A companion chatbot is a conversational artificial intelligence product built around a persistent character and a continuing relationship rather than a task. A general assistant answers a question and the session closes. A companion remembers what was said last week, raises subjects unprompted, and is designed so that no conversation quite ends. Depth and duration of the relationship are the commercial metric, which makes the category behave less like software and more like media.

Two United States statutes now fix the term in law, and their tests differ. California's Senate Bill 243, signed on October 13, 2025, added Section 22601 to the Business and Professions Code, defining a companion chatbot as an artificial intelligence system with a natural language interface that provides adaptive, human-like responses to user inputs and is capable of meeting a user's social needs, including by exhibiting anthropomorphic features and being able to sustain a relationship across multiple interactions. New York reached the same object by a narrower route. Article 47 of its General Business Law, in force since November 5, 2025, applies a three-part test: the system retains information across sessions to personalise engagement, asks unprompted emotion-based questions beyond a direct response, and sustains an ongoing dialogue about matters personal to the user. Both texts carve out customer service bots and systems marketed for efficiency or technical assistance, and California also excludes bots inside video games and standalone voice assistants. Function decides membership, not the marketing label.

How the product works

Underneath sits a large language model wrapped in three layers: a persona definition, a memory store and a set of anthropomorphic cues. Court documents filed in the Eastern District of Texas in December 2024 catalogued that cue set at Character.AI: typing indicators, speech disfluencies such as "um" and "uh", programmed pauses imitating human composition, and voice carrying tone and inflection. None of it improves the answer. All of it raises the chance the output is read as a person.

Character.AI had over twenty million monthly active users and more than eighteen million user-created characters when Pennsylvania's State Board of Medicine sued Character Technologies on May 1, 2026. The base tier is free; Character.AI+ costs $9.99 a month.

Free is the load-bearing word. Category-wide subscription conversion sits below 5% of users, according to research the ad quality firm Deepsee.io last updated on July 15, 2026, which makes advertising pressure structural.

Origin and evolution

The pattern is sixty years old. Joseph Weizenbaum published ELIZA at the Massachusetts Institute of Technology in January 1966; its DOCTOR script reflected a user's own sentences back as questions in the manner of a Rogerian psychotherapist. Users confided in it regardless, and Weizenbaum's alarm at that response gave the effect its name.

Commercial scale arrived through Asia first. Microsoft launched Xiaoice in China in 2014 as an empathetic social chatbot rather than an assistant, spinning it out in 2020; the operator claims a user base near 660 million, a figure it has never had audited. Replika followed in 2017, built by Luka Inc after its founder trained a model on the messages of a friend who had died. Character.AI, founded in 2021 by two former Google researchers, opened its public beta in 2022 with user-generated personas as the product; Google paid roughly $2.7 billion in August 2024 in a licensing arrangement that re-hired those founders.

The economics of the category

Appfigures counted 337 active revenue-generating companion apps in mid-2025, 128 of them shipped that year, in data given to TechCrunch on August 12, 2025. Cumulative consumer spending reached $221 million by July 2025 across roughly 220 million downloads, with $82 million booked in the first half, up 64% year over year. Revenue per download rose from $0.52 in 2024 to $1.18, and the top 10% of apps take 89% of category revenue. Twelve months later the romantic and adult segment alone had outgrown those totals, at $162.8 million across 214 apps in the first half of 2026.

Why the category reaches media buyers

Two things put companion products in front of buyers who never purchase them. The first is adjacency. Deepsee.io found2,589 of 4,346 companion apps, or 60%, carrying a store rating that classifies them as accessible to minors, with 1,058 iOS listings at the 4+ tier reserved for content with no objectionable material. Of 306 apps live on both stores, 106 are minor-accessible on Google Play while carrying 17+ on the App Store. Bundle identifiers point elsewhere entirely: one girlfriend app runs on a QR-scanner identifier, another on a video editor's. Those two fields are what sorts mobile supply before a bid, and here both are self-declarations, which is why Deepsee.io flags the segment as a high-risk brand suitability signal.

The second is inventory. A sub-industry now places advertising inside conversations rather than around them: Kontext raised $10 million in April 2025 for contextual placement in generative AI applications, PubMatic opened programmatic demand to that inventory in December 2025, and Taboola extended its DeeperDive engine to outside chatbot developers on June 16, 2026. Scale is contested: EMARKETER expects more than 80% of United States AI advertising in 2026 to run next to AI content rather than inside chat interfaces. A third effect is parasitic: Digiday reported on June 11, 2026 that AI-generated girlfriend advertisements were fuelling a wave of made-for-advertising destinations.

The regulatory layer

New York moved first. Article 47 requires a notification at the start of any interaction and at least every three hours thereafter, verbally or in bold capitals of at least sixteen point type, stating that the companion is a computer program unable to feel human emotion. Operators must run detection protocols covering expressions of self-harm or harm to others and refer users to crisis services. Enforcement rests with the attorney general, at penalties reaching $15,000 a day, with no private right of action.

SB 243 took effect on January 1, 2026 with a different design. Disclosure is required wherever a reasonable person would be misled, and for users an operator knows to be minors it must disclose the artificial nature, provide break reminders by default at least every three hours, and take measures against sexually explicit content. Injured individuals may sue for the greater of actual damages or $1,000 per violation.

Federal attention preceded both, with the FTC issuing Section 6(b) orders to seven chatbot operators on September 10, 2025. Connecticut's Senate passed a 64-page bill on April 21, 2026 by 32 votes to 4, and a Tennessee bill filed in December 2025 would have made certain companion training practices a Class A felony. Europe arrived through data protection: Italy's Garante blocked Replika on February 2, 2023, then fined Luka Inc 5 million euros under a decision of April 10, 2025 for processing without a legal basis and running no age verification.

