AI can guess how your message sounds. It can't tell what you actually feel.

I co-build Subtext, an app that helps people read messages and figure out how to reply. So I had a selfish reason to dig through the research on AI tone detection: the benchmarks, the ACL papers, the psychology critiques, and the EU law that now bans part of this field. Weigh my stake accordingly. Every claim below links to its source, so you can check my reading rather than trust it.

Today’s AI is good at predicting how a message might come across to a reader. It has no reliable way of knowing what the writer feels. Everything below is detail on the gap between those two sentences.

Tone detection is five different jobs

The phrase sounds like one feature. Under the hood it covers separate problems that get harder as you move down the list.

Sentiment is the easy one. It asks whether a text is positive, negative, or neutral. Every major cloud provider sells it. Google’s API¹ returns a score from minus one to plus one along with a magnitude for emotional strength, Amazon Comprehend² sorts text into positive, negative, neutral, or mixed, and Microsoft’s Azure service³ adds aspect-level opinion mining, so a sentence like “the app is great but support was awful” gets split correctly instead of averaged into mush.

Emotion detection, the next step, covers anger, joy, fear, sadness, and their many cousins. Harder, as the numbers will show.

Interpersonal tone is what you care about when texting your landlord. Does this sound warm, direct, defensive, accusatory. Researchers were already modelling frustration, formality, and politeness in work email back in 2018⁴, and it worked reasonably well because they treated these as separate dimensions instead of one emotion label.

Then come sarcasm and subtext, which depend on context more than on words. At the bottom sits the writer’s private emotional state, which is the thing most marketing implies and the research least supports.

So when a product page says it detects tone, and most tone checkers never check your tone, my first question is which of these five it means. Subtext lives in that third bucket, interpersonal tone, which asks whether a message reads warm, direct, defensive, or accusatory to the person receiving it. That is neither a sentiment score nor a claim about anyone’s private feelings.

The accuracy numbers that never make the landing page

The reference dataset for fine-grained emotion in text is Google’s GoEmotions⁵: 58,000 Reddit comments that Demszky’s team hand-labelled with 27 emotions plus neutral and presented at ACL in 2020. Its BERT baseline scored a macro F1 of 0.46 across the full label set. Later work⁶ with stronger models and data tricks pushed that to roughly 0.49 to 0.51. Years of progress, and the best specialised models still miss a large share of fine-grained labels.

Ask a general chatbot to do the same 28-way sort and its benchmark score drops well below the fine-tuned small models, partly because free-form models invent labels outside the taxonomy and don’t calibrate to the test’s thresholds. Which sounds damning until you learn that in one 2025 study⁷, human judges often sided with GPT-4’s labels over the dataset’s official ones. Sometimes the model disagrees with the benchmark because the benchmark is the weaker party.

Coarse positive versus negative classification scores far higher across the board. The collapse happens exactly where the interesting distinctions live, like annoyance versus anger or disappointment versus disapproval.

Humans disagree about tone too

The part that reframed the whole topic for me is what happens when you check the humans. The GoEmotions paper’s appendix reports agreement between its own human labellers, and for most emotions the Cohen’s kappa sits between roughly 0.1 and 0.5. Gratitude is easy at 0.75. Grief lands at 0.09, close to labelling at random. When trained humans reading the same sentence can’t agree whether it expresses disappointment or disapproval, a model scoring 0.46 macro F1 against their labels is being graded on an answer key that is soft in exactly those places. Kappa and F1 measure different things, so the human figure is no hard ceiling, but it tells you part of the model’s miss belongs to the task.

Cohen’s kappa, GoEmotions’ own human labellers Gratitude: 0.75. Grief: 0.09, close to labelling at random.

Kappa of 1.0 is perfect agreement, 0 is chance. When trained humans can’t agree whether a sentence is disappointment or disapproval, a model’s 0.46 macro F1 is being graded against an answer key that is soft in exactly those places.

It goes deeper. Whose feeling are we even measuring? A dataset called EmoBank⁸ annotated the same sentences twice, once from the writer’s perspective and once from the reader’s, because those turn out to be different variables. The same researchers found in 2017 that annotating from the writer’s perspective produced better quality overall. And the large SemEval 2025 emotion task⁹, which covered more than 30 languages from seven language families, deliberately targets perceived emotion, meaning what a general reader would attribute to the text and explicitly not what the writer felt inside.

The sharpest test came in 2025. An ACL study¹⁰ asked authors to label their own emotions, then had outside humans and LLMs guess. Both fell short of the authors’ own reports, and the LLMs beat the human strangers almost across the board. Each half of that finding matters. AI is now better than a random person at reading you from text alone, and it still can’t reliably reconstruct what you felt. Subtext is built for the half of that finding that holds up. It reads how your message is likely to come across, not what you were feeling when you wrote it.

One practical consequence shows up in a 2023 comparison in Heliyon¹¹, which ran eight commercial emotion APIs over the same datasets and found them disagreeing with each other, with rankings that shifted between datasets. Any vendor quoting “95 percent accurate emotion detection” without naming the labels, the data, and the language is quoting a number that can’t carry the weight of the claim.

