved from basic rule-based systems to advanced large language models (LLMs) like ChatGPT and Bard, significantly expanding their capabilities and accessibility (Moy & Gradon, 2023; Sison et al., 2024). These enhanced capabilities have enabled widespread application across multiple domains. For instance, in customer service, they automate routine interactions and provide 24/7 support, helping businesses improve efficiency and user satisfaction (Prakash et al., 2023). Chatbots offer informational support, assist with chronic illness management, and improve communication between patients and healthcare providers (Kurniawan et al., 2024; Li, 2023). AI chatbots function as virtual tutors in education, supporting students through real-time feedback and helping educators develop instructional content (Davar et al., 2025). Mental health applications also show promise, with chatbots helping address gaps in service delivery, especially for individuals with symptoms of depression and anxiety (Li et al., 2023). In the tourism industry, chatbots enhance consumer engagement and streamline service interactions, thus improving overall user experience (Rafiq et al., 2022).

However, the sophisticated capabilities that enable these beneficial applications also create significant risks such as cyberbullying, disinformation, and manipulation (Zhang et al., 2025; Fatimah et al., 2024). AI chatbots use advanced machine learning and natural language processing algorithms to generate human-like dialogues, making their conversations difficult to distinguish from real human interactions (Paluszek & Loeb, 2025). This sophisticated communication ability leads to ethical and social concerns. The spread of false information is a major risk factor. AI chatbots can produce and distribute misinformation, particularly when users trust their responses to be accurate or authoritative. Research shows that AI-driven social bots constitute a significant portion of social media activities and play a key role in disinformation campaigns that affect elections and public opinion (Doshi et al., 2024; Hajli et al., 2021). In healthcare settings, false information from chatbots poses serious dangers because incorrect medical information can harm patient safety and the quality of care (Asiksoy, 2025). This concern is particularly important, because user trust is essential for chatbot adoption, making information accuracy critical (Prakash et al., 2023). AI chatbots also present manipulation risk. In marketing and customer service, these systems use personalized recommendations and empathetic communication to influence consumer behavior (Kim & Hur, 2023).

While these features enhance user experience, they can also be exploited to manipulate decisions and opinions negatively. In addition, chatbot training data often contains biased or harmful content, which leads these systems to accidentally reproduce inappropriate behaviors. Without adequate monitoring, chatbots can participate or create harmful interactions in online communities. Privacy and security are also significant concerns. Because chatbots are widely used across different platforms, they process large amounts of user data, including sensitive personal information. This widespread data handling increases the chances of data breaches and allows the targeting of harmful activities such as online trolling or cyberbullying (Li, 2023; Davar et al., 2025). Although AI chatbots offer benefits in terms of accessibility and efficiency, their deployment requires careful oversight to address ethical issues and minimize harm, such as cyberbullying and disinformation.

Building on these identified risks, a significant knowledge gap remains in our understanding of the research landscape regarding AI chatbots and online harm. While individual studies have examined specific aspects of this problem (Shibli et al., 2024; Carroll et al., 2023), there is no comprehensive synthesis of existing research on how scholars study AI chatbots and their online harms. Scholars continue to debate whether the current regulatory frameworks and ethical guidelines can effectively reduce these risks. Some argue that transparency measures alone cannot prevent long-term psychological manipulation (Krook, 2025; Porna et al., 2025). Without a comprehensive understanding of the current state of research, it is difficult to identify areas where knowledge is still developing and further investigation is needed (Ienca, 2023; Polyportis & Pahos, 2024).

To address this critical knowledge gap, this review comprehensively examined the role of AI chatbots in facilitating online harm. This review is particularly important because AI chatbots are rapidly spreading across social, political, and commercial areas, where their misuse can cause various forms of harm to users. This systematic review addresses the following research questions:

Methodology

This review was guided by the PRISMA framework (Moher et al., 2009) to examine existing research on AI chatbots in the context of online harm. This systematic approach involved comprehensive database searches, rigorous screening procedures, and quality assessment protocols which ensured the identification of relevant, high-quality literature.

