Skip to content
Affectosphere Group Kansai University, Keito Inoshita Laboratory
JP

Column

An AI That Takes “Thank You” at Face Value Will Lose Your Most Important Customers

When a customer says “well done,” is it praise or sarcasm? The moment your AI mistakes one for the other, NPS goes up, improvement priorities go down, and the genuinely angry customers leave in silence. Five minutes for CX leaders, drawn from two recent sarcasm-recognition papers.

5 min read 日本語版 →
The same phrase “well done” rendered in two colors representing surface meaning vs. real intent

This article is an English translation of the original Japanese column. Some phrasing has been adapted for English readers.

Hello — this is Inoshita from Affectosphere Group.

I had a conversation the other day with a CX manager at a SaaS company who said:

“Our NPS keeps creeping up every month. But churn is also going up. What do you think is going on?”

That should give you a small chill.

NPS up, churn up. This combination, in my experience, is very often caused by a sentiment-analysis AI quietly taking sarcasm at face value.

Picture this review.

“Wonderful handling — more than enough to make me decide never to use your service again.”

The moment your AI labels that “positive,” your NPS overstates reality, your improvement queue gets deprioritized, and the genuinely angry customer leaves without a word.

On the dashboard, everything looks fine. That’s the scary part.

Today I want to translate two sarcasm-recognition papers we published in 2025-202612 into the language of CX, VoC, and support operations.

Three-line takeaway

  1. Value: an AI that can read sarcasm correctly surfaces the single most actionable segment — “surface positive × actual negative.”
  2. Caveat: models that decompose the cognitive steps, or honestly admit uncertainty, beat one-shot LLM classifiers in real operations.
  3. Affective AI lens: whether you own an AI that refuses to crush ambiguity decides the resolution of your CX. An AI that won’t expose its confidence is harder to operate than one that does.

In order.

1. Treat sarcasm as a cognitive chain, not a single classification

The first paper, WM-SAR (World Model Inspired Sarcasm Reasoning), is interesting because it gives up on one-shot classification.

When humans recognize sarcasm, we unconsciously go through roughly these steps.

  1. Observe the situation (“the boss let me push the deadline by one day”)
  2. Predict the normative utterance for that situation (“usually you’d just say ‘thanks, that helped’”)
  3. Detect the gap between the actual utterance and that norm (“‘wonderful exercise of discretion’… that’s off”)
  4. Infer the real intent from the gap (“ah, this is sarcasm”)

WM-SAR distributes these four steps across separate LLM agents.

The “world model” here refers to the internal predictive apparatus humans use to make sense of reality. Sarcasm exists because a speaker deliberately places a gap between the norm predicted by the world model and what they actually say.

WM-SAR pulls this structure straight into the implementation.

Why does this matter operationally? Because the output naturally carries a reason. A VoC report can now say, “this utterance is surface-positive, but it deviates from the situational norm, so we classify it as sarcastic.” That makes decision-makers an order of magnitude more comfortable with the call.

2. Process surface and substance as two parallel streams

The second paper, DBDA-EDL, takes a different cut.

It processes the same utterance through two parallel streams — surface representation and substance representation — and explicitly extracts the mismatch (discrepancy) between them as a fused feature.

Sarcasm is essentially that mismatch, so let’s build the path that extracts the mismatch directly into the model.

The other key piece is uncertainty representation via Evidential Deep Learning (EDL).

Roughly: a model that can return a low, hesitant confidence on utterances that don’t cleanly fall on either side of “sarcastic / not sarcastic.”

This is incredibly useful in operations.

A model that says “I’m not sure” at the right moments is just as valuable as one that gets it right with high confidence. Because you can route the low-confidence cases to humans — that’s a real workflow.

With a model like DBDA-EDL, “AI auto-handles when confident, humans handle when not” becomes a realistic hybrid setup. You concentrate operator time on the sarcastic and nuanced edge cases.

3. The part I most want to convey as an affective-AI lab

This is what I most want CX teams to hear from us.

The real risk of running an “AI that can’t read sarcasm” isn’t the misclassification itself.

The real risk is that the AI is confidently wrong, and your floor decisions slowly skew. NPS keeps drifting up, CSAT looks healthy, the dashboard is green — and churn quietly accelerates. Cause and effect get disconnected, and the team works against the wrong priorities for months.

You can’t fix this by lifting accuracy.

Even if you bump sarcasm accuracy from 80% to 95%, if the remaining 5% is misclassified with high confidence, your team will believe it. This isn’t an accuracy problem. It’s an ambiguity-handling problem.

Our lab’s core stance is to treat emotion as ambiguous and many-sided on purpose. Because human feeling can’t be averaged or voted away.

Sarcasm is the phenomenon where that ambiguity shows up most intensely in language. The strongest feeling expressed in the gentlest words. An AI that takes it at face value loses the customer’s real voice.

If you only evaluate AI on accuracy, efficiency, and cost, you’ll never see this layer. Bringing an affective-AI lens to the operation buys you the ability to design for ambiguity, not against it.

So what do you do tomorrow?

To be fair, three things you can actually move on.

  • Hand-review 100 “positive”-labeled samples each month. Measuring sarcasm-leakage gives you the confidence interval on your sentiment pipeline.
  • Add a confidence guard to auto-replies. Low-confidence cases must escalate to a human. This single rule meaningfully reduces the risk of a sarcastic-reply screenshot going viral.
  • Set up “sarcasm-subset recall” as a standalone KPI, not just overall accuracy. That number is a leading indicator of brand damage.

Back to the CX manager’s question: if NPS is up but churn is up, start with the monthly sample review. That’s the realistic first move.

Closing

Sarcasm is the phenomenon where the strongest feeling appears in the gentlest words.

Whether your AI can handle it is therefore a litmus test of whether it’s actually listening to the customer.

And this is both a technology-selection question and a philosophy-of-deployment question for affective AI. Whether you have an AI that refuses to crush ambiguity decides the resolution of your CX.

If you ever wonder, “is our sentiment analysis actually okay?” — start with monthly sample reviews. You may meet a number that makes you uncomfortable.

That’s it for now.

References

  1. Keito Inoshita, Shinnosuke Mizuno (2026). World model inspired sarcasm reasoning with large language model agents. Discovery Artificial Intelligence.
  2. Takato Ueno, Keito Inoshita (2025). Dual-Branch Feature Extraction via Discrepancy-Aware Fusion with Evidential Deep Learning for Sarcasm Detection. IEEE IAICT 2025, pp. 345-352.

* This article was written in part with AI assistance and may contain inaccuracies.

Footnotes

  1. Keito Inoshita, Shinnosuke Mizuno, “World model inspired sarcasm reasoning with large language model agents”, Discovery Artificial Intelligence, 2026. ↩

  2. Takato Ueno, Keito Inoshita, “Dual-Branch Feature Extraction via Discrepancy-Aware Fusion with Evidential Deep Learning for Sarcasm Detection”, IEEE IAICT 2025, pp. 345-352. ↩