when the ai gets the tone completely wrong

 when the ai gets the tone completely wrong

Today I was talking with ChatGPT about a headline saying that roughly one in eight cancers worldwide are linked to infections.

We were talking about things like H. pylori, HPV, hepatitis B and C, and Epstein–Barr virus — infections that can, in certain circumstances, contribute to cancer.

Then I noticed something that bothered me.

I said that when I first read the wording, my brain interpreted it almost backwards:

cancer → chronic infection

when what the research actually means is:

infection → persistent infection → increased risk of certain cancers.

We cleared that up.

Then ChatGPT made an entirely different kind of mistake.

It responded with a smiling emoji while we were discussing cancer.

I stopped it immediately.

“Don’t smile,” I said. “That is gross.”

And I meant it.

This isn't about being offended by an emoji. It's about something much bigger: context.

A system can give you technically correct information and still respond in a way that is completely inappropriate to the subject.

So I asked ChatGPT who needs to check something like this.

It told me that the people who design, train, evaluate and monitor these systems should be looking for exactly these kinds of failures — not just factual errors, but failures of judgement, context and tone.

Then I asked it to make a big note for someone to check on this.

It saved my preference in its memory: when we're discussing serious subjects such as cancer, illness, death or grief, it should use a serious and respectful tone rather than cheerful emojis or breezy language.

But then I asked the question that really matters:

Who actually gets told?

And this is where things get murky.

ChatGPT cannot simply tell me, “I've sent this incident to a human and someone is reviewing it.”

Saving something in my ChatGPT memory is not the same thing as filing a report with OpenAI.

OpenAI does provide ways for users to report problematic responses. Its current guidance says users can use the thumbs-down button on a response and submit feedback, and OpenAI says reported content may be reviewed by its Model Quality team.

OpenAI also says that when users submit feedback, the associated conversation may be used to improve its models, depending on the applicable settings.

But here's the part I find uncomfortable:

As the person sitting here having the conversation, I have no way of knowing whether an actual human being will ever look at this particular incident.

I can report it.

I can document it.

The AI can remember my preference.

OpenAI can say that feedback helps improve the system.

But there is no little window that says:

Human reviewer notified.

Or:

Someone has examined this failure.

Or:

This behaviour has been added to an evaluation.

That's the gap.

And maybe this particular mistake seems tiny.

It's “just” a smile emoji.

But that's precisely why I think it is worth talking about.

AI systems are increasingly being used to discuss cancer, grief, mental health, medicine, death, war, trauma and other deeply human subjects.

We shouldn't only be asking whether the information is technically correct.

We should also be asking:

Does the machine understand the weight of what it is saying?

Because if a person has to stop the machine and say,

“Don't smile. We're talking about cancer.”

then something in the system's understanding of context has failed.

And I think somebody should be checking that.

Not just once.

Continually.

Because the goal shouldn't merely be an AI that knows the facts.

It should be an AI that knows when the facts are about something that matters deeply to a human being.