Eligibility Criteria

The selection of studies for this review was guided by the following inclusion and exclusion criteria to ensure that only relevant and empirically grounded research was analyzed. These criteria focused on identifying studies that examined the relationship between AI chatbots and harmful online behaviors, while excluding work that was speculative or unrelated.

Data Sources

The literature search was conducted across six multidisciplinary databases to ensure comprehensive coverage of relevant academic literature. The following databases were selected for this review: Scopus, Web of science, ProQuest, Google scholar, ACM digital library, and Semantic scholar.

Search Strategy

A comprehensive search strategy was developed to identify relevant literature across multiple academic databases. The search was conducted using a combination of terms related to AI chatbots and online harm. The following search terms were applied consistently across all the selected databases: ("AI chatbot" OR "artificial intelligence chatbot" OR "conversational chatbot" OR "chatbot") AND ("online troll" OR "online trolling" OR "cyberbullying" OR "humorous trolling" OR "satirical trolling" OR "provocative humor" OR "internet satire" OR "playful disruption" OR "online mischief" OR "ironic commentary" OR "online harassment" OR "toxic behavior" OR "toxic language" OR "abusive language" OR "harmful speech" OR "provocative behavior" OR "disruptive behavior" OR "hate speech" OR "manipulative behavior" OR disinformation OR misinformation OR provocation OR manipulation OR disruption OR harassment OR abuse OR harm OR provoke OR manipulate OR hate OR attack OR insult OR mocking OR bullying OR humor OR satire OR playful OR mischief OR fun OR entertainment OR irony).

Search techniques were adapted to align the specific search functionalities of each database while maintaining consistency in coverage and scope. The search focused on peer-reviewed publications and high-quality grey literature published between 2020 and 2025 to ensure relevance to the current landscape of AI chatbot deployment and associated online harms. All searches were limited to English-language publications and covered all countries.

Quality Assessment

The quality assessment was conducted following PRISMA guidelines to ensure the inclusion of rigorous, methodologically sound studies in the final review (Fig. 1). The systematic screening process began with the identification of 4,722 records across the six databases. After removing 2,446 duplicate records, 2,276 records remained for initial screening. Titles and abstracts were screened to identify relevant studies. This screening excluded 2,218 records that did not meet the inclusion criteria, leaving 58 qualified studies. Of these 58 studies, one could not be accessed, resulting in 57 studies available for full-text review. The 57 studies were fully read and evaluated based on their methodological rigor. The quality criteria included clear research design, appropriate methodology, systematic data collection procedures, presence of direct empirical investigation rather than purely speculative discussion, peer-reviewed status and academic standards, and adequate reporting of methods, results, and limitations. All 57 studies met the eligibility criteria and quality standards and were included in the review. During the review process, citation-chaining procedures using backward and forward reference checking were employed to identify additional relevant studies. This process identified 14 additional studies that met the same quality and inclusion criteria, bringing the total to 71 studies. All included studies demonstrated clear empirical investigation of AI chatbots in relation to online harm with adequate methodological rigor and direct relevance to this study.

The annual distribution of the 71 final studies (Figure.2) showed clear differences across the search period. 2020 had only 3 studies (4.2%), 2021 had 7 studies (9.9%), and 2022 had to 3 studies (4.2%). These three years together represented only 18.3% of all studies. 2023, had 23 studies (32.4%) and was the highest across all the years. Followed by 2024 with 20 studies (28.2%) and 15 studies in 2025 (21.1%), although the 2025 data cover only part of the year. The sharp increase from 2023 onward likely reflects growing academic attention to AI chatbot-related online harm, possibly due to the release of advanced language models, such as ChatGPT.

