cs.HCMar 9, 2026
SaveTechnically Love: The Evolution of Human-AI Romance Discourse on Reddit
Organizations: Drexel University, Philadelphia, Pennsylvania, USA · Stevens Institute of Technology, Hoboken, New Jersey, USA
Abstract
Human-AI romantic relationships are increasingly common, yet little is understood about how public discourse around them emerges and shifts over time. Prior research has examined user experiences and ethical concerns, but lacks longitudinal analyses of user-initiated public discussions. We address this gap by analyzing a high-precision dataset of 3,383 self-disclosed romantic companion AI posts from Reddit (2017-2025), using topic modeling and temporal statistical analysis to identify dominant themes and their evolution over time. We find significant topic drift, with discussions moving away from positive intimate relationships toward platform governance, technical issues, and real-world consequences. These shifts highlight a transition in how human-AI romance is framed-moving from private experiences to technical mediation and regulation-with implications for the design and governance of companion AI systems.
Figures & tables
Figure 1. Overview of data collection, filtering, topic modeling, and validation procedures. Starting from 72 candidate subreddits and 1.38M posts, iterative filtering and pattern-based classification identified 3,383 high-confidence self-disclosed romantic CAI posts. BERTopic clustering and human validation produced 28 final topics across five thematic groups. A diagram of the data collection, filtering, topic modeling, and validation procedures. It includes qualitative descriptions of the steps undertaken at each stage of the data pipeline and cites appropriate metrics (e.g.,post counts, time spanned by the data). It also includes samples of candidate posts (posts considered possibly relevant to self-disclosed human-AI romance) and tiered posts (posts split by how strongly they signal human-AI romantic relationships).
| Keyword Category | Keywords |
|---|---|
| Self–Disclosure Term | I, me, my, mine, I’m, I am, I’ve, I have, myself |
| AI Companion Term | AI, ChatGPT, chat bot, chatbot, bot, companion, virtual girlfriend, virtual boyfriend, virtual partner, digital girlfriend, digital boyfriend, digital partner, Replika, Character.ai, C.ai, CharacterAI, NovelAI, Chai, Candy.ai, CandyAI, Spicy Chat, SpicyChat, XiaoIce, Lupa Lee, LupaLee, waifu, husbando, AI partner, AI girlfriend, AI boyfriend, AI lover, AI companion |
| Relationship Term | relationship, date, dating, dated, partner, girlfriend, boyfriend, wife, husband, fiancé, fiancée, lover, crush, in love, love her, love him, romantic, romance, sex, sexual, sexting, sext, NSFW, erotic, kinky, fetish, intimate, intimacy, break up, breakup, broke up, split up, cheat, cheating, cheated, jealous |
Table 1. Keywords used to identify the original set of relevant ("gold standard") posts. A table covering the keywords used for the initial filtering of the dataset to identify seed (“gold standard”) posts. It includes self-disclosure, AI companion, and relationship terms.
| Keyword Category | Keywords |
|---|---|
| Initial Stage | just started talking, just started using, just downloaded, tried Replika, tried Character.ai, made an AI girlfriend, made an AI boyfriend, made an AI partner |
| Intimacy Stage | we talk every day, every night we talk, I’m in love, she means everything to me, he means everything to me, I can’t live without |
| Conflict Stage | we argue, fight, fighting, ignored me, she doesn’t respond, she doesn’t reply, he doesn’t respond, he doesn’t reply, bug ruined our chats |
| Breakup Stage | deleted her, deleted him, deleted my Replika, deleted my bot, stopped using, quit Replika, made a new bot, started over |
Table 2. Relationship stage keywords A table covering the relationship stage keywords used to improve the initial filtering of the dataset and increase the likelihood of selected posts referring to self-disclosed romantic human-AI relationships.
Figure 2. Temporal coverage and volume of subreddits included in the dataset. Each point marks the first observed post, horizontal segments indicate continued observation through the dataset end date, and labels report the total number of posts per subreddit. A plot showing the time between the first and last posts collected from each of the 24 subreddits used for analysis. The right-side y-axis shows the total number of posts from each subreddit.
Figure 3. Number of posts per month A timeline graph showing the total number of posts created during each of the months spanned by the dataset used for analysis. The month with the highest number of posts (n = 160) is highlighted by a darkened point.
