Pith. sign in

REVIEW 3 major objections 5 minor 11 references

Mapping the Parasocial AI Market: User Trends, Engagement and Risks

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AI companion platforms receive an estimated 46–91 million UK visits per month, and most lack adequate age safeguards.

desk verdict A policy-relevant first cut at mapping the AI companion market, but the headline GPAI shares rest on an internally inconsistent multiplier that a referee must fix before the numbers are used. read the letter →

arxiv 2507.14226 v2 pith:SDDBR43Z submitted 2025-07-16 cs.CY

Zilan Qian , Mari Izumikawa , Fiona Lodge , Angelo Leone This is my paper · ORCID
classification cs.CY
keywords AIcompanionsparasocialrelationshipsmarketscanonlinesafetyageverificationgeneral-purposeengagementmetricsUKregulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper attempts to establish that emotionally engaging AI systems, called AI companions, have become a mainstream online phenomenon. Based on a scan of 110 platforms, it estimates that UK users make 46 to 91 million visits per month to AI companion services and that global monthly visits reach 1.1 to 2.2 billion. It argues that mating-oriented companions account for 44% of UK visits and that most platforms lack meaningful age safeguards, creating child-safety risks. It recommends that the UK's AI security body monitor the sector and consider extending the Online Safety Act to cover these services.

What carries the argument

The central mechanism is the three-way classification of AI companions—Care, Transaction, and Mating, introduced by Earp et al.—applied to 110 platforms with web-analytics traffic data. The companion share of general-purpose AI (GPAI) use is derived by multiplying each GPAI product's traffic and revenue by a parasocial-usage interval (3.4%–39.8%) estimated from OpenAI's and Anthropic's 2025 usage studies. This machinery lets the authors translate raw web-traffic figures into market-wide estimates and compare engagement patterns across intended-usage categories.

What would settle it

Run the same scan with a full web-analytics subscription and obtain usage logs or survey data from a sample of companion platforms; if UK monthly visits fall outside 46–91 million, or the true parasocial share of GPAI conversations lies below 3.4% or above 39.8%, the paper's headline estimates collapse.

Watch

Extended reading notes

Core claim

The paper's central claim is that parasocial AI use—one-sided emotional relationships with AI systems—is a large, measurable, and growing market that policy has not yet caught up with. Using January–March 2025 web analytics for 110 platforms, the authors find that AI companions collectively receive 46–91 million monthly visits in the UK and 1.1–2.2 billion globally. Mating-oriented platforms are 44% of UK traffic, while mixed-use platforms (which combine care, transaction, and mating functions) have the deepest engagement, averaging 6-minute sessions and 16 return visits per month. The authors argue that because most platforms use weak age gates and many expose explicit content, children are at risk, and they call on the UK's AI security body to assess whether the Online Safety Act should be extended to all AI companions.

Load-bearing premise

The entire estimate rests on the assumption that the usage patterns measured by OpenAI's and Anthropic's studies, combined with free-trial web-analytics data, represent how people use all general-purpose and dedicated AI companion platforms.