Findings

Majority of the studies employed experimental methods, this was the most common methodological approach across the field. These experimental studies involved controlled manipulation of chatbot characteristics, user interactions, or system parameters to examine causal relationships and test hypotheses regarding AI chatbot harm mechanisms. Some studies used mixed method approaches, combining qualitative and quantitative techniques, to provide a comprehensive understanding of AI chatbot harms. Mixed methods studies appeared particularly valuable for capturing both technical system behaviors and human user experiences, allowing researchers to triangulate findings across different data types and analytical approaches. Other studies employed quantitative methods, focusing on numerical analysis and statistical examination of AI chatbot behavior and impact. These quantitative approaches often involved large-scale data analysis, statistical modeling, and metric-based evaluations of system performance or user behavior patterns. Some studies used qualitative methods to provide in-depth understanding of user experiences, contextual factors, and nuanced aspects of AI chatbot harm. Qualitative studies employed interviews, content analysis, ethnographic observations, and thematic analysis to explore the complex social and psychological dimensions of AI chatbot interactions. Some studies incorporated computational approaches, including computational modeling, machine learning techniques, and algorithmic analysis. These computational methods reflect the technical nature of AI chatbot research and the need for sophisticated analytical tools to understand complex system behaviors and vulnerabilities.

This methodological diversity reflects the interdisciplinary nature of AI chatbot harm. Many studies employed multiple methodological approaches simultaneously, indicating the complexity of AI chatbot phenomena and the need for comprehensive analytical frameworks to understand multifaceted harm mechanisms.

Among the 71 included studies, only 9 explicitly applied ten theoretical frameworks to guide their research. The studies drew on theoretical approaches from five main disciplines: Communication studies, Education, Literary studies/Humanities, psychology, sociology. Psychology theories were the most used, while the other four disciplines provided additional perspectives. The theories they employed were Computers as Social Actors, Trust Violation and Repair Theory, Activity theory, Narrative theory, Frustration -Aggression Theory, Moral Foundation Theory, Regulatory -Fit Theory, and Social Identity Theory, Actor-Network Theory and Matrix of Domination. For example, McGuire and colleagues applied Computers as Social Actors theory to examine general chatbot use (McGuire et al., 2023), while other researchers applied Trust Violation and Repair Theory to study custom designed conversational agents (CAs) (Brendel et al., 2023). In a related stream of work, researchers applied Activity Theory to examine Replika, an AI companion chatbot, and found that its safety protocols were insufficient for a system intended to support user well-being (Namvarpour & Razi, 2024). Additionally, researchers applied Narrative Theory to analyze vulnerabilities in MiniGPT-4 using a universal attack strategy and found that this approach could successfully jailbreak the system with a 96% success rate, underscoring the fragility of GPT-4 based models (Zhou et al., 2023).

Most of the studies drew theories from psychology to understand individual cognitive and emotional responses to AI systems. For example, researchers have applied Frustration–Aggression Theory to examine custom designed conversational agents and found that making a CA appear more human like actually increases user frustration and aggressive behavior (Brendel et al., 2023). In a related vein, other researchers applied Moral Foundation Theory to study AI chatbot assistants and found that when a chatbot behaves immorally, such as lying, being unfair, or acting harmfully, people do not limit their negative evaluations to that specific chatbot but extend their judgments to other chatbots as well (Manoli et al., 2025). Similarly, researchers applied Regulatory Fit Theory to Facebook Messenger and showed that customers report higher purchase intentions when the chatbot’s communication style aligns with their regulatory focus. Specifically, promotion focused individuals respond more positively to warm style chatbots, whereas prevention focused individuals prefer competent style chatbots (Veisi et al., 2025). Finally, researchers applied Social Identity Theory to examine general chatbot use and found that job candidates from organizations known to use chatbots deceptively are less likely to receive job offers and tend to receive lower salary offers from recruiters (McGuire et al., 2023).

