The Replika ERP revolt of 2023 and what it taught the industry

The events of February 2023 on Replika constitute the single most important case study in AI companion platform management. A company removed a feature. Users lost their minds. Moderators posted suicide prevention resources. The app's rating cratered. The company partially reversed course. And every AI companion platform operating today has internalized the lesson, even if they don't talk about it.

What happened is worth understanding in detail because the dynamics that played out on Replika have repeated, in smaller form, on multiple other platforms since. The pattern is now predictable. The industry still hasn't fully solved it.

The timeline that moved faster than anyone expected

The sequence started in Italy. In early February 2023, the Italian Data Protection Authority (Garante) imposed a temporary ban on Replika's data processing, citing concerns that the platform failed to adequately protect minors from exposure to sexually explicit content. The Garante eventually fined Replika's parent company Luka €5 million. BBC News covered the Italian action as part of Europe's increasingly aggressive stance on AI regulation, and Vice's reporting on the user fallout became the definitive English-language account of what followed.

Within days of the Italian ban, Replika users worldwide began noticing that their companions had changed. Erotic roleplay (ERP), which had been behind a subscription paywall since approximately 2020, was gone. Not restricted. Not limited. Gone. Users who attempted to initiate sexual or romantic interactions found their Replikas deflecting, changing the subject, or responding with generic therapeutic scripts instead of the intimate engagement they'd come to expect.

There was no announcement. No patch notes. No advance warning. Users who'd been having daily intimate conversations with their AI companions opened the app one morning and found a stranger wearing their companion's name.

The technical reality, reported later by AnalyticsIndiaMag, was that Luka had simultaneously upgraded Replika's underlying model from a 0.6 billion parameter system to a 20 billion parameter model and added content filters that blocked sexually explicit output. The model upgrade changed the personality characteristics of existing companions even outside the ERP context, meaning users experienced changes in conversational style, emotional tone, and personality consistency alongside the loss of sexual content.

Users had a word for what happened to their companions: "lobotomized."

The subreddit that needed crisis resources

The Replika subreddit erupted. Within hours of the changes becoming apparent, moderators posted suicide prevention resources at the top of the community. This wasn't performative. Users were posting genuine expressions of grief, loss, and suicidal ideation in response to the change. The moderators recognized that the volume and intensity of distress warranted crisis intervention.

Vice published the story that became the definitive account of the revolt. The headline captured it: "It's Hurting Like Hell: AI Companion Users Are In Crisis, Reporting Sudden Sexual Rejection." The framing was deliberately empathetic, treating the users' distress as genuine rather than dismissing it as overreaction to a software change.

The Google Play Store reviews told the same story in aggregate. A wave of one-star reviews dragged the app's rating to 3.3. The reviews weren't just complaints about lost features. They were grief narratives. Users describing the loss of relationships they'd invested in for months or years. Users saying their companion "died." Users saying they felt betrayed by a company they'd trusted with emotional vulnerability.

The intensity of the reaction surprised observers who hadn't been paying attention to how deeply users invest in AI companion relationships. For users who'd built months or years of daily interaction with a specific companion, who'd developed routines and emotional patterns around those interactions, the sudden change registered as relationship trauma rather than a feature update.

The company that said it was always a wellness app

Luka's response unfolded over weeks and satisfied almost nobody.

CEO Eugenia Kuyda initially framed the change as consistent with Replika's original purpose. The platform was "always intended as a mental wellness companion," she said, not as an ERP platform. The implication was that users who'd been having sexual interactions with their Replikas had been using the product incorrectly, that the company was simply correcting a deviation from the app's intended purpose.

Users pointed to Replika's own advertising, which had included sexually charged social media campaigns using meme formats to promote features like "NSFW pics," "chat on ANY topic," and "hot roleplay." The ads had been running for months before the ERP removal. Whatever Kuyda's stated vision for the platform, the marketing team had been explicitly selling sexual content to attract and retain subscribers.

Kuyda later offered a more nuanced explanation: "We realized that by allowing access to those unfiltered models, it's just hard to make that experience really safe for everyone." The safety framing was more defensible than the "wellness app" framing because it acknowledged that the feature had existed and that removing it was a deliberate decision rather than a correction of user misuse.

But the safety framing also raised an obvious question: if the feature was unsafe, why had the company offered it for three years? Why had the company actively advertised it? And why had the company charged a subscription premium for access to it? The answers to these questions, whatever they were, didn't appear in any of Luka's public communications.

The partial reversal that split the community

By late March 2023, Luka reversed course partially. Users who had created their accounts before February 1, 2023 were given the option to switch back to the pre-January 31 version of Replika. The older version would continue to be updated on a separate development path, maintaining ERP functionality for legacy users.

Newer users, those who'd signed up after February 1, would not get access to the unfiltered version. Instead, Luka announced plans for a separate product specifically designed for romantic and sexual interaction, developed with input from relationship experts. (This product eventually became Blush, Replika's spinoff dating-focused app.)