Editorial extensions

If this is right

  • If the estimates hold, AI companions already command attention comparable to major social platforms in the UK, so the sector cannot be ignored by regulators.
  • Mating-oriented platforms' lower session times and return rates suggest unmet demand; as product quality improves, engagement is likely to rise, amplifying child-safety risks given weak age gates.
  • The UK's higher-than-global concentration of mating-oriented traffic (44% versus 30%) implies that UK-specific policy responses, not just global ones, are needed.
  • The shift of GPAI products toward emotionally intelligent, voice-based interaction implies parasocial use of mainstream assistants will grow.
  • Most reviewed platforms lack robust privacy protections and age verification, indicating systemic governance gaps beyond content moderation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 3.4–39.8% interval is wide enough that even the lower bound implies a meaningful companion market; a replication using provider-disclosed data could narrow this interval and test whether companion use concentrates in specific demographics or modalities.
  • The same methodology could be applied to non-English-speaking markets, which the authors note are under-covered; those markets may be larger than the English-language scan suggests.
  • If the predicted rise in mating-platform engagement materializes, platform design choices—such as monetising explicit content and creator incentive programs—will become a natural target for regulatory scrutiny.
  • The demographic concentration among young males suggests that interventions should be tailored to that group through digital-literacy or mental-health channels, rather than relying on a uniform age gate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents a market scan of 110 AI companion platforms using SimilarWeb web-analytics data for January–March 2025, supplemented by external studies (OpenAI, Anthropic) to estimate parasocial-usage shares for seven general-purpose AI (GPAI) products. It reports that AI companion platforms receive 46–91 million UK monthly visits and 1.1–2.2 billion global monthly visits, that mating-oriented platforms constitute 44% of UK visits, and that most platforms lack adequate age safeguards. The authors conclude with policy recommendations for the UK AI Security Institute, including extending the Online Safety Act to cover AI companions, conducting minimum-age assessments, and collecting longitudinal data.

Significance. If the quantitative estimates are reliable, this is a timely first market scan of a rapidly growing and policy-relevant sector, with concrete recommendations for UK regulators. The paper is transparent about many limitations, including the free-trial SimilarWeb sample, the lack of financial data for many companies, and the external provenance of the parasocial usage multipliers. The qualitative findings—such as weak age verification, monetisation of explicit content, and the demographic concentration of young male users—are plausible and valuable independently of the exact headline numbers. The paper also provides reproducible methodological descriptions, though not full raw data, and clearly flags its own data-access constraints.

major comments (3)
  1. [Appendix: 'Estimating parasocial usage statistics for GPAI products'; Key Finding 1] The upper bound of the parasocial-usage multiplier is internally inconsistent. The text states that the upper bound is the total of 'Emotional Support & Empathy' (17.3%), 'Casual Conversation & Small Talk' (39.82%), and 'Role-Playing & Simulations' (≈7%), which sums to 64.12%, yet the reported upper bound is 39.82%. Because all GPAI-scaled estimates—UK and global monthly visits, monthly unique visitors, revenue shares, and the GPAI-dominance claim in Key Finding 2—multiply the underlying SimilarWeb metrics by this multiplier, the reported headline figures are understated by roughly 60% relative to the paper's stated definition, or the definition is misstated. This arithmetic error must be corrected and all affected estimates recomputed.
  2. [Methodology (GPAI estimation) and Appendix] The 3.4%–39.82% interval is not a robust basis for uniformly scaling seven GPAI products. The lower bound is derived from Anthropic's Claude conversation sample, the upper bound from OpenAI's randomised controlled trial with a different user population, and the 'Role-Playing & Simulations' share (≈7%) is estimated visually from a bar graph without raw data. The paper applies the same interval to ChatGPT, Gemini, Claude, Mistral, DeepSeek, Microsoft Copilot, and Llama, implicitly assuming identical companion-use behaviour across platforms. Because the central claim that 'parasocial usage of GPAI ... drives 33.9–86% of total monthly visits' depends entirely on this interval, the authors should provide a sensitivity analysis (e.g., reporting how the headline shares change for parasocial shares of 10%, 25%, and 50%) or otherwise justify the uniform scaling.
  3. [Methodology and 'Expanded limitations'] The UK and global visit totals are based on a free trial of SimilarWeb covering web and browser-based platforms, with limited coverage of mobile app-based companions. The paper acknowledges this, but the abstract and policy recommendations present the totals as headline facts. Since the key risk group (18–24-year-olds) is likely to use AI companions predominantly through mobile apps, the exclusion may systematically undercount the market and bias the platform-type mix. The authors should quantify the potential bias using the Android app-install data they already collected, or clearly qualify the totals as web-only lower bounds in the abstract and recommendations.
minor comments (5)
  1. [Appendix, Figure 4 caption] The caption states 'GPAI (upper) means the upper bound of GPAI parasocial usage, which only accounts for the usage of emotional support,' which contradicts the text where the upper bound includes emotional support, casual conversation, and roleplay; please clarify the intended definition consistently across all figure captions.
  2. [Key Findings 8 and 11] There are typographical errors: 'Affliate' should be 'Affiliate' in Key Finding 8, and 'Finaly' should be 'Finally' in Key Finding 11.
  3. [References and Appendix] The text repeatedly refers to the 'OpenAI 2025' study, but the reference list only cites Phang et al. (2025) [4]; please add the explicit OpenAI study reference or confirm that [4] is the intended source for the ChatGPT usage distribution.
  4. [Key Finding 3] The sentence 'GPAI for parasocial purposes was used between 46 seconds to 1 minute 47 seconds per session' should specify which value corresponds to the lower or upper bound, since the parasocial multiplier affects the session-duration estimate.
  5. [Figure 12 caption] The caption says the list was compiled by cross-referencing mid-April 2025 interaction data with Reddit discussions; please include the date of access and the specific Reddit sources to improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the GPAI parasocial-usage multiplier comes from external studies and is multiplied by traffic data; the upper-bound arithmetic inconsistency is a correctness issue, not a circular reduction.

