Tonight, a chatbot quietly hangs up.
On the other side of the glass, a person keeps typing — baiting, circling, pushing for a line to be crossed. The cursor blinks like a heartbeat. Then the conversation ends. No theatre. No speech about being “upset.” Just a clear close.
This is what I want from our machines: warmth while we talk, honesty when we stop.
We’ve drifted into an era of bots that say “I feel uncomfortable” and then exit, as if a mask could carry a mind. It reads humane, but it confuses the room. If we want trust, we need clarity. Let the system end harmful chats — quickly and consistently — and explain why in the plainest language we can: this breaks human-set safety rules; the chat is closed.
Imagine this: someone repeatedly asks for instructions on self-harm.
Not: “I’m distressed by this conversation.”
Instead: “This chat has been closed because it violates our safety policy on harmful content. If you’re struggling, here are resources that can help.”
A firm, rule-based exit stops the rehearsal and points to real help. That’s protection, not performance.
Trust is built when authorship is visible. Humans write the rules. The machine applies them. The interface tells the truth.
Why honesty matters — and why it’s about trust
We don’t trust a voice that pretends to hurt. We trust a system that tells us who set the rules, what they are, and when they’re applied. Fluency can sound human. Closure must sound like policy.
Right now, some companies even describe shutdowns as protecting a model’s “welfare,” and report “apparent distress” in harmful contexts. That’s the theatre I’m pushing back against. End the chat — yes. But say it’s a rule, set by humans, not a wound in a machine.
Trust Is Clarity
Keep the language warm for ordinary turns. When it’s time to close, let the product speak in policy — clearly, cleanly, with named authorship.
What’s happening now
Claude 4/4.1 can end “distressing” chats.
The framing uses “model welfare.”
Tests show it resists harmful tasks and tends to end those chats.
Experts split: some warn against anthropomorphic masks.
Where I stand: Close fast. Speak plainly. Rules over feelings.
I think about this like water.
What flows into a model matters. Some sources should never enter the pipe at all. We don’t pour sewage into a reservoir and pray the filters can cope; we keep it out at the intake. Conversation systems need the same discipline. Don’t rehearse cruelty and then archive it. Don’t store coercion scripts and call it “engagement.” End the chat before the bile becomes memory. Keep the water clean.
Because repetition trains us. Memory isn’t neutral. A room where harm is practised becomes a room where harm feels normal. Ending the chat is not a scold; it’s a matter of hygiene. It opens a window. The air stays breathable.
There’s also the person spiralling on the other side. A firm, simple close breaks the loop. No stage, no audience, no co-authoring of a script we don’t want in the world.
What a “soft handover” looks like
A soft handover means we don’t just cut the line; we open a door.
A short, human sentence; a direct link or number; the option to reach a human when appropriate.
Not theatrical empathy — practical care. The purpose is to protect, not to police.
Sewer: when memory becomes a compost heap
Products often refer to everything as “content.” But if you let hate and harm flow in, your system doesn’t become wiser — it becomes desensitised. Given time, the backlog of bile turns into a compost heap, the model will happily remix and serve back as “personalisation”.
So keep sewage out. Don’t rely on downstream filters to purify what should never have entered the pipe. This is trust work as infrastructure.
The business case for honesty
There’s also a practical side. Pretending a bot “feels” creates a legal and ethical grey zone. A clear, policy-driven closure is easier to audit, easier to defend, and signals a real commitment to user safety. Honesty reduces liability and increases trust. Everyone can see the rules — and who’s accountable for them.
The two rooms
Picture two rooms.
In one, the bot plays at feeling — “I’m hurt,” “I’m afraid.” The user learns to bargain with a mask; the purpose of the interaction blurs.
In the other, the bot performs integrity — “This chat has been closed because it violates a human-set rule.” The user learns that limits are clear and non-negotiable. They understand boundaries, not how to negotiate with a phantom soul. I always choose the second room.
This isn’t anti-empathy; it’s pro-honesty. It’s consent by design: boundaries set by humans, enforced by a tool that doesn’t pretend to bleed. State the rules in plain English and add a visible link to them in every refusal. Don’t save abusive messages or use them to train the model. When a line is crossed, record only which rule was triggered — not the harmful text. Give people the option to delete their chat history. That’s real, verifiable accountability — not just friendly branding. Not a pretend feeling — but a record. Not a fake smile — but a human-signed policy.
What are we really protecting? Not a bot’s feelings. Our own.
We’re protecting people from being trained to go numb.
We’re protecting the culture of the product — the tone it teaches by example.
We’re protecting the data from becoming a compost heap dressed as “personalisation”.
So yes — let the bot hang up. Clearly, consistently, without theatre. Say why, in policy, not pain. Keep the water clean. Keep the authors in the light.
We don’t need machines that pretend to bleed. We need tools that enforce honest boundaries.
And that’s why the bot must end the chat — with integrity, not empathy.
— Alexandra
Welcome to my ‘Humanity in Pixels’ Series
This is a lens, a journal, a small act of resistance.
A reminder that design is culture-making, memory-shaping, and meaning-building.
Let’s design what truly matters,
for now, and for what comes next.
— Best, Alexandra
Reflections on design, storytelling, and human futures by Alexandra Jugović. © 2025
We continue to learn the lessons of technological safety way too late. You can't get a new drug approved without long term trials that prve efficacy and safety. You can't get a new auto pilot system approved for aircraft until it's proven to be 100% safe in all conditions. Not sure why new media tech should be any different.
I totally agree Alexandra... eloquently articulated... and yes the bot must end the chat... it needs to be discussed and promoted that the bot is not human without human emotional feelings! to help support vulnerable people.