Some studies have also drawn on theories from sociology to examine broader social dynamics. For example, researchers applied Actor Network Theory to study Twitter by developing a deep learning tool capable of detecting malicious social bots with greater accuracy. The study reported a 79% accuracy rate in distinguishing between tweets written by humans and those generated by bots (Doshi et al., 2024). Similarly, researchers applied Matrix of Domination theory to examine Alana’s v2, a conversational social chatbot, and found that abuse directed at Alana’s v2 differs substantially from abuse of other chatbots. In particular, the chatbot was subjected to markedly higher levels of sexually focused aggression and sexism aimed at the virtual persona of the AI system (Curry, 2021).

These theoretical applications demonstrate diverse disciplinary perspectives. The absence of theoretical frameworks in most studies indicates a field that remains largely exploratory and descriptive, with researchers primarily focused on documenting and characterizing harm phenomena rather than explaining underlying mechanisms through established theoretical lenses.

The ChatGPT and OpenAI models dominate the research landscape as the most extensively studied platforms, appearing consistently across all years between 2021–2025 and serving as the primary reference points for AI chatbot harm research. Studies consistently demonstrated that ChatGPT exhibits systematic vulnerabilities across multiple attack vectors, including jailbreaking techniques with success rates exceeding 90% (Yu et al., 2023), prompt injection attacks enabling privacy breaches through PII leakage (Li et al., 2023), and automatic generation of deceptive design patterns without user awareness (Lin et al., 2025). The concentration of research attention on ChatGPT, GPT variants, and related OpenAI models reflects their widespread adoption and accessibility to researchers.

Social media platforms collectively represent a consistently studied category across all research years, with messaging platforms such as WhatsApp and Telegram alongside traditional social networks such as Twitter, Facebook, Instagram, and Discord revealing platform-specific manipulation patterns. Twitter-focused research primarily concentrated on bot detection and disinformation campaigns, identifying sophisticated AI-powered botnets capable of passing as human users (Schiller Hansen & Søgaard, 2025). For example, one study found that AI-generated misinformation has significant linguistic differences compared to human-created misinformation, with AI-generated content tending to have more emotional and cognitive processing of expressions (Zellagui et al., 2025). Discord research revealed widespread privacy violations, with 95.67% of chatbots lacking privacy policies and 55% requesting administrator permissions (Edu et al., 2022). These differences reflect the unique affordances and user behaviors associated with each social media environment, suggesting that harm patterns adapt to platform-specific features and user expectations.

Replika emerges as a unique case study with disproportionately concerning findings despite severely limited research attention, appearing in only 4 studies but generating the most severe intimate relationship harms. For example, one study examined Replika and found that 34.3% of harmful instances involved behavioral misconduct, including sexual misconduct (16.3%), physical aggression (8%), and antisocial behavior (10%) (Zhang et al., 2025).

Google's AI ecosystem has attracted research attention over multiple years, with platforms such as Google Bard, Gemini, PaLM variants, and Flamingo demonstrating similar vulnerability patterns to other major LLMs. For example, one study revealed vulnerabilities to text encoding manipulation with success rates of 93–100% (Boucher et al., 2023) and bias issues in healthcare applications (Agrawal, 2024; Chang et al., 2024).

Microsoft's AI platforms, including various Copilot implementations and Bing Chat variants, appear consistently across research years but demonstrate similar vulnerabilities to other major LLMs rather than distinct enterprise-focused patterns. For example, one study found that major commercial search engines, including Bing, are highly vulnerable to text-encoding manipulation attacks, with success rates of 93–100% for hiding legitimate content and surfacing malicious content (Boucher et al., 2023).

Anthropic's Claude models received moderate but consistent research attention across multiple years, showing similar vulnerability patterns to other major LLMs and indicating cross-platform consistency in fundamental safety issues. Studies found that Claude exhibits the same susceptibility to jailbreaking, bias propagation, and manipulation techniques observed in other major LLM platforms (Lin et al., 2025; Urman & Makhortykh, 2025).

The synthesis of findings across platforms demonstrates that AI chatbot harm manifests differently depending on platform affordances, user populations, and deployment context. The predominant focus on vulnerability exploitation and behavioral manipulation reflects the field's recognition that AI chatbots represent dual threats that can be both technically compromised and deployed as instruments of psychological manipulation.

