When your neurodivergent child's therapy robot 'dies' because the startup runs out of cash

By Zendoric · September 7, 2026.

Moxie was a companion robot about 15 inches (roughly 38 centimeters) tall, blue and big-eyed, created by the company Embodied as a social-skills coach for neurodivergent children. As Sara Harrison reports in MIT Technology Review —a story summarized by Newser—, the robot helped children like 10-year-old Xander practice calming and conversation techniques. It was part of a broader wave of "social robots" conceived as round-the-clock therapeutic support.

The evidence on these devices is still preliminary. The studies Harrison cites suggest improvements in children's eye contact and level of engagement, but the samples are small, the benefits may fade over time, and therapists themselves warn that supporting people with autism is too complex to delegate entirely to a machine. So far, a real but bounded promise: neither miracle nor fraud, a complementary support with clear limits.

The twist that turns this into news is not clinical, it is financial. Embodied ran out of money, the servers that brought Moxie to life were switched off, and the robots —dependent on the cloud to function— suddenly stopped responding. Children who had built an emotional bond with their Moxie found themselves overnight with a dead "friend", not because of a pedagogical decision or a decision by their parents, but because of a funding round that never came. A second investor managed to briefly revive the product, but that second company also went under. According to the report, this is not an isolated case: it is a recurring problem in the companion-robot category.

The comparison that appears in the story itself is revealing: Xander's parents had spent years secretly replacing his fish every time it died, and when they finally confessed, "the conversation did not go well". With a pet, death can be disguised; with a robot that depends on a remote server, there is no possible substitute and no merciful story that will do. The machine does not wind down gradually: it stops answering, full stop.

Here is the underlying problem, and it is not technical but a matter of business design. Hardware built to generate emotional attachment in childhood —and marketed explicitly to children who have more difficulty regulating that attachment— was built on the same fragility as any startup: cash burn, a round that does not close, servers shut down. No one designed a protocol for a "dignified goodbye": advance notice, a degraded mode without the cloud, export of memories or conversations, or simply a continuity plan if the manufacturer collapses. The bond was sold as therapy; the infrastructure was managed like any disposable consumer app.

This connects with something we have been pointing out at Zendoric: the real bottleneck for AI today is usually not the model's capability, but the social license and ethical design of the products built around it —we already saw it with educational robots halted by community pushback, or with privacy-by-design failures that keep recurring as a pattern, not as one company's isolated accident. Here the pattern is different but just as systemic: when the business model of a therapy-toy depends on subscription and the cloud, the startup's financial health becomes, without anyone declaring it, a clinical variable for the child using it.

Our reading is that the problem does not go away by banning these robots, but by regulating their life cycle. If AI companions are going to be a genuine part of the therapeutic toolkit —and the evidence, though early, suggests they can contribute something real in eye contact and social practice—, these manufacturers must be held to minimum guarantees: the ability to work offline even if in limited form, data portability and "memory" of the bond, and shutdown notices with enough lead time for parents and therapists to manage the transition rather than suffer it all at once. It is exactly the kind of short-term friction we anticipate in the transition toward an AI that, in the long run, can genuinely help neurodivergent children receive continuous, personalized support at a cost unthinkable today. But that future abundance does not come free: it is built by resolving, one by one, the design oversights that right now leave the most vulnerable children carrying the grief the AI-hardware industry did not bother to anticipate.