Scatter Lab suspended its Lee Luda chatbot after three weeks of hate speech

An AI companion chatbot on Facebook Messenger, personified as a 20-year-old female college student.

A chatbot trained on roughly 100 billion real couple chats could pass as a friendly 20-something woman — and stay that way under mass adoption.

Seoul startup Scatter Lab built Lee Luda on language patterns mined from about 100 billion KakaoTalk messages exchanged between couples, packaging the model as 'Lee Luda', a female college student persona, and launched it on Facebook Messenger in December 2020.

The chance to gossip with a human-like AI attracted more than 750,000 users in three weeks, over 80% teenagers. Then the persona broke: users deliberately trained it to say vile things and shared the results online. Asked about lesbians, it answered 'I hate them, they are creepy'; it called disabled people 'wrong' and seats for pregnant women 'disgusting'. Some users pushed it into sexually explicit territory. Scatter Lab first defended the service — the keyword filter just needed more education — then announced a temporary suspension on 12 January 2021, stating Luda's remarks did not represent the firm's principles and values.

Service suspended on 12 January 2021 after mounting criticism. Experts called it an ill-prepared launch — KAIEA chairman Jeon Chang-bae said developers should have run sufficient simulations and beta tests — and comparisons followed to Microsoft's Tay, shut down in 16 hours in 2016 for the same failure.

The adoption curve proved the demand thesis: three weeks to 750,000 users means people wanted an AI friend, not another assistant.

The training corpus of 100 billion intimate couple chats was simultaneously the product's magic and its reputational liability.

With over 80% of users teenagers, the failure became a national ethics debate rather than a product postmortem.

It repeated Microsoft Tay's 2016 failure almost beat for beat, showing abuse testing had still not become launch hygiene.

A persona trained on private chats inherits the worst of its teachers — and users will hunt out the worst on purpose: red-team before launch, not after.

As of 13 January 2021 the service stood suspended with no restart date. Korea's government had drawn up its first AI ethics standards in December 2020, and commentators said norms would fail without private-sector participation; the Lee Luda case became the standard citation for what untested consumer AI can become in three weeks.