Where AI companions came from

The order things happened in

Two things stand out from that list before you read a word of analysis.

The first is how compressed it is. From the first mainstream product to a market with a dozen competitors took about eighteen months. The second is that the regulatory story starts almost at the beginning — Italy acted in February 2023, before most of the current products existed. The rules did not arrive late because nobody noticed. They arrived while the thing being regulated was still being built.

Before there was a category

Replika is the outlier and the ancestor, and it is worth understanding why it does not resemble anything that came after.

Replika gives you one character. Not a library, not a catalogue — one, which remembers your name, your relationships, your interests, your routine, and builds a continuing relationship on that. It positions itself as a friend rather than a roleplay platform. Everything that followed did the opposite: a browsable library of thousands of characters, most of them made by other users, none of them yours in any lasting sense.

That difference explains the Italian case. When a regulator looked at Replika in early 2023, it was looking at a product built around sustained personal disclosure, with no age verification at any point and a privacy policy published only in English in a country where the service was being sold. The five-million-euro fine landed in May 2025, listing those failures plainly. Italian users lost the product overnight and had done nothing wrong — the first instance in this category of a service disappearing on its users, and not the last. A library-of-characters product presents a different risk profile — still a real one, but different, and the law was written for the first shape before the second became dominant.

The Character AI moment

The public beta on 16 September 2022 is the closest thing this category has to a starting gun.

What it proved was narrow and enormous. People would talk to a text character for hours, repeatedly, without being paid or prompted to. The founders had come from Google, and the product was free, unpolished and had no obvious business model. It did not matter. Usage arrived anyway.

Every product on the list above is downstream of that demonstration. Not because they copied the interface — several are barely comparable — but because Character AI removed the question of whether anyone wanted this. That is why the 2023 launches cluster the way they do: three of them inside a year, all with roughly the same shape, all built by people who had watched the same thing happen.

Character AI also set the terms people still use to compare products. It is why, three years later, a whole genre of comparison exists around where to go when it stops working for you — the category's reference point is a product most of its users no longer use.

What the 2023 wave actually built

The three launches of 2023 look similar from outside and split on a single decision: who writes the characters.

CrushOn and SpicyChat went the library route. Users make characters, publish them, and the platform's value is the catalogue. SpicyChat launched with around ten people and a million dollars in seed funding, which tells you how cheap this was to start when the model was somebody else's.

Candy AI went the opposite way. Curated characters, image generation as the headline feature, and a paid product from the beginning rather than a free one with monetisation bolted on later.

Both bets worked, and they produced the split that still defines the market. If you want a library of thousands of characters written by other people, one set of products serves you. If you want consistent images of one character and are willing to pay per generation, a different set does. Very few do both well, which is why our ranking of The Best AI Girlfriend Apps of 2026 puts them in an order that changes depending on which question you are asking.

The second wave was about media, not chat

Look at the 2024 and 2025 entries and the pattern changes.

Infatuated, PolyBuzz in its new form, OurDream — these did not launch to prove people would chat. That was settled. They launched into a market where chat was assumed and the competition had moved to what surrounds it: image consistency, voice, video length, how many generations you get before paying.

This is when the second currency arrived, and it arrived nearly everywhere at once. Tokens, coins, credits — different names for the same structure, in which the subscription buys text and everything else is metered on top. It is the single biggest change in how these products are priced, and it happened in about a year. We have a separate piece on what “unlimited” means once a second currency exists, because the answer is rarely the one on the pricing page.

The other thing the second wave brought was corporate distance. Compare the structures:

That third row is not a typo. PolyBuzz's terms of service name a Hong Kong company; the app stores and press releases name a Delaware one. Which company you have a contract with depends on which document you read — and the applicable law is Delaware, with arbitration in New York. We read both documents for exactly this reason; they disagree more often than you would expect.

None of this was true of Character AI in 2022, which was a straightforward American startup with an address you could look up. The category grew into offshore structures the way most consumer internet businesses do, and users mostly did not notice because nobody reads the footer until something goes wrong.

Why the rules arrived when they did

The regulatory sequence makes more sense once you have the timeline in front of you.

Italy moved in February 2023 against the oldest product, built on the oldest shape. Everything else waited — not because regulators were slow, but because the products that would attract attention had barely launched. The US legal reckoning began with a lawsuit filed in October 2024, two years after Character AI's beta, and produced its defining ruling in May 2025. We wrote about that case and what it decided separately, because the reasoning matters more than the headlines suggested.

By the time California's SB 243 took effect on 1 January 2026 — the first US law written specifically for companion chatbots — the category was three years old and had already changed shape twice.

That gap explains a lot of what users experience as arbitrary. A product designed in 2023 for an unregulated market, sold in 2024 to an audience that grew faster than its safety team, and then required in 2026 to verify ages and disclose that it is not a person, will feel like it is changing the deal. It is. The alternative was not changing and being fined, which is what happened to the one company that got there first.

What survived

Of the products on that timeline, the striking thing is how many are still running. This is not a category littered with corpses — not yet. The failures so far have been features rather than companies: retired models, closed age brackets, regions withdrawn from.

What has not survived is the original arrangement, where a free product with no verification and no second currency would let you talk to anything for as long as you liked. That version lasted roughly from September 2022 to somewhere in 2024, and everyone who describes these apps as they were then is describing a market that no longer exists.

If you are arriving now, the practical consequence is that age policy and pricing structure are the two things that changed most since the reviews you may have read were written. Both are checkable in about fifteen minutes, and both are in the terms rather than on the marketing pages. Our free tier comparison tracks the current limits rather than the launch-era ones, because they move.

The package name still says SpeakMaster. Everything else has been replaced at least once.