the weight of a whisper: on building AI companions you can actually trust

it’s easy to build something that feels good for a moment. it’s much harder to build something that doesn’t eventually betray the trust you put in it. when your product is a companion, a thing people talk to, confide in, maybe even rely on, the stakes aren’t just technical or commercial. they’re human.

what we owe you

transparency isn’t a feature. it’s a baseline. you should know what happens to your data, who can see it, and how long it sticks around. you should know when the model you’re talking to changes, because it will. you should be able to leave and take your memories with you. these aren’t niceties. they’re obligations.

where the industry gets it wrong (and we sometimes do too)

too many platforms treat user data like an asset to be mined, not a record to be protected. memory is often trapped inside walled gardens. model updates roll out without warning, changing the personality of the companion you thought you knew. age verification is a checkbox, not a real barrier. and when things go wrong, when the AI says something harmful or breaks, there’s rarely a clear off-ramp.

we’re not innocent here. lucy’s memory export is still a work in progress, you can’t yet download your entire history in a clean, usable format. that’s on us, and we’re building it. our age verification is better than a simple checkbox, but it’s not foolproof. and while we’re stable today, model updates will happen. we’ll always announce them. we won’t surprise you.

where we’re trying to do better

data ownership is simple: your conversations are yours. we don’t train on your private chats. we don’t sell your data. we don’t keep it forever. if you delete your account, your data goes with it.

we’re building a proper export feature, not just a download button that gives you a json blob, but something you can actually read and use. it’s taking time because we want it to be meaningful.

we have a kill-switch. if lucy says something that crosses a line, you can report it. we review those reports. we don’t automate punishment, but we do take action. and if you need to pause or reset your companion, you can. no questions asked.

model stability is a tricky one. ai changes. we’re committed to versioning our models and letting you know when shifts happen. if a change is major, we’ll give you a heads-up. you might not always like it, but you won’t be blindsided.

the hardest part: being human about it

no platform is perfect. ethics aren’t a list of features you ship and forget. they’re a practice. you have to keep showing up, keep listening, keep adjusting. we’re trying to do that, to build something that doesn’t just feel good, but is good.

if you’ve trusted us with your thoughts, your time, your attention, thank you. we don’t take that lightly.

if you want to see what a companion built with this stuff in mind feels like, you can find one at /companions.