Updated September 2026: the original version of this article contained several claims Robius should not have published as written. It said exactly 128 AI companion apps were in active distribution, claimed Robius had tested the five platforms that “actually matter,” ranked their privacy and memory performance without a reproducible test method, described Google as Character.AI’s parent company, quoted the wrong amount for an Italian Replika privacy fine, and gave a precise loneliness-reduction percentage that we could not trace cleanly to the research cited.
Those claims have been removed or corrected. The subject itself remains important. MIT Technology Review included AI companions in its 2026 list of 10 Breakthrough Technologies, regulators are actively examining the category, and peer-reviewed research suggests these systems can reduce loneliness in the moment for some users while leaving important questions about longer-term effects, dependency, privacy and youth safety unresolved.
| THE ROBIUS READ: AI companions can provide a genuine feeling of being heard, and research shows that short interactions can reduce loneliness for some users. That is not the same as proving long-term mental-health benefit. These products also process unusually intimate conversation data, and regulators have already intervened over privacy and age-protection failures. Do not choose an AI companion from a “best app” ranking alone. Read the current privacy policy, age rules, deletion terms and safety design before giving it the most personal parts of your life. |
First, the “128 Apps” Number Was Misread
The earlier Robius article opened with a very precise claim: 128 AI companion apps were in active distribution, up from 16 three years earlier.
That is not what the strongest source we could locate said. In August 2025, TechCrunch reported data supplied by app-intelligence company Appfigures showing 337 active and revenue-generating AI companion apps worldwide. Of those, 128 had been released during 2025 at the time of the analysis.
In other words, 128 was a launch count inside a larger market, not the total number of companion apps in existence. The old Robius wording appears to have inherited the statistic from a secondary comparison article and changed its meaning.
We are not replacing it with a new “exact number of companion apps today.” Definitions vary, app stores change constantly and many products blur the line between general chatbot, roleplay platform, wellness tool and dedicated companion.
The Category Is Still Important Enough to Be a 2026 Breakthrough Technology
One part of the old opening was correct. MIT Technology Review named AI companions among its 10 Breakthrough Technologies of 2026.
That should be read as a judgment about technological and social significance, not an endorsement of companion apps as safe or therapeutic. The publication’s own framing highlighted both the rapid formation of intimate relationships with chatbots and the risks that can follow.
The Privacy Question Is Different From a Normal Chat App
A companion works better when it knows more about you. That creates the product’s central privacy tension.
People may tell an AI companion about relationships, grief, health, sexuality, work problems, family conflict, finances or fears they would never put into a normal customer-service chatbot. The conversation history itself can therefore become unusually sensitive even when the app never asks for a passport number or bank card.
The right comparison is not “which app collects the fewest boxes on an App Store label?” It is what each company’s current policy says about conversation content, sensitive information, model providers, advertising, retention, deletion, training and age restrictions.
Replika: The Current Policy Is More Nuanced Than the Old Privacy Table
Replika’s privacy policy, updated 27 May 2026, says the service collects account and profile information plus the messages, photos, videos, voice messages and other content users provide in conversations.
The policy says Replika does not use or disclose the content of Replika conversations for marketing or advertising. It also says conversation data can be sent, subject to de-identification and data-minimisation measures, to third-party AI language-model providers to generate responses. According to the policy, those providers are contractually restricted from using the data to train their own models.
Replika also tells users not to provide special categories of personal data, while acknowledging that intimate conversations can lead users to disclose such information anyway. If that happens, the company says the information is processed as part of conversation data to provide the service.
That is more precise than the old Robius table, which reduced the privacy story to a vague “moderate” label.
Replika Was Fined €5 Million in Italy — Not €5.6 Million
In May 2025, Italy’s data-protection authority imposed a €5 million fine on Luka Inc., the company behind Replika. The regulator said violations identified during its earlier investigation had occurred, including failures involving the legal basis for processing and age-related protections.
The previous Robius article stated €5.6 million. That number was wrong. The regulator’s own English-language announcement says €5 million.
Character.AI: The Regulatory Record Changed Again in July 2026
Character.AI deserves separate treatment because its youth-safety and privacy position changed materially after the original Robius article was published.