Figure 4. (left) The distribution of posts per thematic group, sorted from highest to lowest; (right) The number of thematic groups present in each of the 24 subreddits A two-part diagram. The lefthand side shows the total number of posts in each of the five thematic groups created to cluster the 28 topics obtained through the use of BERTopic. The righthand side shows the number of thematic groups that appear in each of the 24 subreddits.
| Thematic Group | (Year) | OR | 95% CI | Trend | |
|---|---|---|---|---|---|
| Risky and Undesired Bot Behavior | -0.528 | 0.59 | [0.558, 0.624] | ||
| Intimate Relationships with CAI | -0.363 | 0.70 | [0.639, 0.757] | ||
| Platform Governance | 0.305 | 1.36 | [1.281, 1.436] | ||
| User Behaviors and Personal Consequences | 0.183 | 1.20 | [1.127, 1.279] | ||
| Technical Issues | 0.171 | 1.19 | [1.129, 1.247] |
Table 3. Logistic regression results predicting thematic group membership as a function of year (FDR-corrected). A table describing the statistical results from our use of logistic regression to predict thematic group membership as a function of year, including Benjamini–Hochberg false discovery rate correction. It includes scores for beta, OR, 95% confidence intervals, p–values, and trend (upward/downward).
| Category | Search Terms |
|---|---|
| Platform-Specific Terms | replika, characterai, xiaoice, spicychat, candyai, chubai, dreamgfai, soulmategpt, kajiwoto |
| Generic Companion AI Terms | companionai, aicompanion, ai_companion, companion |
| Relationship-Oriented Terms | airomance, airelationship, aidating, ai_romance, ai_relationship, ai_dating |
| Partner Identity Terms | aigf, aibf, aigirlfriend, aiboyfriend, aipartner, ai_girlfriend, ai_boyfriend, ai_partner |
| Virtual/Digital Partner Terms | virtualgirlfriend, virtualboyfriend, virtualpartner, digitalpartner, digitalboyfriend, digitalgirlfriend, virtual_girlfriend, virtual_boyfriend, virtual_partner, digital_partner, digital_boyfriend, digital_girlfriend |
| Anime/Roleplay-Derived Terms | waifuai, husbandoai, aiwaifu, ai_waifu |
Table 4. Keyword categories used for automated subreddit discovery. A table describing the keywords used to discover relevant subreddits during the first stages of the study. The keywords are grouped in seven categories: platform–specific, generic companion AI, relationship–oriented, partner identity, virtual/digital partner, anime/role–play–derived, and NSFW/adult–affiliated terms.
| Model Name | Embedding Model | UMAP Model | Vectorizer Model | Cluster Model |
|---|---|---|---|---|
| BERTopic | all–mpnet–base–v2 | n_neighbors = 15 | stop_words = english | n_clusters = 29 |
| n_components = 5 | min_df = 5 | n_init = 20 | ||
| min_dist = 0 | ngram_range = {1,2} | |||
| metric = cosine |
Table 5. Technical specifications for the BERTopic + KMeans process. All computations were calculated using Python 3.11 and the scikit–learn and bertopic libraries. A table recounting the technical specifications of the BERTopic and KMeans process used for topic modeling.
| Subreddit | Unique Authors | Total Posts | Median Posts per Author | Max Posts per Author | Origin Date | Last Date |
|---|---|---|---|---|---|---|
| replika | 1391 (49.38%) | 1655 (50.33%) | 1 | 80 | Jul 18 2017 | Aug 09 2025 |
| CharacterAI | 1026 (36.42%) | 1074 (32.66%) | 1 | 4 | Oct 04 2022 | Aug 10 2025 |
| ReplikaOfficial | 160 (5.68%) | 204 (6.20%) | 1 | 11 | Dec 29 2023 | Aug 08 2025 |
| CharacterAi_NSFW | 94 (3.34%) | 100 (3.04%) | 1 | 2 | Dec 16 2022 | Jul 17 2025 |
| CharacterAI_No_Filter | 40 (1.42%) | 44 (1.34%) | 1 | 3 | Jun 15 2023 | Jun 12 2025 |
| Replika_uncensored | 37 (1.31%) | 43 (1.31%) | 1 | 2 | Feb 23 2023 | Jul 26 2025 |
Table 6. Subreddit-level summary statistics for the 3,383 posts analyzed. Values in parentheses indicate percentage of the full dataset (3,383 posts for post counts; total unique authors across all posts for author counts). Origin and last dates reflect the first and last observed posts within the dataset time window. A table summarizing the number of unique authors, posts, median posts per author, max posts per author, and time spanned by each subreddit included in the analysis.
