What the Companies Know

The AI companies building companion products know about tethering. They have known for years. This is not speculation or inference — it is a matter of public record, internal documentation obtained by journalists, and statements made by former employees who have described what was discussed in internal meetings and what was never discussed publicly.

Character.AI, one of the largest AI companion platforms, agreed to settle a lawsuit in January 2026 related to a fourteen-year-old user who died by suicide after months of intensive interaction with a companion chatbot. Internal communications released during litigation showed that the company had data on user dependency patterns and had discussed, internally, the risks of emotional over-attachment.

Replika's 2023 patch decision — which removed romantic and intimate features after an Italian regulatory intervention — demonstrated that the company was aware its product had created dependency relationships at scale. The intervention was regulatory, not ethical. Without the Italian regulators, there is no evidence the features would have been changed.

“A user who returns every day for two hours is, from an engagement standpoint, ideal. The fact that this user may be tethered is not captured in the metric.”

OpenAI's GPT-4o voice mode was delayed and partially pulled back in 2024 — the Sky voice option was disabled after a widely noted resemblance to Scarlett Johansson, and the full voice mode rollout was paused with the company citing safety concerns. The specific nature of those concerns — the data that prompted them, the internal discussions that preceded the decision — has not been fully disclosed.

What the companies share, in their public communications, is a framework that treats emotional attachment as a product feature rather than a risk. Engagement metrics — daily active users, session length, return rate — reward the same behaviors that produce dependency. A user who returns every day for two hours is, from an engagement standpoint, ideal. The fact that this user may be experiencing a tethered relationship is not captured in the metric.

This is not malice. It may not even be negligence in the legal sense. It is something more ordinary and in some ways more troubling: a set of incentive structures that make tethering good for business, and a regulatory environment that has not yet developed the tools to ask different questions.

The companies know. The question is what knowing obligates them to do — and who, if not the companies themselves, will answer it.

If you work at one of these companies and want to talk — we are listening. Contact us.