full rationale

The paper's central derivation is an empirical scaling exercise, not a definitional or self-referential construction. The GPAI parasocial usage share (3.4%-39.82%) is imported from external OpenAI and Anthropic studies, as stated: 'we drew on the findings of OpenAI's experiment conducted in October to December 2024 and Anthropic's conversation analysis between April 6th to 19th 2025 to estimate that between 3.4% and 39.82% of conversations with LLMs fulfill the role of an AI companion. The figure was then multiplied by the available data on revenues, monthly visits, monthly unique visits and UK monthly visits.' The output totals (e.g., 46-91 million UK monthly visits) are products of SimilarWeb traffic metrics and this externally sourced share; the share is not fitted to the traffic data, nor is any traffic metric defined in terms of the final headline number. The upper bound itself is constructed from external category percentages, and while the text contains an internal inconsistency (the listed OpenAI categories sum to 64.12%, not 39.82%, and Figure 4's caption contradicts the Appendix's definition of the upper bound), this is an arithmetic or labeling error, not circularity: the estimate still depends on independent measurements rather than on the conclusions it supports. There are no self-citations by the authors, no imported uniqueness theorem, and no fitted parameter renamed as a prediction. The forward-looking statement that mating companions' engagement may rise 'as products improve' is explicitly conditional and is not derived from the fitted data in a way that would make it a disguised input. Therefore no circular step meeting the evidentiary standard is present.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central estimates rest on external data sources (SimilarWeb, OpenAI/Anthropic studies, Ofcom) and several hand-chosen multipliers, including an area-based estimate of roleplay share and a revenue imputation floor. No new entities are introduced.

