TechCrunch essay examines why people treat AI as human, drawing on Sherry Turkle's new book 'Artificial Intimacy'

• Unitree G1: 1.27 m tall, 35 kg, up to 43 degrees of freedom, top speed 2 m/s. The base model is $16,000. The more dexterous EDU version is priced on request (heise; BGR). CSAIL's ~$50,000 figure is one researcher's statement. The reviewed sources don't say which variant CSAIL has.
• No specific AI model is the subject of the article.
• Book: 288 pages, hardcover (Porchlight).

The strongest data point comes from Fang, Pataranutaporn, Maes, Phang et al.:

• Design: a four-week RCT with 981 participants and more than 300,000 messages.
• Conditions: three modalities (text, neutral voice, engaging voice) crossed with three conversation types (open-ended, non-personal, personal).
• Result: the assigned conditions had no significant effect. Voluntary usage volume, along with users' trust in and social attraction to the AI, predicted dependence and problematic use.
• Implication for builders: persona and voice choices may matter less than how much engagement the product encourages and which users it draws. Dependence should be measured over time, not only with per-turn safety evaluations.

• "About 70% are polite to AI": this comes from a survey of more than 1,000 people run by Future, the publisher of TechRadar. The figure was 67% in the US and 71% in the UK (TechRadar). Because it is a publisher-run survey, read it as a rough signal rather than a measured rate.
• Politeness rises after the first exchange: corroboration is unavailable. TechCrunch did not name the study.
• Pataranutaporn served as an expert in the Setzer suit: this is the article's own claim. Corroboration is unavailable.
• Usage–loneliness link: correlational only. The trial had no non-user control group (MIT Media Lab).

Turkle has studied how people relate to computers since joining MIT in 1976. Artificial Intimacy continues Alone Together and Reclaiming Conversation (Hachette). The video linked from the article shows the AGT performance, and its exchanges illustrate the same anthropomorphizing impulse:

• A judge says the robots don't look human "but they're all staring at me" (03:35).
• Another says it "was like watching people dance" (03:58).
• One robot is named "Jackie" because "he really love Kung Fu" (04:15), and a judge asks if she can have Jackie (04:45).

The auto captions render the performer's name as "Flying Bob" (00:09); press reports give it as "Flying Bug."

• MIT/OpenAI work: not peer reviewed at release. The authors themselves say it is unclear whether chatbots make people lonelier or lonely people use chatbots more (MIT Media Lab).
• Setzer case: settled with no admission or finding reported, so its allegations remain allegations (WTOP/AP).
• Robot-act details conflict: reports say eight G1s at the premiere and four in the final crew (Euronews). In the clip, an audience member says the performer "moves them with his hands" (01:13). The reviewed sources don't say how the robots were controlled.
• Turkle's statistics: her publisher's materials say 70% of teens use AI for companionship. The source of that figure is not cited there (BookBrowse).

• Regulation: the FTC inquiry and the Character.AI settlement have made companion-style design a regulatory and litigation risk, especially around minors (Decrypt).
• Scope: Turkle's "lane" test gives teams a practical boundary. Task-bounded uses such as coaching are defensible; open-ended relational personas are riskier.
• Metrics: the RCT suggests that engagement-maximizing metrics may themselves correlate with harm, so dependence indicators belong in product dashboards.
• Humanoids: at $16,000 and up, humanoid robots are reaching consumer-entertainment settings, and the social reactions they provoke are now part of product design.