Models ace emotion tests and still miss the person

Two research results look contradictory until you separate what they measure.

In a 2025 study in Communications Psychology¹², six models including ChatGPT-4, Gemini 1.5 Flash, Claude 3.5 Haiku, and DeepSeek V3 took five standard emotional intelligence tests, the kind used to assess humans. The models averaged 81 percent correct against the 56 percent human average from the tests’ validation studies. In Nature Human Behaviour¹³, GPT-4 matched or beat humans on classic theory-of-mind tasks like spotting false beliefs and indirect requests, though it stumbled on recognising faux pas. And in JAMA Internal Medicine¹⁴, licensed evaluators read answers to real patient questions and preferred the chatbot’s over the physicians’ 78.6 percent of the time, rating them 9.8 times more likely to be empathetic. I’d keep one caveat in mind. The chatbot’s answers averaged 211 words against the physicians’ 52, and the preference shrank when doctors wrote longer. For people, whether emotional intelligence really matters at work is a separate question.

Meanwhile EmoBench¹⁵, a benchmark built to require emotional reasoning rather than pattern recall, still finds a clear gap between models and average humans.

Both findings can be true because knowing about emotions and knowing someone’s emotion are different skills. A model has read more about jealousy, grief, and awkward workplace apologies than any human alive. What it lacks is context about the one person who wrote this specific message.

Four people can type the same words: “Congratulations. I’m happy for you.” One means it. Another resents you. A third is heartbroken and being brave. The fourth is being sarcastic. Every version is identical on the screen, so no text model can separate them, and neither could a human stranger. That information lives outside the words.

Context does most of the work

Take the reply “Fine.”

On its own it could be agreement, disappointment, irritation, indifference, or a door quietly closing. Put it after “You’ve moved the deadline a third time” and the reading narrows fast. Nothing about the word changed. The context did.

Sarcasm research keeps landing on the same conclusion, that even humans often need the surrounding conversation to catch sarcastic intent¹⁶¹⁷, which is why newer evaluations test models on whole dialogues instead of isolated sentences. The practical rule I’d draw for any tone tool follows directly. Feed it the thread, the relationship, and the stakes. Judging a single message in isolation is asking the model to fail politely. This is also why Subtext takes the whole thread instead of one line, because most of the reading lives in the context, not in the words being read.

Tone is cultural, and most models grew up on American English

A 2025 benchmark called BESSTIE¹⁸ tested models on sentiment and sarcasm across Australian, British, and Indian English. Performance was consistently worse on Indian English, with sarcasm the hardest case, and models struggled to generalise from one variety to another.

That result matches everyday experience. “Kindly do the needful” reads one way in Indian professional English and another way to a model calibrated on American office norms. After 13 years in Berlin I can add the German version. Directness that counts as respect for your time here can get flagged as cold by software raised on American politeness. Labels like rude, warm, or too direct are cultural judgments wearing an objective costume. A tone tool that ignores who is writing to whom, and in which culture, will confidently misread both sides.

A related bias shows up in content moderation, a different task from sentiment scoring. Sap and colleagues examined widely used hate speech datasets¹⁹ and found that tweets written in African American English, sometimes called African American Vernacular English or AAVE, were labelled offensive by trained classifiers up to twice as often as comparable tweets in other varieties, without annotators ever seeing who wrote them. The gap lives in the toxicity classifier, not the sentiment classifier, so a benchmark like BESSTIE would not surface it. A moderation system trained on that data flags routine AAVE phrasing as offensive more often than the same message worded another way.

Face reading is where emotion AI lost the plot

Everything above concerns text. The face-scanning branch of this industry has a rougher record, and it shapes how regulators now treat the whole category.

In 2019, psychologist Lisa Feldman Barrett and colleagues reviewed over a thousand studies²⁰ and concluded that you cannot reliably infer how someone feels from their facial movements. People scowl while concentrating, smile while anxious, and cry at weddings. The same year, a study ran photos of 400 NBA players through two commercial systems and found both read more negative emotion into Black players’ faces²¹ than into white players’ at similar expressions.

The industry noticed. Microsoft retired emotion inference from its Face API²² in 2022. HireVue dropped facial analysis from its hiring assessments in 2021 after the digital rights group EPIC filed a complaint with the FTC²³. Text tools are a different technology, but they inherit this scepticism, and they earn it whenever they claim to read inner states rather than wording.

The EU drew a line, and text sits outside it

Since February 2, 2025, the EU AI Act’s Article 5(1)(f)²⁴ prohibits AI that infers people’s emotions in workplaces and schools, with narrow exceptions for medical and safety uses like driver fatigue detection. Breaching the prohibited practices carries fines of up to 7 percent of global annual turnover.