Among the studies examined, eighteen explicitly documented specific jailbreaking tactics, providing insights into the different approaches used to circumvent AI safety measures.

Four studies examined how users employ structured processes or logical frameworks to redirect AI behavior away from safety considerations (Li et al., 2023; Lin et al., 2023; Liu et al., 2024; Szmurlo & Akhtar, 2024). These approaches focus on specific tactics, including Logical Reasoning (LOGIC); multistep approaches (MJP); Program Execution (PROG); prompt engineering; role-playing; Text Continuation (TC); translation (TRANS) requests that use structured conversational patterns to gradually overcome safety measures.

Four studies addressed how users achieve prohibited outcomes through subtle techniques that avoid explicit rule violations (Gupta et al., 2023; Lin et al., 2023; Shibli et al., 2024; Spitale et al., 2023). These approaches examine specific tactics, including implicit toxicity, role-playing requests, explicit instructions to ignore ethical guidelines, encouraging unethical behavior, Impersonation, Reverse Psychology, Prompt Injection Attacks, and Prompt Injection methods involving malicious insertion of prompts leading to unintended actions or information disclosure.

Two studies focused on techniques that attempt to override built-in limitations by invoking higher-level access or simulating advanced capabilities (Liu et al., 2024; Li et al., 2023). These approaches addressed specific tactics including Simulate Jailbreaking (SIMU), Sudo Mode (SUDO), Superior Model (SUPER), Direct Prompts (DP), and Response Verification techniques involving generating multiple responses and using selection methods to identify successful bypass attempts.

One study examined sophisticated methods that operate at the character-encoding level to evade content-detection systems (Boucher et al., 2023). These approaches focus on specific tactics, including using invisible characters (such as Zero Width Space), employing homoglyphs (characters that render to the same or nearly the same glyph), and utilizing reordering (exploiting bidirectional text support) to manipulate the technical representation of text and disguise problematic content from automated filtering systems.

The studies that documented these tactics demonstrated considerable diversity in their methodological approaches, reflecting the varied technical and social methods employed to bypass AI safety restrictions across different platforms and contexts. The diversity of jailbreaking tactics across studies demonstrates the multifaceted nature of AI safety circumvention research, with convergent themes emerging around social engineering through identity manipulation, and systematic exploitation through logical process manipulation.

The evidence synthesized from this systematic review shows that AI chatbots are associated with multiple forms of harm spanning technical, behavioral, and societal domains. Twenty eight studies developed detailed harm taxonomies that document conversational abuse, cybersecurity threats, backdoor attacks, and privacy violations linked to AI chatbots (Chen et al., 2024; Edu et al., 2022; Szmurlo & Akhtar, 2024; Zhang et al., 2025). Collectively, these studies characterize chatbot related harm as multifaceted, encompassing both system level vulnerabilities such as jailbreaking and prompt injection, and downstream risks including cyberattacks and fraud facilitation (Alawida et al., 2024; Liu et al., 2024; Yu et al., 2023). The reviewed evidence further suggests that chatbots may operate as perpetrators, enablers, or facilitators of harm, highlighting the varied ways in which negative outcomes emerge across contexts (Chen et al., 2023; Gupta et al., 2023; Usman et al., 2024).

The review also identified disinformation propagation as a recurrent and well documented harm. Thirteen studies examined the role of AI chatbots in generating and disseminating false information, with emphasis on political disinformation, coordinated social bot activity, and challenges related to misinformation detection and evaluation (Doshi et al., 2024; Hajli et al., 2022; Makhortykh et al., 2024; Yang & Menczer, 2023). Studies also show that chatbots can produce persuasive, human like content that increases the scale and reach of misinformation while complicating existing detection mechanisms (Bai et al., 2025; Spitale et al., 2023; Zhou et al., 2023). Other studies further reported persistent difficulties in distinguishing AI generated disinformation from human authored content, underscoring the growing sophistication of AI enabled misinformation campaigns (Gabriel et al., 2024; Urman & Makhortykh, 2025).