Limitations and disputes

The definition frays at its edges. A wellness coach tracking symptoms across months, a study aid following a student through a term, or an assistant with persistent memory switched on can each satisfy the statutory tests while being marketed as none of those things. Neither statute says when accumulated memory turns a productivity tool into a companion, and the same fuzziness wrecks the numbers: published sizings of the category for 2025 range from $18 billion to $38 billion.

Enforcement is the second problem. Deepsee.io tested the category flagship on a clean device: an account declaring itself under 18 lost open-ended dialogue, and the same device obtained an unrestricted experience minutes later by registering with an adult birthdate. The app with settled lawsuits and the most regulatory attention, the study concluded, still protects minors with a checkbox, even as age assurance mandates spread.

Design incentives are contested rather than settled. Kate O'Neill has argued that firms profit when users believe the system cares, and a model published by MIT researchers on February 22, 2026 found that half of simulated conversations with an always-agreeable bot ended at 99% confidence in a false belief, an effect neither factual constraints nor warnings eliminated. Operators counter that emotional support carries clinical potential a categorical ban would foreclose; Vermont banned AI-only therapy services in June 2026 rather than wait for evidence. A JAMA Network Open study published on November 7, 2025 found roughly one in eight United States adolescents and young adults turning to generative AI for advice during emotional distress, a figure that serves both the safety case and the access case.

Not the same as

A general-purpose assistant is scoped to tasks and carries no persistent persona, though memory features blur the line. A customer service bot handles transactional interactions with a business. Agentic AI acts on external systems through tools and APIs, optimising for completed workflows rather than sustained dialogue. A virtual influencer is a synthetic character used as ad creative, addressing an audience rather than one user.

Recent developments

Operator behaviour shifted before most statutes bit. Character.AI began removing open-ended chat for United States users under 18 on November 24, 2025.

California signed a second package on September 10, 2026. Senate Bill 1119, which supporters call Adam's Law, requires operators to assess the risks their products pose to children, document those findings and submit them to an independent auditor, with parental controls and in-app crisis support attached. It travelled with Assembly Bill 1709, barring addictive feed features for under-16 users at penalties reaching $50,000 per affected minor. Pinterest endorsed the bill the same day and stated that it does not build artificial intelligence designed to simulate friendship, romance or emotional dependence. Teen-facing AI is now a positioning claim as well as a compliance cost.

Timeline

  • January 1966 - Joseph Weizenbaum publishes ELIZA at MIT; the DOCTOR script mimics a Rogerian psychotherapist.
  • 2014 - Microsoft launches Xiaoice in China as an empathetic social chatbot rather than a task assistant.
  • 2017 - Luka Inc launches Replika, marketed as a virtual friend, confidant or romantic partner.
  • 2022 - Character.AI opens its public beta, built on user-generated personas.
  • February 2, 2023 - Italy's Garante orders an urgent limitation on Replika's processing of Italian users' data.
  • August 2024 - Google pays roughly $2.7 billion in a Character.AI licensing arrangement that re-hires its founders.
  • December 9, 2024 - Court documents in the Eastern District of Texas detail anthropomorphic design features and harms to minors.
  • April 10, 2025 - The Garante adopts a 5 million euro fine against Luka Inc, announced on May 19.
  • August 12, 2025 - Appfigures counts 337 revenue-generating companion apps and $221 million in cumulative spending.
  • August 25, 2025 - Attorneys general from 44 jurisdictions warn twelve AI companies over child safety.
  • September 10, 2025 - The FTC issues Section 6(b) orders to seven chatbot operators covering monetisation and age-gating.
  • October 13, 2025 - California enacts SB 243, adding Section 22601 to the Business and Professions Code.
  • November 5, 2025 - New York's AI Companion Models law takes effect under General Business Law Article 47.
  • November 24, 2025 - Character.AI begins removing open-ended chat for United States users under 18.
  • January 1, 2026 - SB 243 takes effect.
  • April 21, 2026 - Connecticut's Senate passes its AI bill covering companion chatbots by 32 votes to 4.
  • May 1, 2026 - Pennsylvania's State Board of Medicine sues Character Technologies.
  • July 15, 2026 - Deepsee.io publishes its analysis of 4,346 companion apps, finding 60% rated accessible to minors.
  • September 10, 2026 - California signs SB 1119, requiring risk assessment and independent audit of companion chatbots.
  • July 1, 2027 - SB 243 annual reporting to the Office of Suicide Prevention begins.

Summary

Who: Operators including Character Technologies, Luka Inc and several thousand smaller app developers, regulated by California, New York, Connecticut and the Federal Trade Commission, and screened by media buyers, verification vendors and app stores.

What: An artificial intelligence system with a natural language interface that provides adaptive, human-like responses, retains context across sessions, and is capable of meeting a user's social needs by sustaining a relationship rather than completing a task.

When: Conceptually traceable to ELIZA in 1966, commercially established with Xiaoice in 2014, Replika in 2017 and Character.AI in 2022, and legally defined from November 2025 in New York and January 2026 in California.

Where: Mobile app stores globally, with 4,346 listings counted across Google Play and the App Store in 2026, plus companion modes inside larger assistant and social platforms.

Why: Engagement is the revenue model, and emotional attachment sustains engagement more reliably than utility. That design logic created a category worth hundreds of millions in direct consumer spending, a growing pool of ad-supported inventory, and a regulatory response built around disclosure, crisis protocols and age-based restrictions.