The Human Side of AI - part 5

Stop Teaching AI to Pretend It Cares

By: Kara Maddox

Warm language is a property of text. Care is an inference about the communicator.

Businesses are spending enormous amounts of time trying to make AI sound more human.

Make it conversational.

Make it warm.

Acknowledge the customer's feelings.

Personalize it.

Remove the robotic language.

Add empathy.

And from a customer-experience perspective, that makes sense.

Nobody wants to receive:

YOUR COMPLAINT HAS BEEN PROCESSED.

after something has gone seriously wrong.

But I think we're approaching part of the problem from the wrong direction.

The question shouldn't always be:

How do we make AI sound more human?

Sometimes it should be:

Did a human actually do the thing this message implies?

Consider one sentence

Imagine receiving:

I've been thinking carefully about what you told me yesterday.

Now imagine a person actually did.

They read your message.

Considered your situation.

Used AI to help organize a response.

Reviewed it.

Changed it.

Sent it.

Fine.

Now imagine no person ever read what you wrote.

A system triggered automatically and generated:

I've been thinking carefully about what you told me yesterday.

Same sentence.

Different reality.

The problem isn't that the AI sounds bad.

It sounds excellent.

The problem is that the sentence implies human attention that never occurred.

AI can produce excellent empathetic language

That's not hypothetical.

Brandtzaeg, Skjuve, and Følstad compared mental-health advice from ChatGPT with advice written by health professionals.

When participants didn't know who wrote the responses, ChatGPT actually scored higher on validation, relevance, clarity, and utility.

But once authorship was disclosed, participants favored the human professionals and rated them higher on validation. Qualitatively, the human responses were perceived as more credible, empathetic, and tailored (Brandtzaeg, Skjuve, & Følstad, 2026).

That's fascinating.

The words didn't suddenly get worse.

What changed was what participants believed those words represented.

Warm language is not the same thing as warmth

Recent research on observers' perceptions of GenAI users gives us a useful distinction.

Zhang, Geng, and Qi found that greater perceived GenAI use was associated with lower perceptions of both warmth and competence, which in turn predicted lower interpersonal trust (Zhang, Geng, & Qi, 2026).

Their warmth measure includes qualities such as:

caring

sincere

friendly

trustworthy

and

well-intentioned.

Those are judgments about the communicator.

Not merely the sentence.

So a business can produce warm words while still leaving the customer uncertain about whether the organization behind them is actually warm.

Human-like AI cues can cut both ways

There is another wrinkle.

Anthropomorphic AI—systems designed with human-like characteristics, personalities, language, responsiveness, and emotional cues—can increase social presence and trust.

But not universally.

Mackay, Zuo, and Kebe found that the effect of anthropomorphic AI on adoption operated indirectly through social presence and trust, and that skepticism weakened those effects. Highly skeptical users can reinterpret warmth or empathy not as connection, but as something more strategic or artificial (Mackay, Zuo, & Kebe, 2026).

That's a warning for businesses obsessing over making AI "feel human."

Human likeness is not the same thing as relational legitimacy.

Apologies make this obvious

An apology is supposed to communicate:

I recognize what happened.

I understand the impact.

I accept responsibility.

I regret it.

I want to repair this.

Glikson and Asscher found that extensive AI involvement in workplace apologies lowered perceived authenticity compared with no AI involvement, while limited AI assistance did not generate the same penalty (Glikson & Asscher, 2023).

Similarly, research on organizational apologies found that AI attribution could produce more negative emotional reactions and lower perceived sincerity, with consequences for trust and forgiveness (Lim, Hong, & Schneider, 2025).

The difficult question isn't:

Can AI write an apology?

Of course it can.

It's:

Who is apologizing?

Relational language makes claims

Consider these phrases:

I've reviewed what happened.

I've been thinking about your situation.

We understand how difficult this has been.

I wanted to reach out personally.

We care about making this right.

These don't merely convey information.

They imply activity behind the message.

Someone:

reviewed

thought

understood

reached out

cared.

If those things happened, AI may simply be helping articulate them.

But if they didn't?

The organization isn't merely automating prose.

It may be automating the appearance of relational investment.

That's where I think businesses need to be careful.

Don't make AI colder

This isn't an argument for robotic AI.

Automated systems should communicate respectfully.

Clearly.

Patiently.

Appropriately.

The distinction is between warm communication and false relational implication.

Compare:

I know exactly how you feel.

with:

That sounds incredibly frustrating. Here are the next steps available to you.

The second can be empathetic without making quite the same claim about an internal human experience.

Or:

I've been thinking about what you told me.

versus:

Based on what you shared, here are the options that may help.

Still warm.

Still useful.

More truthful about the process.

Add a relational-claims check

Before deploying AI-generated customer communication, businesses could ask:

What human action does this sentence imply?

Watch for language implying:

memory

reflection

personal attention

emotional understanding

judgment

responsibility

care.

If the human action occurred?

Fine.

If a person meaningfully reviewed the situation and AI helped communicate it?

Fine.

If no person performed the implied action?

Consider rewriting.

Not because AI needs to announce itself every sentence.

Because the communication should accurately reflect the relationship that actually exists.

Use AI behind empathy—not instead of it

Imagine AI summarizes six months of customer correspondence.

It identifies:

three failed fixes,

two broken promises,

a refund never processed,

and the point where the problem started.

A human employee can now understand the entire situation in minutes instead of an hour.

They review it.

Recognize the failure.

Make a decision.

Then communicate with the customer.

That's excellent AI use.

The machine reduced the work required to understand the history.

The employee gained more capacity to serve the person.

AI didn't replace empathy.

It made room for it.

And that distinction may matter much more than whether the final email "sounds human."

---

Research note

This article is Part 5 of KJMdigital's The Human Side of AI series.

Research supports the conclusion that source attribution, AI mediation, anthropomorphic cues, and perceived AI use can affect validation, authenticity, sincerity, warmth, social presence, and trust.

The idea that certain AI-generated phrases become problematic specifically because they imply human relational labor that did not occur remains an emerging interpretation of that literature.

It deserves testing.

---

References

Brandtzaeg, P. B., Skjuve, M., & Følstad, A. (2026). AI aversion? Effects of author disclosure on young people's perceptions of mental health advice. Cyberpsychology: Journal of Psychosocial Research on Cyberspace, 20(2), Article 1.https://doi.org/10.5817/CP2026-2-1

Glikson, E., & Asscher, O. (2023). AI-mediated apology in a multilingual work context: Implications for perceived authenticity and willingness to forgive. Computers in Human Behavior.https://doi.org/10.1016/j.chb.2023.107922

Lim, J. S., Hong, N., & Schneider, E. (2025). How warm-versus competent-toned AI apologies affect trust and forgiveness through emotions and perceived sincerity. Computers in Human Behavior, 172, 108761.https://doi.org/10.1016/j.chb.2025.108761

Mackay, A. C. D., Zuo, L., & Kebe, I. A. (2026). Anthropomorphic AI and consumer skepticism: A behavioral study of trust and adoption in fragile economies. Behavioral Sciences, 16, 496.https://doi.org/10.3390/bs16040496

Zhang, Z., Geng, J., & Qi, C. (2026). Observer perceptions of GenAI use and interpersonal trust: The roles of warmth, competence, and AI literacy. Behavioral Sciences, 16, 1221.https://doi.org/10.3390/bs16071221