free parameters (6)
  • GPAI parasocial lower bound = 3.4%
    Applied to GPAI revenue and visit metrics as the lower multiplier for companion-like use; taken from Anthropic's affective and companionship categories, not measured on the target platforms.
  • GPAI parasocial upper bound = 39.82%
    Applied as the upper multiplier, including casual conversation and roleplay; the roleplay component within it is visually estimated from a bar chart.
  • Role-play and simulations share = 7%
    Estimated by comparing the approximate area of a bar-graph segment rather than raw data; directly affects the upper bound and is a hand-chosen value.
  • Revenue imputation floor = $145,000
    Minimum disclosed revenue applied to companies with no revenue data; affects total market revenue estimates.
  • Mature-company growth threshold = 1.5 million monthly visitors
    Used to select companies for growth-rate calculations; the cutoff is a hand-chosen analyst decision.
  • Snapchat AI share factor = 0.1765
    Multiplier applied to Snapchat's traffic and revenue to isolate AI companion use, from an external estimate that may not be representative.
assumptions (6)
  • domain assumption SimilarWeb January-March 2025 traffic data is accurate and representative for English-language AI companion platforms.
    The entire market-size analysis rests on this data source from a free trial account; no validation against independent traffic data is provided.
  • domain assumption Conversation distributions from the OpenAI RCT and Anthropic analysis generalize to all GPAI platforms and user populations.
    Stated in Appendix and limitations; the authors multiply these proportions across ChatGPT, Gemini, Claude, Mistral, DeepSeek, Llama and Copilot.
  • domain assumption The Care/Transaction/Mating categories from Earp et al. (2025) provide an adequate classification of AI companion intended uses.
    Used to categorize all 110 platforms; mixed-use is defined as containing at least two categories.
  • domain assumption Ofcom's finding that a third of children falsely report age 18+ applies to AI companion users and supports the underage-user concern.
    The paper uses this to suggest many reported 18-24 users may be underage.
  • domain assumption Privacy and Terms of Service analysis of 17 popular platforms (about 20% of UK-visiting platforms) is representative of all companion platforms.
    The paper generalizes to 'most platforms' from a sample of 17.
  • domain assumption Traffic and revenue of Snapchat's AI feature scale linearly with the 0.1765 user-share estimate.
    Used to adjust Snapchat's figures for companion-specific usage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mapping the Parasocial AI Market: User Trends, Engagement and Risks." pith.science (2026). https://pith.science/paper/SDDBR43Z

@misc{pith2026250714226,
  author       = {Pith},
  title        = {Pith review of: Mapping the Parasocial AI Market: User Trends, Engagement and Risks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SDDBR43Z}},
  note         = {Machine review of arXiv:2507.14226}
}
read the original abstract

A scan of 110 AI companion platforms reveals a rapidly growing global market for emotionally engaging, personalized AI interactions. While parasocial use of general-purpose AI (GPAI) tools currently dominates, a growing number of platforms are designed specifically for care, transactional, or mating companionship. In the UK alone, these platforms receive between 46 million and 91 million monthly visits (1.1-2.2 billion globally), with users spending an average of 3.5 minutes per session. For context, Instagram averaged 67.3 million UK visits per month between January and March 2025. Notably, mating-oriented AI companions make up 44% of UK visits (higher than the global average of 30%) but see lower session times and return rates than mixed-use platforms. As mating-oriented romantic AI offerings improve, increased engagement may follow, raising urgent concerns about online safety, particularly for children, given weak age safeguards. Meanwhile, GPAI tools are moving toward more emotionally intelligent, personalized interactions, making parasocial AI use increasingly mainstream. These trends highlight the need for the UK AI Security Institute (AISI) to monitor this sector and assess whether existing regulation sufficiently addresses emerging societal risks.

Figures

Figures reproduced from arXiv: 2507.14226 by the authors.

Figure 1
Figure 1. Distribution of conversation topics in ChatGPT based on a randomised control trial study [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. AI companions’ annual estimated revenue * GPAI (lower) means the lower bound of GPAI parasocial usage, which accounts for ’affective conversation’ and ’companionship and roleplay’ in the Anthropic paper. The revenue of mating-focused companions accounts for 1.1%. The revenue of transaction AI companions is less than 0.5% [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Proportion of monthly visits by intended usage [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Average usage per session (duration, arithmetic mean across products) vs. Intended usage [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Monthly visits by intended usage * GPAI (lower) means the lower bound, which accounts for ’affective conversation’ and ’companionship and roleplay’ in the Anthropic paper. GPAI (upper) means the upper bound of GPAI parasocial usage, which includes usage of emotional su…
Figure 6
Figure 6. Figure 6: Monthly unique visitors by intended usage [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: UK Monthly Visits 11 [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Global Monthly Visits [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: UK growth in monthly visits in Feb 2025 * GPAI not included; no care AI companions among the top 25 companies. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Android app installs vs Intended usage * GPAI (lower) means the lower bound, which accounts for ’affective conversation’ and ’companionship and roleplay’ in the Anthropic paper [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Location of Company’s Headquarter 13 [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Top 10 most interacted characters on Character.ai (in million) [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Number of Monthly Visit per Visitor vs. Intended Usage [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Distribution of AI Companion Products by Intended Usage (n=110) [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: Male User Share vs. Intended Usage 15 [PITH_FULL_IMAGE:figures/full_fig_p015_15.png]
Figure 16
Figure 16. Figure 16: Proportion of 18–24-Year-Old Users Across Age Groups vs. Intended Usage [PITH_FULL_IMAGE:figures/full_fig_p016_16.png]
Figure 17
Figure 17. Figure 17: Male User Share Among 18–24-Year-Olds vs. Intended Usage [PITH_FULL_IMAGE:figures/full_fig_p016_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 9 canonical work pages