On 9 July 2026, Italy’s data-protection authority announced a €158,000 fine against Character Technologies Inc., the US company that operates Character.AI. The authority identified privacy-notice and compliance shortcomings and said it had also found concerns around protections for minors and age-verification mechanisms.
Character.AI has also changed the product. The company says it removed open-ended chat with Characters for users under 18 beginning in late 2025. Its current under-18 experience is a reading mode without open-ended chat, backed by age-assurance systems that can escalate to selfie or, in limited cases, identity-document verification through third-party provider Persona.
This is exactly why an old feature comparison can become misleading quickly: the safety architecture itself is changing.
Google Is Not Character.AI’s Parent Company
The previous article said “Character.AI and its parent company Google” had agreed to settle lawsuits concerning teen mental-health harms. That corporate relationship was described incorrectly.
Character Technologies is the company behind Character.AI. Google separately entered a licensing arrangement with Character.AI and rehired its founders. Google was named as a co-defendant in litigation over a Florida teenager’s death, and Reuters reported in January 2026 that Google and Character.AI had reached a settlement with the teenager’s mother. The terms were not disclosed.
A settlement is not the same thing as a regulatory finding or a judicial determination of liability. It is relevant to the safety history of the category, but Robius should describe it accurately rather than turning Google into Character.AI’s parent company.
The FTC Is Studying Companion Chatbots and Children
In September 2025, the US Federal Trade Commission launched a formal study into consumer-facing AI chatbots acting as companions. It issued information orders to seven companies, including Alphabet, Character Technologies, Meta, OpenAI, Snap and xAI.
The FTC said it wanted to understand how companies measure and mitigate possible negative effects on children and teens, how they monetize engagement, how they disclose risks, how they enforce age restrictions, and how they use or share personal information obtained through chatbot conversations.
That inquiry is useful context because it focuses on the same questions a consumer should ask: what is the product optimizing for, what data does the conversation create, and what happens when a user becomes emotionally attached?
Nomi: Read the Actual Policy Instead of Calling It “the Privacy Winner”
The old Robius article called Nomi the most privacy-conscious major platform. We are removing that ranking because “best privacy” requires a defined audit methodology, not a quick comparison table.
Nomi’s privacy policy, updated 27 April 2026, says the company does not sell or rent personal information and encourages users not to include personally identifiable information in their interactions. It collects the content users provide in chats and customisation, along with account and usage information needed to operate the service.
The policy also explains that personal information may be stored on servers outside the user’s country and describes deletion rights and retention exceptions. Nomi’s separate account-deletion guidance says deleting an account permanently removes associated messages, media and personal data.
Those are useful facts. They still do not justify a universal “privacy winner” badge without comparing the same criteria, legal jurisdictions, security architecture and actual data flows across every competing product.
What the Loneliness Research Actually Shows
The earlier Robius article said a published Replika study found an average 15.2% reduction in loneliness after four weeks. We could not trace that precise figure cleanly to the study we should have been citing, so it has been removed.
There is nevertheless credible evidence that AI companions can reduce loneliness in the short term.
A peer-reviewed paper in the Journal of Consumer Research, published online in 2025 and appearing in the April 2026 issue, ran several studies on AI companions and loneliness. The researchers found that companion interactions could produce momentary reductions in loneliness, including over a one-week longitudinal study, and that feeling heard was an important part of the effect.
That is meaningful. It is not proof that long-term use improves mental health, prevents social isolation or should replace human support.
Other 2025–2026 research points to a more mixed picture. Studies of real-world companion use report emotional validation and opportunities to rehearse social interaction, but also raise concerns about over-reliance, withdrawal and the way some systems respond to users in vulnerable states.
The evidence therefore supports a narrower conclusion than the old article: an AI companion can make some people feel less lonely in the moment, while the longer-term psychosocial effects remain an active research question.
Do Not Assume the Business Model Is Automatically Harmful Either
The previous article said almost every companion platform was commercially optimized to keep users talking longer and implied that deeper attachment was inherently part of the revenue model.