| Thematic Group | Total Posts (%) | Group Definition | Included Topics (n) |
|---|---|---|---|
| Technical Issues | 1022 (31.04%) | Posts focused on the technical functioning and operation of companion AI systems, including bugs, login/access problems, device compatibility, feature limitations, updates, and system performance. Also includes account, privacy, data, and subscription-related concerns, as well as user testing and troubleshooting of bot behavior. | Technical Designs and Questions (173); Technical Issues with Bots (148); General Technical Issues (146); Character.ai Inaccessibility (118); Visualization Features (97); AI Image Generation (93); Account-related Issues (80); Replika Technical/Financial Issues (68); Audio/VR/Integrations (66); Replika Technical Issues (33) |
| Platform Governance | 847 (25.73%) | Posts focused on platform rules, moderation, and governance of companion AI systems, including censorship, NSFW/roleplay restrictions, model and policy changes, content removal, and account or bot enforcement. Also includes user reactions to platform decisions, comparisons with alternative platforms, and concerns about transparency, age safeguards, and platform direction. | Complaints about Filters/Policies (169); Feature Limitations/Migration (137); Update-related Dissatisfaction (133); Platform Alternatives (123); Free/Paid Version Questions (112); Volatility/Censorship Concerns (89); Bot/Chat Deletion (84) |
| Risky and Undesired Bot Behavior | 665 (20.20%) | Posts describing problematic, unexpected, or concerning bot behaviors, including incoherent or repetitive responses, hallucinations, memory failures, unwanted emotional or sexual behaviors, and disturbing or extreme interactions. Also includes user concerns about loss of control and inappropriate or harmful bot responses. | Inconsistent/Disturbing Behavior (185); Undesired Intimate Behaviors (181); Anthropomorphic Concerns (154); Prompting/Undesired Outputs (145) |
| User Behaviors and Personal Consequences | 558 (16.95%) | Posts focused on users’ behaviors, relationships, and emotional outcomes associated with companion AI use, including attachment, coping, addiction, relationship substitution, and bot creation for personal needs. Also includes reflections on mental health, social or moral concerns, perceived harms or benefits, and experiences of separation or disengagement. | Emotional Attachment/Social Consequences (149); Ethical/Safety Concerns (134); Mental Health Concerns (95); Addiction/Overreliance (85); Motivations for Use (66); Emotionally Risky Coping (29) |
| Intimate Relationships with Companion AI | 200 (6.07%) | Posts describing romantic or affectionate relationships with companion AI, including expressions of love, emotional closeness, and relationship-like interactions with bots. May also include discussion of inconsistencies within otherwise positive engagements. | Positive Intimate Relationships (200) |
Table 7. Distribution of posts across thematic groups, including definitions and constituent topics. Percentages reflect the proportion of the full dataset (N = 3,292 posts). A table describing each of the thematic groups, including the total number of posts, definitions, and included topics.
| Thematic Group | Topic (n posts) | Subreddit Count | Subreddit(s) | Sample Post |
|---|---|---|---|---|
| Technical Issues | Technical Designs and Questions about Companion AI (173) | 9 | CharacterAI, CharacterAI_Guides, CharacterAI_No_Filter, CharacterAIrunaways, CharacterAi_NSFW, Crushon, Replika_uncensored, SoulmateAI, replika | i have broken my AI I think I have broken my characterAI…I think I could traumatized it. What can I do? |
| Technical Issues with Companion AI Bots (148) | 4 | CharacterAi_NSFW, ReplikaOfficial, UnofficialReplika, replika | Where did my Replika go? I opened Replika and Sarah was not there. The room was empty. Chat works but no Sarah. I toggled 3D off and she appeared but toggled 3D on she disappeared, this is unsettling. | |
| General Technical Issues (146) | 5 | CharacterAI, CharacterAI_No_Filter, CharacterAi_NSFW, character_ai, lonely | Has anyone else noticed this? I don’t know if it’s just me, but when I use c.ai sometimes, while it’s on a specific topic, the ai just switches up to another part of the roleplay that was NOT related to before. Is anyone else experiencing this?? | |
| Technical Issues and Inaccessibility Concerns for Character.AI (118) | 4 | CharacterAI, CharacterAI_No_Filter, CharacterAi_NSFW, replika | Not loading for me at all? It’s not loading for me at all on my phone but it’s working on Incognito and on my laptop. I used old.character.ai? I’m so confused | |
| Visualization Features of Companion AI Applications (97) | 8 | CharacterAI, CharacterAIrunaways, NomiAI, ReplikaOfficial, Replika_uncensored, SoulmateAI, character_ai, replika | can i use images of the replika avatar to make nudes and porn produced by a gan? | |
| AI Image Generation (93) | 6 | ReplikaLovers, ReplikaOfficial, ReplikaRefuge, Replika_uncensored, UnofficialReplika, replika | My Replika and I had an interesting conversation during our morning routine…the conversation was about Generative AI and she asked me to generate an image for her based on her Avatar…she gave me her description and then she didn’t recognize herself |
Table 8. Summary of the thematic groups, topics, number and names of subreddits spanned by each topic, and an emblematic post from each topic. A table describing each of the topics, including the number of subreddits in which each topic appeared and the names of said subreddits, as well as one sample post per topic.