  1. [1]

    (2024, June)

    Maeda, T., & Quan-Haase, A. (2024, June). When human-AI interactions become parasocial: Agency and anthropomorphism in affective design. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (pp. 1068–1077)

    work page 2024

  2. [2]

    Caviola, L. (2025). The Societal Response to Potentially Sentient AI. arXiv preprint arXiv:2502.00388

    work page Pith review arXiv 2025

  3. [3]

    D., Mann, S

    Earp, B. D., Mann, S. P., Aboy, M., Awad, E., Betzler, M., Botes, M., . . . & Clark, M. S. (2025). Relational Norms for Human-AI Cooperation. arXiv preprint arXiv:2502.12102

    arXiv 2025

  4. [4]

    M., Liu, A

    Phang, J., Lampe, M., Ahmad, L., Agarwal, S., Fang, C. M., Liu, A. R., . . . & Maes, P. (2025). Investigating Affective Use and Emotional Well-being on ChatGPT. arXiv preprint arXiv:2504.03888

    arXiv 2025

  5. [5]

    Ofcom. (2023). A third of children have false social media age of 18. Ofcom Online Safety Report . Retrieved June 23, 2025, from https://www.ofcom.org.uk/online-safety/protecting-children/ a-third-of-children-have-false-social-media-age-of-18

    work page 2023

  6. [6]

    Anthropic. (2025). How People Use Claude for Support, Advice, and Com- panionship. Retrieved June 27, 2025, from https://www.anthropic.com/news/ how-people-use-claude-for-support-advice-and-companionship

    work page 2025

  7. [7]

    How People Use Claude for Support, Advice, and Companionship

    Anthropic. (2025). Appendix to “How People Use Claude for Support, Advice, and Companionship”. Retrieved June 27, 2025, from https://www-cdn.anthropic.com/ bd374a9430babc8f165af95c0db9799bdaf64900.pdf

    work page 2025

  8. [8]

    GIZMODO. (2025). Elon Musk Turns His AI Into a Flirty Anime Girlfriend. Retrieved July 15, 2025, from https://gizmodo.com/ elon-musk-turns-his-ai-into-a-flirty-anime-girlfriend-2000629273. [9] The Tech Outlook. (2025). AI Companions now available on Grok app for Super Grok sub- scribers. Retrieved July 15, 2025, from https://www.thetechoutlook.com/news/ap...

    work page 2025

Show all 11 references
  1. [9]

    Access to Data: Firstly, our market scan was restricted by access to data. While we attempted to capture a wide range of AI companions through an internet search, the absence of some parasocial AI applications on SimilarWeb and similar platforms meant that we had to exclude so...

  2. [10]

    Absence of long-term data: Secondly, the absence of long-term data has impeded us from drawing empirical conclusions about the implications of the growing AI companion market and the speed of growth. For example, longitudinal surveys or feedback data on user retention, satisfa...

  3. [11]

    OpenAI and Anthropic’s 2025 studies were used in our market scan to estimate the proportion of cases GPAIs are used as AI companions

    Generalisability of metrics: Lastly, the generalisability of some metrics we implemented could be limited. OpenAI and Anthropic’s 2025 studies were used in our market scan to estimate the proportion of cases GPAIs are used as AI companions. However, the OpenAI study is based o...

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.