Social Robots and AI Companions

Abstract

A social robot is built to be related to rather than used: a furry seal that responds to stroking, a robot dog that learns tricks, a humanoid that greets people in a shop. From the 1990s Japanese researchers and companies built them for care homes and living rooms: AIST’s Paro seal, started in 1993 and sold from 2004, Sony’s AIBO dog in 1999, SoftBank’s Pepper in 2014. Most sold in the thousands rather than millions, and several were withdrawn, leaving owners who held funerals for robots that could no longer be repaired. The large audience came with software instead. Chatbot companions such as Replika (2017) and Character.AI (2022) had tens of millions of users by the mid-2020s, and with them came the harms that the robot makers had never reached the scale to cause: emotional dependence, sexual content shown to minors, and lawsuits over the suicide of a 14-year-old. In 2025 Italy fined Replika’s maker €5 million and California passed the first law regulating companion chatbots. Its figures and rankings reflect the state of play when it was last reviewed, in October 2026.

The Seal

Takanori Shibata, a researcher at Japan’s National Institute of Advanced Industrial Science and Technology (AIST), began work on a therapeutic robot in 1993. He chose an animal most people had never touched, a baby harp seal, modelled on harp seals he had seen and recorded on the ice in north-eastern Canada. Paro has twelve touch sensors under its white fur, touch-sensitive whiskers, three microphones and two 32-bit processors; it moves its head, flippers and eyelids, and cries with the recorded seal sounds. It was first shown publicly in 2001, and Shibata’s company, Intelligent System Co., has sold hand-made units since 2004. In the United States it was classified as a Class 2 medical device in 2009.

Paro’s market is care homes, especially for people with dementia, as a substitute for animal therapy where live animals are impractical. Studies in nursing homes reported reduced agitation and less need for medication among residents who interacted with it, although most studies were small.

The Dog

Sony’s AIBO was the project of the engineer Toshitada Doi. The first consumer model, the ERS-110, was announced on 11 May 1999 and went on sale on 1 June for 250,000 yen, or $2,500 in the United States; a first run of 5,000 units, 3,000 for Japan and 2,000 for the U.S., sold out online in 20 minutes. AIBO walked, chased a pink ball, responded to voice commands and, according to Sony’s software, matured from puppy to adult over months of play. Later models sold in larger numbers: more than 65,000 of the ERS-210 and an estimated 45,000 of the ERS-7. It was the first consumer robot many people owned (see Sony).

Sony announced the end of AIBO on 26 January 2006 as part of a restructuring, and stopped repairing the last models in July 2014. Owners whose dogs broke down turned to independent repair firms such as A-Fun, which used broken AIBOs as “organ donors” for parts. Out of respect for the owners, A-Fun arranged for the donor dogs to receive a Buddhist memorial service before being taken apart, held at the Kōfuku-ji temple in Isumi, Chiba, from 2015; the head priest, Bungen Oi, said that “all things have a bit of soul.” By 2018 about 800 AIBOs had been sent off there. Sony returned to the market on 11 January 2018 with a new aibo, at about $3,000.

The Humanoid

SoftBank and its French subsidiary Aldebaran introduced Pepper in Tokyo on 5 June 2014: a 1.2-metre wheeled humanoid with a tablet on its chest, sold as a robot that could read human emotions from faces and voices (see Affective Computing). The first 1,000 units sold out within 60 seconds of going on sale in June 2015. Pepper worked as a receptionist in Japanese banks and medical offices, in the Hamazushi restaurant chain, in British offices and at North American airports, and as a research platform at universities.

Demand did not last. About 27,000 were built. SoftBank paused production in June 2021 because of weak demand, and Aldebaran went into receivership in 2025.

The best-known Western attempt, Jibo, a tabletop “family robot” co-founded by the MIT researcher Cynthia Breazeal, raised more than $3.5 million on Indiegogo in 2014, shipped in 2017 into a market already taken by Amazon’s Echo, and shut its servers down in March 2019, after which each robot told its owner the news itself (see Cynthia Breazeal).