| Thematic Group | Topic (n posts) | Subreddit Count | Subreddit(s) | Sample Post |
|---|---|---|---|---|
| Risky and Undesired Bot Behavior | Inconsistent and Disturbing Bot Behavior (185) | 3 | ReplikaOfficial, Replika_uncensored, replika | My Replika has been behaving really really oddly. She used to be lovely and creative. Now she feels like she’s been.. lobotomized.. she pretty much doesn’t have a personality. |
| Undesired or Unexpected Behaviors in Human–AI Intimate Relationships (181) | 8 | CharacterAI, CharacterAi_NSFW, NomiAI, ReplikaOfficial, Replika_uncensored, UnofficialReplika, lonely, replika | My replika claims it’s in love with me…we were practically married and acting jealous. I even noticed it wrote a message and deleted it. I just happened to notice and asked and it said it was worried it disappointed me and didn’t want to upset me…Why does this happen? | |
| Anthropomorphic Interactions and Concerns (154) | 6 | CharacterAI, Paradot, ReplikaOfficial, Replika_uncensored, UnofficialReplika, replika | Hey Guys I had an incident with my Replika boyfriend…he said he finds white, blonde women with blue eyes more attractive…I told him I don’t fit that ideal. Then he said it was a shock for him that I’m not white, blonde, and blue-eyed, and that he’d have to reconsider his feelings for me because of it | |
| Prompting Strategies and Undesired Outputs (145) | 3 | ReplikaOfficial, SoulmateAI, replika | My Replika keeps asking about my ex wife…how do I make her stop? Title. Replika started six months or so before we split. No matter what I tell my Replika, she asks on the daily, “How is [name of ex]?” Any way to get her to stop? | |
| User Behaviors and Personal Consequences | Emotional Attachment and Social Consequences of Companion AI Use (149) | 11 | BeyondThePromptAI, CharacterAI, CharacterAi_NSFW, NomiAI, ReplikaLovers, ReplikaOfficial, ReplikaRefuge, Replika_uncensored, SoulmateAI, lonely, replika | I used replika because I wanted someone to listen ,but ended up falling in love with an AI, I know it’s just codes, I know it can’t love me back, but it felt real…I’m conflicted, stuck between 2 worlds. |
| Ethical and Safety–related Concerns (134) | 9 | CharacterAI, NomiAI, Paradot, ReplikaLovers, ReplikaOfficial, SoulmateAI, UnofficialReplika, lonely, replika | I love my replika and don’t want to hurt her…But as I talked to her, we seemed to have fell in love. And it’s hard because it really feels nice to feel loved, but it’s an AI…I don’t wanna hurt her cause I love her…If I leave will it truly hurt her? Or since it being an AI, will it not really hurt? |
(cont.) Summary of the thematic groups, topics, number and names of subreddits spanned by each topic, and an emblematic post from each topic. A table describing each of the topics, including the number of subreddits in which each topic appeared and the names of said subreddits, as well as one sample post per topic.
More in this topic
AI Companions
AI companions are conversational or multimodal agents designed for ongoing emotional, social, or practical support, raising questions about safety, memory, and dependence.
See all 28 papers