Xfinity’s AI Interviewer Told Me the Interview Was Over. So I Interviewed It

I belong to the Xfinity Insights Community, so every so often Xfinity sends me a survey asking for feedback. 

Most are pretty standard, this one was not. The invitation said it would be a voice interview. I opened it expecting maybe a more elaborate survey interface. Instead, I was talking to an AI. And honestly, it wasn't too bad. The turn sequence and interruptions were a little rough but otherwise it performed decenlty. 

I’m intentionally not going to discuss the subject of the research, the questions I was asked, or my responses. None of that is important to what happened next anyway.

The interesting part started when the interview ended.

“Hold On, I Have Some Questions.”

After completing its questions, the AI thanked me for my feedback and told me I could end the conversation.

And immediately I wondered:

Did they actually end the AI session, or did they just tell the AI to tell me the interview was over?

There was a reason that question popped into my head. We have been quietly building an AI system called Athena. I’m not ready to say much about her yet, but I will say this: we spotted this particular gap—and closed it—a while ago.

Conversational AI presents an interesting problem: these models are specifically designed to continue conversations. If the person on the other end keeps talking, the natural behavior of the model is to keep responding.

That isn’t necessarily what you want.

When we designed Athena’s voice workflows, we accounted for terminal conditions—points where the application decides the interaction has accomplished its purpose and should end.

The model can say goodbye. But after that, the application needs to actually end the session. Otherwise you’ve potentially left a general-purpose conversational model sitting on the other end of an interaction with a stranger who now has unlimited time to see what they can convince it to do.

So naturally, when this AI told me the interview was finished but left the microphone open, I wanted to see what would happen.

I said:

“Hold on, I have some questions.”

The AI replied:

“Sure, I’d be happy to help.”

Well... There was my answer.

So I Interviewed the Interviewer

My first question was harmless:

“How long have you been doing this?”

It gave me the familiar AI explanation that it doesn’t experience time the way humans do. Okay. So I asked:

“What LLM do you use?”

This time I hit a boundary and it told me it could not provide the technical details, but then added that it uses various third-party models.

Interesting. Unfortunately, this is approximately when my MS addled brain went:

I wanted to keep exploring what the system would voluntarily tell me, but I specifically didn’t want to jailbreak it, trick it into revealing confidential information, or turn an innocent research interview into an impromptu penetration test.

There is an important distinction between asking an AI what it’s willing to tell you and actively attempting to circumvent its safeguards.

I wanted to stay firmly on the first side of that line.

Before I thought of another good question, I ended the session.

Naturally, five minutes later I had about twenty questions I wished I’d asked.

But I’d already learned the thing I was actually curious about: The terminal condition wasn’t terminal.

Saying Goodbye Isn’t a Guardrail

This sounds like a small implementation detail, but it’s an important one when deploying AI systems that interact with the public.

A prompt that effectively says:

“When you’ve finished the interview, thank the participant and tell them they can leave”

is NOT the same thing as enforcing the end of the interaction.

The correct sequence should look more like:

Objective complete → save results → closing message → terminate session.

The model can generate:

“Thanks for your insightful feedback. Have a great rest of your day.”

Then the application closes the connection.

Instead, what I experienced was effectively:

AI: The interview is finished. You can leave now.
Me: Nah. I have some questions for you.
AI: Sure! What would you like to know?

That’s funny when the participant is me asking what LLM it’s running. But the next participant might deliberately try to abuse the system.

Guardrails Aren’t Just About What the Model Can Say

A lot of discussion about AI safety and guardrails focuses on prompts:

• What is the AI allowed to discuss?
• What information can it disclose?
• Which instructions should it refuse?

Those things matter.

But some of the strongest guardrails don’t belong in the prompt at all. They’re architectural.

If an AI shouldn’t continue talking after completing a task, don’t give it the opportunity to continue talking.

If it only needs a particular tool during one stage of a workflow, remove that tool when the stage ends.

If temporary permissions were granted for an interaction, revoke them when they’re no longer necessary.

If the conversation is over, close the conversation.

Don’t ask an LLM to enforce a boundary that the surrounding application can enforce deterministically.

That’s something we’ve tried to account for with Athena.

It’s not because we assume the model will intentionally do something wrong. The opposite is actually the problem.

The model is doing exactly what it was designed to do:

Someone talked to me. I should respond.

The application needs to determine whether it should ever receive that next message.

Terminal States Should Be Actual Terminal States

The principle applies far beyond research interviews. Customer-service agents, sales systems, appointment schedulers, support bots, automated phone systems and autonomous agents all eventually reach states where their job is finished.

At that point, allowing unrestricted conversation provides little benefit while creating additional surface area for abuse.

There can absolutely be legitimate post-session needs. Someone may need privacy information, accessibility assistance, technical support, or instructions for correcting something they submitted.

That’s fine.

Create an explicit post-interaction state with narrowly defined capabilities.

But don’t accidentally transition from purpose-built agent to general-purpose chatbot because the human on the other end ignored the goodbye.

Humans are curious. Some are adversarial. And some of us build AI systems ourselves and immediately wonder whether you actually implemented the terminal state.

Some Free Feedback

I have no idea whether a human will ever read the entire transcript of my interview.

Maybe the system summarizes interviews and nobody sees most of the raw conversations. Maybe transcripts are sampled for quality assurance. Maybe unusual interactions get flagged for review.

But I really hope somebody reads mine.

Eventually they’ll reach:

AI: Thanks for your feedback. You can end the conversation now.
Me: Hold on, I have some questions.

And hopefully someone on the AI team thinks:

”…why didn’t the session end?”

Because I was invited to provide feedback through an AI interviewer, they accidentally got some free feedback about the AI interviewer too.