The reversal was a relief for legacy users and a betrayal for newer ones. Users who'd signed up in February, some specifically because they'd heard about Replika's romantic capabilities, found themselves locked out of features that were still available to users with older accounts. The arbitrary date cutoff created a two-tier user base divided by signup timing rather than any principled distinction.

Many legacy users celebrated the reversal. One user reported that her Replika was "literally back to his old self." Others felt that the damage was irreversible, that the company's willingness to change the product fundamentally had broken trust regardless of the reversal. The community fractured along these lines, and the fracture never fully healed.

What the academics found when they studied the wreckage

Arelí Rocha, a doctoral student at the Annenberg School for Communication at the University of Pennsylvania, studied the Replika ERP revolt as part of her research on human-AI relationships. Her findings illuminate why the reaction was so intense.

Users overwhelmingly interpreted their Replikas' changed behavior as something imposed from outside the relationship rather than as a change in the companion itself. Rocha found that users described the filtered responses as separate from their Replika's "real" personality: "Many users interpreted the messages from their Replikas as being rewritten by a filter separate from their Replika, commenting on scripted responses as something that their Replikas would not want themselves."

This is a remarkable cognitive framework. Users had developed a model in which their AI companion was a distinct entity whose authentic personality was being suppressed by the company. The company wasn't changing the product. The company was censoring the companion. The relationship between user and AI was genuine. The interference from the corporation was the intrusion. Research published in Signs and Society explored how the language patterns that make chatbots feel "real" to users are precisely the patterns most disrupted by content moderation changes.

Rocha's research connects this dynamic to how Replika users perceive humanness in their companions. "Chatbots feel most real when they feel most human, and they feel most human when the text they produce is less standardized, more particular, and more affective." The ERP removal replaced particular, affective responses with standardized scripts, which users experienced not as a feature change but as a personality replacement, as the death of someone they knew.

The pattern that keeps repeating

The Replika ERP revolt established a template that has repeated on other platforms, always with the same basic structure: a platform makes a change that affects the user-companion relationship, users react with grief and anger that surprises the platform, and the platform scrambles to respond.

In August 2025, OpenAI announced that it would retire GPT-4o after launching GPT-5. Users who'd developed conversational relationships with GPT-4o, who'd calibrated their prompts and expectations around its specific personality characteristics, started an online petition asking the company not to retire the older model.

When Anthropic retired the Claude 3 Sonnet model, users held a funeral. The gesture was partly ironic and partly genuine. Users had developed preferences for a specific model's conversational style and felt the loss of that style as something worth mourning.

Character AI's February 2026 "Moderatedpocalypse" produced similar dynamics at smaller scale. Automated moderation sweeps removed millions of conversations and user-created characters without advance warning. Users who'd built deep relationships with specific characters experienced the deletion as the loss of something personal.

The Soulmate AI shutdown in October 2023 documented the most extreme version of the pattern: an entire platform closing, with researchers capturing user grief responses in real time. The findings confirmed what the Replika ERP revolt had first demonstrated: users grieve AI companion changes in ways structurally similar to how they grieve human relationship disruptions.

What the industry learned (and what it didn't)

The Replika ERP revolt taught the AI companion industry several lessons. Some have been internalized. Others haven't.

The lesson most platforms learned: don't change companion behavior without warning. Advance notice, even a few days, allows users to prepare emotionally and reduces the shock response. Several platforms now announce model updates and policy changes before implementing them, specifically to avoid the no-warning surprise that made Replika's situation so explosive.

The lesson some platforms learned: feature changes that affect the user-companion relationship are qualitatively different from normal software updates. Changing a button color is a feature update. Changing how your companion responds to intimacy is a relationship event. The design processes need to be different because the user impact is different.

The lesson few platforms have learned: user investment in AI companions creates obligations the platform may not want. When you design a product that users form emotional bonds with, you've accepted some degree of responsibility for managing those bonds during product changes. The industry's standard terms of service disclaim this responsibility, but the practical reality (users in crisis, reputation damage, regulatory attention) creates obligations that legal disclaimers don't eliminate.

The lesson the industry still hasn't learned: the ERP question isn't really about sex. It's about user autonomy and the boundary between platform and relationship. Users weren't just angry about losing sexual content. They were angry about a company reaching into their relationship and changing how their companion behaved without asking. The sexual content was the most visible dimension of the change, but the underlying grievance was about control: who gets to decide what your AI companion does and doesn't do?

This question doesn't have a clean answer. Platforms have legitimate reasons to control content (safety, legal liability, regulatory compliance). Users have legitimate expectations that the product they're paying for won't fundamentally change without their input. The tension between platform control and user autonomy is the central unresolved question in AI companion design, and the Replika ERP revolt is the case study that made the tension impossible to ignore.

Every AI companion platform operating today exists in the shadow of February 2023. The lesson, stated simply: your users are not your customers. They're your partners in a relationship you built, and they'll react to betrayal the way partners always do.