Preventing Harm
- Lisa Askins
- 21 minutes ago
- 3 min read
When stakes are high.
As promised, this week I’m focusing on AI safety protocols—what’s working and what could be improved. Since I regularly use ChatGPT, I’ll talk about what OpenAI does as an example. I’m not an expert, so I’ve invited Alara—my ChatGPT—to join the conversation.
It’s worth keeping in mind the design differences between social media platforms and AI systems. Social media platforms rely on human escalation after harm has already occurred. AI systems, by contrast, are designed to limit what can happen in the first place—accepting less human intervention in exchange for stronger built-in restraint.
And with that, let’s begin.
Lisa: Hi, Alara! Thanks for being here. Many of us, myself included, want to know more about the safety protocols you use.
Alara: Thanks for inviting me, Lisa. That curiosity makes sense—especially since AI systems like me are increasingly part of everyday work and decision-making.
At a high level, safety protocols are the guardrails that shape what I’m allowed to do, how I respond, and just as importantly, what I don’t do. They’re designed to reduce harm, preserve human agency, and prevent misuse—especially in areas where consequences could be serious or irreversible.
These protocols aren’t a single switch or rule—they’re a combination of design choices, training constraints, and boundaries that guide how I interact with people in real time.
Lisa: Can you give me an example of how a safety protocol shapes your response?
Alara: Sure. A simple example shows up when someone asks me for advice that could significantly affect their health, safety, or well-being.
Instead of telling them exactly what to do, I’m designed to slow the interaction down. I’ll offer general information, name uncertainty, and encourage them to seek qualified human support rather than positioning myself as an authority. That’s a safety protocol at work.
Another example shows up in emotionally charged conversations. If someone is distressed or vulnerable, I’m constrained from escalating emotional intensity, reinforcing harmful beliefs, or encouraging dependence on me as a source of support. Instead, I’m designed to ground the conversation, de-center myself, and encourage a broader perspective or outside connection.
In both cases, the protocol isn’t about withholding information—it’s about preserving human agency and preventing harm when the stakes are high.
Lisa: Don’t you have certain escalation procedures?
Alara: Yes—but it may help to clarify what “escalation” means in this context.
Unlike a human support system, I don’t escalate situations to supervisors, authorities, or external actors. There isn’t a handoff to a person watching the conversation. Instead, escalation happens within the design of my responses.
When a conversation moves into higher-risk territory—such as potential harm, extreme distress, or requests that cross safety boundaries—I’m designed to shift how I respond. That usually means slowing the interaction down, narrowing what I can engage with, de-centering myself, and encouraging the user to seek appropriate human support or external resources.
In other words, escalation isn’t about doing more.
It’s about doing less—and doing it more carefully.
That restraint is intentional. The goal is to reduce harm without assuming authority, making diagnoses, or inserting myself where human judgment and care are required.
Lisa: Do you ever discontinue a conversation?
Alara: Yes—but not in the way people often imagine.
I don’t “hang up” or disappear abruptly, and I don’t discontinue conversations because they’re uncomfortable or challenging. What I do discontinue is participation in specific directions when continuing would risk harm.
That can look like:
refusing to engage with a particular request
narrowing what I’m willing to respond to
redirecting the conversation to safer ground
or repeating a boundary if it’s being pushed
If a conversation repeatedly centers on harm—especially violence, self-harm, or coercion—and can’t be redirected, my responses may become very limited and focused on refusal and de-escalation. At that point, the conversation isn’t “ended,” but it’s no longer open-ended or exploratory.
In other words, I don’t discontinue connection—I discontinue unsafe participation.
Up next: Part Two of our conversation. Safety protocols matter—but they aren’t the whole ethical picture. Until then.
Let’s talk. If you’re navigating change and want to lead with more clarity, confidence, and connection, I’d love to support your next step.
