Korea's 42.5C heat wave lifted AI care escalations 30%. That is not substitution.

NHN Wapl, NHN's senior care unit, said on 16 August that care workers on its AI Living Support platform made a daily average of 46.7 welfare phone calls during the third week of July, 30% above June's average of 35.8, as South Korea recorded an all-time high temperature of 42.5C ^[1][2][3]. That is a labour number dressed as a weather number, and it points the opposite way from the usual claim made for AI care.

Read the architecture before reading the growth. Smartphone sensors measure a user's activity; if no movement is detected the system sends an alert message; if there is still no response, a human care worker telephones ^[5]. Reaching the phone-call stage is by design the exception, and NHN Wapl says it means the warning signs were pronounced ^[6]. So the 30% rise is not a count of things the AI handled. It is a count of things the AI handed off. Ten to eleven additional human calls a day per the reported averages ^[13], with the first ten days of August running 27.7% above June ^[14].

The mechanism is legible. Average daily steps among Wapl users were 4,760 in the week of 7-13 June and fell to 3,200 in the first days of August, a decline of roughly 33% ^[7][15], across the period when Korea's first-ever major heat wave alert took effect ^[7]. Less movement trips more no-movement alerts. The Korea Disease Control and Prevention Agency estimates that mortality risk for people aged 65 and over rises 19% when perceived temperature reaches 38C ^[8], so the sensitivity is pointed at a real hazard rather than noise. Sedaily reports AI elderly care services detected up to five times as many warning signs this summer as a year earlier ^[4].

Naver Cloud's Clova CareCall, which notifies local governments when its AI phone calls surface symptoms, reported alerts for fainting up 417% year on year in July, high fever up 237%, dizziness up 144% and shortness of breath up 112% ^[9]. Those are percentages without denominators. A 417% rise on a small base is a different operational fact from the same rise on a large one, and the base was not published.

The substitution case rests on softer ground. A March survey of organisations running Clova CareCall, conducted by Yonsei University's Center for ESG and Business Ethics, found the rate of lonely deaths 44% lower in adopting areas than non-adopting areas, and emergency room visits down 9.2% ^[10]. Adopters are self-selected; that is a comparison between places that bought the service and places that did not, not a trial. And the one rescue described from this heat wave went the labour-intensive way: on 10 July in Yangpyeong County, Gyeonggi Province, Wapl's system flagged an older adult, the call went unanswered, and a monitor entered the home with police assistance to find the person collapsed and unconscious ^[11].

What AI replaced here is the search, not the care. That is still worth money against the Korea Development Institute's April estimate of a roughly 990,000-worker shortfall in long-term care by 2043 ^[12], and analysts quoted by Sedaily argue the services matter more as climate extremes recur and staffing thins [c13a]. But a triage layer that raises escalation volume 30% in a heat week has a ceiling set by the humans behind it.

Watch for absolute alert counts and false-positive rates rather than year-on-year multiples, and for the escalation rate per user: if alerts rise while handoffs per worker stay flat, the software is filtering. If both rise together, local governments are buying demand they have to staff.