Replika's founder built a private grief bot a year before it

The design

Before Replika existed as a product, its founder built a private chatbot for a smaller circle of grief. Eugenia Kuyda, co-founder of the AI startup Luka, describes the origin herself in a Bloomberg Businessweek interview: her close friend Roman Mazurenko 'was killed in an accident, walking across the street,' and months later she 'plugged in all the texts that we sent each other' to produce a bot that could 'text with me just the way Roman would have.' A longer account in The Verge, reported independently by Casey Newton, dates Mazurenko's death to 28 November 2015, a car accident in Moscow, and Kuyda's public Facebook announcement of the finished bot to 24 May.

What the evidence says

Newton's reporting adds detail Kuyda's own account does not: ten of Mazurenko's friends and family contributed more than 8,000 lines of text to train a Russian-language neural network built on Luka's existing restaurant-recommendation technology, and Kuyda 'edited out a handful of messages she believed would be too personal to share broadly' from her own contribution before training began. The two sources also diverge in emphasis: Kuyda's own words frame the project as reconstruction, 'a simulacrum to bring her best friend back to life,' while Newton records a friend, Vasily Esmanov, publicly calling the released bot 'half-baked' and warning it repeated a mistake he attributed to a television drama about grief.

What it asks of people

The project asked ten grieving friends and family members to hand over years of private correspondence with someone who had just died, for a purpose Newton reports one contributor, Sergey Fayfer, called uncertain in advance: 'the question wasn't about the technical possibility. It was: how is it going to feel emotionally.' It then asked the wider public, once released on Facebook, to decide whether to engage a simulation of a friend at all; Newton reports several people who declined.

Privacy and safeguards

Kuyda's own editing of her contributed messages before training is the only stated safeguard either source describes, a personal judgment call rather than any policy applied to the other nine contributors' texts. Neither source states what retention or deletion practice, if any, governed the contributed messages once training was complete.

• What safeguard, beyond one person's own editorial judgment, should govern text contributed by grieving family members for a project like this?
• How does Kuyda's first-person account compare to what Newton's independent reporting adds or complicates?
• What does this project's direct line to Replika suggest about consent practices in the commercial product that followed?

Two accounts, one from Kuyda and one independently reported, agree on the bot's origin in grief and diverge on how comfortably it should be judged.