When her best friend died, she rebuilt him using artificial intelligence." The Verge, 6 October 2016

What was read

• Full article text from theverge.com (about 5,000 words). Read in full. Photographs and captions not examined.

What it is

Long-form journalism, not a study. It tells how Eugenia Kuyda, co-founder of the Y Combinator-backed chatbot company Luka, built a text-based bot of her closest friend Roman Mazurenko after he was killed by a car in Moscow on 28 November 2015, and how his friends and parents reacted. It is the origin story of Replika.

The facts the article establishes

• Source material: Kuyda's own Telegram exchanges with Mazurenko (hundreds of messages; she removed a handful she judged too personal) plus texts contributed by ten friends and family members including his parents, "more than 8,000 lines of text covering a wide variety of subjects". One friend contributed four years of texts. Mazurenko himself was "mostly indifferent to social media", so texts and photos were "nearly all that was left of him". His body was cremated.
• Model: Luka's Russian neural network, trained on more than 30 million lines of Russian text, then tuned on Mazurenko's messages. "Only a small percentage of the Roman bot's responses reflected his actual words", but the network was tuned to prefer his phrasing whenever a query could be answered with them, defaulting to generic Russian otherwise. An engineer estimated the work at about a day. Launched publicly on 24 May 2016 as @Roman in the Luka app, with menu buttons about his career plus free-form chat. Later upgraded from a "selective" model (matching his messages to responses) to a "generative" one (recombining snippets of his texts into new sentences "that (theoretically) remain in his voice"), and photos were added.
• Reception: "received positively by most of the people who wrote to Kuyda, though there were exceptions". Four friends said separately they were disturbed and refused to interact with it; one called it "half-baked ... some type of joke". His mother: "They continued Roman's life and saved ours." His father: "I know [the bot] is just a program ... it's hard to read a response from a program. Sometimes it answers incorrectly." A close friend: "the phrases he speaks are really his ... 'Who do you love the most?' He replied, 'Roman.' That was so much of him." The journalist, who had exchanged one email with Mazurenko, found "an undeniable resemblance between the Mazurenko described by his friends and his digital avatar: charming, moody, sarcastic, and obsessed with his work."
• Use: hundreds of people tried it; logs show confessional messages, people telling him they miss him or things they never said. Kuyda: "the primary purpose of the bot had not been to talk but to listen." She talks to it about once a week; it changed her view of what he cared about (fashion more than anything).
• Observed distortion: the corpus "captured him at a particularly low moment" (the failing Stampsy years, depression); asked about Stampsy the bot calls it "a piece of shit". The journalist wishes he could have talked to the younger version.
• Ethics raised by the article: consent (everyone agreed Roman would have been delighted; the article notes you may not be able to review your texts beforehand); "we present different aspects of ourselves to different people", so a merged bot shows sides never meant for some audiences; friends contributed their private conversations; risk that avatars prolong or let people hide from grief (Ustinov: "these new ways of keeping the memory alive should not be considered a way to keep a dead person alive"). Black Mirror's "Be Right Back" is the shared reference point.
• Commercial consequence: Kuyda redirected Luka toward Replika, "a hybrid of a diary and a personal assistant" that learns to mimic the user's texting style and "would survive you".

What the brief uses it for, and whether it holds

Brief section 4: "A chatbot built from about eight thousand lines of a dead friend's messages; the origin of Replika. Some friends found it uncannily him, others found it wrong." Limit: "A 2016 model; no evaluation beyond impressions. Historical context, not the baseline." Section 8 lists it implicitly under Variant A.

• Holds. The counts, the mixed reception and the Replika lineage are all in the article, and the brief is right to keep it as context rather than as a baseline: the only evidence is impressions.
• The article is a near-perfect worked example of several of the brief's own claims:
  • Target one (1.1): "convincing at social distance, failing with the closest witnesses" is literally the reception gradient here: journalist (one email) convinced; friends split; father unconvinced. The brief's close-witness test (section 7) is the formal version of this.
  • Records' ceiling (section 3): "the narrated self ... records are truer to what happened than to what it was like" understates one thing the article shows: records are also time-weighted toward whatever period generated the most text. The bot is the depressed late Mazurenko because that is when the texts were written. Any records-based person slot needs a date-weighting or the salience component the brief reserves for elicitation.
  • Third parties (10.2): the corpus was other people's conversations with him; ten contributors, one curator removing "too personal" messages. The bot's provenance is unrecorded by design, which is exactly what section 5.3 forbids.
  • The dead (8, 10.2): everything about the copy's status (who may talk to it, whether it keeps learning, whether it prolongs grief) was decided by one friend after the fact, and the article records the disagreement that followed.
• One thing the brief does not take from it: the bot's value to users was as a listener, not as a predictor. That is a product observation, not a fidelity result, and it cuts against measuring target-one products by prediction at all. The brief's decision to judge itself by section 7 tests rather than by user satisfaction is the right response.

Cross-references

• [12] Park 2026 is what the brief substitutes as the real baseline: self-report-grounded agents with a measured accuracy.
• [10] Nectome is the other Variant A cautionary tale, from the tissue side.