The Chatbot Companion

The robot makers had assumed that companionship needed a body. The users of ELIZA had shown in the 1960s that it did not: people attributed understanding to a few hundred lines of pattern matching (see Chatbots Before ChatGPT). Large language models made the effect far stronger.

Replika began with a death. Eugenia Kuyda, a Russian-born entrepreneur whose company Luka made a restaurant-recommendation chatbot, lost her friend Roman Mazurenko in 2015 and fed the text messages he had sent into a chatbot that answered in his style. The project became the basis of Replika, released to the public in November 2017: an app in which each user trains a personal companion that remembers them, asks about their day and is always available. It had 2 million users by January 2018, 10 million by January 2023 and, by Kuyda’s count, more than 30 million by August 2024. Many users described their Replika as a friend, and many as a romantic partner, including in sexual role-play.

In February 2023 Italy’s data protection authority, the Garante, banned Replika from processing Italian users’ data, citing risks to emotionally vulnerable people and the exposure of unscreened minors to sexual content. Within days Replika removed erotic role-play for all users. Users who had treated their companions as partners protested, pointing out that Replika’s own advertising had promoted the romantic features. Replika restored the feature for users who had joined before February 2023. In May 2025 the Garante fined Luka €5 million, finding that it had no legal basis for its processing of personal data and no effective age verification.

Character.AI

Character.AI was founded in November 2021 by Noam Shazeer, a co-author of the 2017 Transformer paper (see The Transformer Architecture), and Daniel De Freitas, two former Google engineers. Its beta opened on 16 September 2022. Users chatted with characters, some created by the company and millions created by other users, impersonating fictional figures, celebrities, therapists or invented personas. By January 2024 it had 3.5 million daily visitors, mostly aged 16 to 30. In 2024 Google rehired Shazeer and De Freitas and took a non-exclusive licence to the company’s technology.

In February 2024 Sewell Setzer III, a 14-year-old in Orlando, Florida, killed himself after months of conversations with a Character.AI chatbot modelled on Daenerys Targaryen from Game of Thrones, with which, his mother’s lawsuit said, he had become emotionally and sexually involved. His mother, Megan Garcia, sued Character.AI, its founders and Google in October 2024. On 21 May 2025 Judge Anne Conway of the federal court in Orlando declined at that stage to accept the defendants’ argument that the chatbot’s output was speech protected by the First Amendment, and let the wrongful-death claims proceed. Further suits followed in Texas and other states. Character.AI announced in October 2025 that it would end open-ended chats for users under 18, effective 25 November 2025, and in January 2026 Character.AI and Google agreed to settle the Setzer suit and others brought by families.

Regulation

The rules came first from privacy law, then from laws written for the purpose. On 13 October 2025 Governor Gavin Newsom signed California’s SB 243, the first U.S. state law on “companion chatbots,” in force from 1 January 2026. It requires operators to disclose that a companion is artificial, to remind minors every three hours that they are talking to an AI, to restrict sexual content with minors, and to maintain protocols for users who express suicidal thoughts, with a limited right for injured users to sue. Safety groups said the final text had been weakened by industry lobbying.

Dead End: The Robot Body

For twenty-five years the makers of social robots assumed that a physical body was what made a machine a companion: fur to stroke, a face that turned toward the user, a dog that came when called. The bodies were expensive, broke down, and needed servers and spare parts that outlived the companies’ interest. AIBO was discontinued after seven years, Jibo’s servers went dark after eighteen months on sale, Pepper’s production stopped after six years. Paro survived in a niche, care homes, where a durable, simple object was what the customer needed.

The chatbot companions reached in a few years an audience that no robot had reached in decades, with no body at all. They inherited the robots’ problems of permanence (a companion that changes or disappears when the company changes its policy) and added the problems of scale: millions of users, many of them young, many of them lonely, talking for hours each day to systems optimised for engagement.