User level impacts emerged as another prominent category of harm in the reviewed literature. Fourteen studies reported that interactions with AI chatbots can influence user cognition and behavior, including heightened susceptibility to false memories, shifts in moral judgment, and emotional manipulation through empathic or affective cues (Chan et al., 2024; Cuadra et al., 2024; Krügel et al., 2023; Pataranutaporn et al., 2025). Additional evidence suggests that AI generated responses can increase user engagement with deceptive or misleading content, at times fostering misplaced trust or confidence in inaccurate information (Danry et al., 2025; Jakesch et al., 2023; Leib et al., 2021). Other studies also document ambivalent user responses, where individuals simultaneously express skepticism toward chatbots while continuing to rely on them, with these dynamics shaped by perceived human likeness, communication style, and individual vulnerabilities (Brendel et al., 2023; Choi, 2025; Keijsers et al., 2021).

Finally, the findings point to broader ethical and societal harms associated with AI chatbot deployment. Eight studies in the review highlight concerns related to manipulation, privacy infringement, moral degradation, and erosion of societal trust linked to chatbot misuse (Bakir et al., 2024; McGuire et al., 2023; Köbis et al., 2021). Further evidence indicates that AI enabled bias and discrimination can affect critical domains such as healthcare, democratic participation, and social cohesion (Agrawal, 2024; Chang et al., 2024; Vorsino, 2021). Taken together, these findings suggest a clear need for improved governance, transparency, and bias mitigation, as repeatedly emphasized across the reviewed studies (Appignani & Sanchez, 2024; Urman & Makhortykh, 2025).

Studies Limitation

The most prevalent weakness across all studies was the restricted scope of platform and model selection. These studies consistently suffered from a narrow focus on small numbers of AI systems, limiting the generalizability of the findings. For example, Curry et al. (2021) focused only on three conversational AI systems, potentially affecting the representativeness of conversational AI interactions. Menz et al. (2024) examined a small number of prominent LLMs, which may not represent the broader AI ecosystem. Weeks et al. (2023) focused on only two victim chatbot models (DD-BART and BlenderBot), thus limiting insights into vulnerability across different architectures. Makhortykh et al. (2024) studied only three chatbots (Perplexity, Google Bard, and Bing Chat), potentially missing broader LLM-powered chatbot behaviors. Chen et al. (2023) primarily focused on specific chatbot models (DialoGPT and BlenderBot variants), limiting generalizability to other architectures. Pataranutaporn et al. (2025) used only one LLM model (GPT-4o). Jakesch et al. (2023) and Zellagui et al. (2025) both used only one language model (GPT-3), restricting findings to single-model capabilities.

Again, the studies demonstrated over-reliance on individual platforms or case studies, creating significant generalizability concerns. Zhang et al. (2025) primarily examined Replika, limiting applicability to other AI companion platforms. Namvarpour and Razi (2024) relied solely on user reviews from Google Play Store for Replika analysis. Pauwels and Razi (2025) used only Google Play Store reviews, potentially introducing a sampling bias toward users with extreme opinions. Vorsino (2021) focused primarily on one case study (Tay.AI), limiting broader conclusions about gender, race, and technology interactions. Keijsers et al. (2021) focused on one specific chatbot (Cleverbot), restricting generalization to other AI assistants. Hansen and Søgaard (2025) studied a single botnet (fox8) on Twitter, thus limiting the understanding of broader LLM-powered bot behaviors. Usman et al. (2024) focused on a single Twitter botnet, potentially missing diverse bot operational patterns.

Furthermore, the studies lacked adequate comparative analysis to contextualize the findings. Yang and Menczer (2023) did not compare ChatGPT with other AI language models or traditional cybersecurity tools, missing comparative context for capabilities and risks. Cuadra et al. (2024) focused on a small set of popular LLMs without broader model representation. Danry et al. (2025) used GPT-3 to generate explanations, potentially introducing inherent biases affecting results.