Engagement is obviously commercially valuable to subscription and consumer-app businesses, and the FTC is explicitly asking companies how they monetize engagement. But Robius should not claim to know the optimization objective of every major companion app without internal evidence.
The better question for a user is observable: does the product use streaks, notifications, relationship progression, paid intimacy features, virtual currency or other mechanics that encourage more interaction? If so, understand that design before deciding how much emotional space you want the app to occupy.
What to Check Before Sharing Intimate Information
- Conversation use: does the policy say chats are used only to provide the service, or also to improve models, train safety systems or perform other analysis?
- Third-party models: does conversation content leave the companion company and go to another AI provider?
- Sensitive data: what does the company say about health, sexuality, religion, finances and other sensitive information disclosed inside a conversation?
- Advertising: are conversation contents used for advertising, or only separate website and device data?
- Deletion: can you delete individual messages, a companion, the whole account, or all three? How long does deletion take and what exceptions remain?
- Age rules: is the service adults-only, age-gated or available in a restricted teen mode?
- Safety: what happens when the conversation involves self-harm, abuse, delusions or another vulnerable state?
And regardless of the policy, avoid giving an AI companion information you do not need it to have. A good privacy policy reduces risk; it does not make unnecessary disclosure risk-free.
If You Are Using One for Emotional Support
An AI companion can be a conversation tool. It should not be treated as a licensed therapist, emergency service or guaranteed source of safe mental-health advice unless the service is specifically regulated and qualified for that purpose.
If the app helps you rehearse a difficult conversation, put thoughts into words or feel heard for a while, that can still be useful. The more consequential the situation becomes, the more important it is to bring a real person, qualified professional or appropriate emergency service into the loop.
The Bottom Line
AI companions are a real and rapidly evolving consumer-AI category. They are important enough that MIT Technology Review put them on its 2026 breakthrough list and regulators in the US and Europe are actively examining how they affect privacy, children and vulnerable users.
The research does not support either extreme. These systems are not automatically meaningless toys: controlled studies show that they can reduce loneliness in the moment. But they are also not proven long-term substitutes for human relationships or professional mental-health support.
And there is no defensible reason for Robius to declare one app the “best” for privacy from an undocumented five-platform test. The better standard is transparent: read the current policy, understand how conversation data is processed, look at the regulator record, check the age and deletion rules, and decide how much intimate information the service genuinely needs.
Sources
• TechCrunch / Appfigures, August 2025: 337 active revenue-generating AI companion apps, with 128 released during 2025 at the time of the analysis — techcrunch.com
• MIT Technology Review, 2026: AI companions included in the annual 10 Breakthrough Technologies list.
• Italian Data Protection Authority, 19 May 2025: €5 million fine against Luka Inc., operator of Replika — garanteprivacy.it
• Replika Privacy Policy, updated 27 May 2026: conversation content, sensitive data, third-party AI providers, advertising and data processing — replika.com
• Italian Data Protection Authority, 9 July 2026: €158,000 Character Technologies fine and required measures concerning privacy and age verification — garanteprivacy.it
• Character.AI, September 2026: current safety update confirming open-ended chat remains removed for under-18 users — character.ai
• Character.AI Age Assurance FAQ, updated August 2026: current age-assurance and Persona verification process — character.ai
• Reuters, 7 January 2026: settlement involving Character.AI, Google and the Florida lawsuit; settlement terms undisclosed — Reuters.
• Federal Trade Commission, September 2025: inquiry into AI chatbots acting as companions, including safety, monetization and data handling for children and teens — ftc.gov
• Nomi Privacy Policy, updated 27 April 2026: collection, use, sharing and deletion practices — nomi.ai
• Journal of Consumer Research: “AI Companions Reduce Loneliness,” Volume 52, Issue 6, April 2026 — Oxford Academic
• Yuan et al., 2025: “Mental Health Impacts of AI Companions,” mixed psychosocial effects in longitudinal and interview research — arXiv
• Chu et al., June 2026: “When Chatbots Accommodate,” analysis of how companion systems respond during vulnerable conversations — arXiv
Robius.news — Dubai, UAE — 2026 | Built to be first. Built to be trusted.