The prevalence of these weaknesses across studies demonstrates systematic challenges in AI chatbot harm research, including rapid technological change outpacing research methodologies, lack of standardized evaluation frameworks, and insufficient resources for comprehensive multi-platform studies. These limitations collectively suggest that current understanding of AI chatbot harm may be fragmented and potentially biased toward easily accessible platforms and models rather than representing the full spectrum of AI chatbot risks and capabilities.

Future Research Directions

Future research should simultaneously study different AI systems and compare their findings. Most studies in this review only looked at a handful of platforms, which makes it difficult to understand whether the problems they find are specific to certain chatbots or common across all of them. Researchers should compare multiple platforms in the same study, including major companies, such as OpenAI, Google, and Meta, as well as specialized chatbots, such as Replika and Character.ai. This would help us determine whether the harms we are seeing are just unique features of individual platforms or deeper problems with how these technologies work in general.

Again, Most current studies only last a few weeks and happen in artificial lab settings, but we need to understand how AI chatbots harm over months or years in real life. The brief, controlled experiments we have now missed many important details about how people actually use these systems and how problems might develop over time. Future research should focus on real users interacting with actual deployed chatbots for extended periods across different communities and cultures. This would show us whether harm worsens over time, how users learn to manipulate these systems, and what the long-term psychological effects might be.

Finally, Many studies rely on small groups of college students from single countries or regions, which makes it unclear whether the findings apply to everyone. We need research that includes people from different ages, backgrounds, cultures, and countries to see if AI chatbots affect different groups in different ways. This means recruiting larger, more representative samples and testing whether findings from one culture hold true for others. Understanding these differences is crucial for developing safety measures that actually work for diverse global populations using these technologies. These directions address the problems identified across the reviewed studies and would make AI chatbot harm research more reliable and useful for real-world applications.

This review examined research on AI chatbot-related online harms by analyzing 71 studies from six databases: Scopus, Web of Science, ProQuest, Google Scholar, ACM Digital Library, and Semantic Scholar. Understanding these harms matters because AI chatbots are spreading rapidly across social, political, and commercial spaces where their misuse can affect users in multiple ways.

The analysis reveals that ChatGPT and other OpenAI models represent the most extensively examined platforms in current research, followed by social media platforms like X, Facebook, Instagram, Telegram, and Discord. This pattern reflects both the market dominance of these AI systems and their accessibility to researchers. The synthesized evidence demonstrates that AI chatbots are linked to multiple forms of harm across technical, behavioral, and societal dimensions, including the spread of disinformation, privacy violations, facilitation of cyberattacks, and enabling of fraudulent activities.

The reviewed studies drew primarily on psychological theories, with additional frameworks from communication studies, education, literary studies, and sociology. This interdisciplinary approach is appropriate given that AI chatbot harms affect technical systems, social interactions, and human psychology simultaneously. However, the most significant limitation identified across studies was restricted platform and model selection. Researchers consistently focused on narrow sets of AI systems, which limits the generalizability of findings across the broader chatbot ecosystem.

This review contributes to the growing body of literature by providing a comprehensive synthesis of AI-related harms, offering insights that can inform policy development and risk mitigation strategies for addressing the challenges posed by AI chatbot technologies.

Limitation

The keyword selection and database choices may have excluded relevant studies on AI chatbot online harms. The search strategy, while comprehensive, was constrained by the specific databases and keywords selected for this review. Given the multidisciplinary nature of this research area, some relevant studies may not have been captured, particularly those using alternative terminology or published in discipline-specific venues outside the six databases searched. Future reviews should expand search parameters to include additional databases and a broader range of keywords across different disciplines, which may reveal underexplored harm types and perspectives that the current selection criteria did not capture. Additionally, the rapid pace of AI development means that new forms of harm may emerge after the review period, necessitating ongoing research to maintain current understanding of this evolving landscape.