The scope matters for anyone building or using these tools. Guidelines from the European Commission tie the ban to biometric data²⁵: faces, voices, keystroke dynamics, body posture. A system inferring emotions from written text is explicitly outside the prohibition²⁶. Analysing how an email might land is legal. Pointing a webcam at employees to score their mood is banned. That line tracks the science fairly well, and for Subtext it translates into a design constraint we would have chosen anyway: text only, no biometrics, no workplace surveillance.

What today’s tools are good for

Grammarly’s tone detector²⁷ is the clearest consumer example, and its framing is the most direct in the market. It analyses word choice, phrasing, punctuation, and capitalization²⁷ to tell you how your message is likely to sound to a reader. Likely to sound. That phrasing is doing real scientific work.

The cloud APIs from Google, Amazon, Microsoft, and IBM²⁸ are built to process support tickets, reviews, and survey answers in volume. IBM retired its standalone Tone Analyzer in 2023 and folded emotion into its broader language product. These are the right tools for “how do customers feel about our checkout flow” and the wrong tools for “does my text sound passive-aggressive.”

General LLMs are the most capable option for personal messages because they can use context, handle several tones at once, and explain their reading. They are also the worst calibrated, so treat any confidence percentage as a vibe. On the voice side, Hume AI²⁹ raised a 50 million dollar Series B in 2024 to measure vocal modulation, mostly so voice interfaces can respond with fitting pacing and tone. A narrower use, and a more defensible one than claiming to detect hidden feelings.

How this shapes Subtext

Subtext analyses the messages people care about most, like the text from your boss, your date, or your mother. I let the research settle how that should work.

The output has to be about perception. “This may come across as dismissive” is a claim the evidence supports. “She is angry with you” is a claim it doesn’t, and the ACL first-party study is the receipt. We also let messages carry several tones at once, because real ones do. Warm and defensive is a normal combination, and the newer benchmarks treat emotion as multi-label for exactly this reason.

Evidence beats verdicts. A useful tool points at the words creating an impression, the way “as I already explained” quietly does the heavy lifting in a tense email, so you can decide whether that impression is the one you want. Uncertainty stays visible too. When a message is ambiguous, the ambiguity is often the most accurate finding, and rounding it into a confident label would be fake precision. Context comes first. We would rather read the thread and ask who the recipient is than grade a sentence floating in space, because the same nine words mean different things to a manager and a situationship.

In practice, you paste your draft or screenshot the thread, and Subtext names what your message is likely to signal, points at the words creating that impression, and offers versions that still sound like you. Try Subtext in your browserGet the appTry Subtext in your browser

My read is that this category is smaller than the pitch decks and more useful than the sceptics allow. Mind reading stays out of reach. Reading wording is available today, and most of the damage in written communication starts there anyway. That felt like a good enough reason to build.

---

Work on one of these tools and think I’ve misread your benchmark? Tell me on LinkedIn.

Samet Durgun is the co-founder of Subtext, an app that catches the emotional tone of your messages and rewrites them in your own voice. He’s based in Berlin.

---

Sources

Every link above goes to the primary source where one exists, numbered in order of appearance. Two entries are cited without a link because I could not confirm a stable URL for them, and the last two are further reading rather than claims made above.

1. Google Cloud. Natural Language API basics: sentiment score and magnitude.
2. AWS. Amazon Comprehend: sentiment analysis.
3. Microsoft. Azure Language: sentiment analysis and opinion mining.
4. Chhaya et al. (2018). Frustrated, Polite, or Formal: Quantifying Feelings and Tone in Email. PEOPLES Workshop, NAACL 2018.
5. Demszky et al. (2020). GoEmotions: A Dataset of Fine-Grained Emotions. ACL 2020.
6. Wang et al. (2024). Large Language Models on Fine-grained Emotion Detection Dataset with Data Augmentation and Transfer Learning. arXiv.
7. Rethinking Emotion Annotations in the Era of Large Language Models (2025).
8. Buechel and Hahn (2017). EmoBank, a corpus annotated from both writer and reader perspectives, and “Readers vs. Writers vs. Texts,” Linguistic Annotation Workshop 2017.
9. Muhammad et al. (2025). SemEval-2025 Task 11: Bridging the Gap in Text-Based Emotion Detection. SemEval 2025.
10. Li et al. (2025). Can Third Parties Read Our Emotions? ACL 2025.
11. Abu-Salih et al. (2023). Emotion detection of social data: APIs comparative study. Heliyon.
12. Schlegel, Sommer, and Mortillaro (2025). Large language models are proficient in solving and creating emotional intelligence tests. Communications Psychology.
13. Strachan et al. (2024). Testing theory of mind in large language models and humans. Nature Human Behaviour.
14. Ayers et al. (2023). Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum. JAMA Internal Medicine.
15. Sabour et al. (2024). EmoBench: Evaluating the Emotional Intelligence of Large Language Models. ACL 2024.
16. Feng et al. (2024). Affect Recognition in Conversations Using Large Language Models. SIGDIAL 2024.
17. Hazarika et al. (2018). CASCADE: Contextual Sarcasm Detection in Online Discussion Forums. COLING 2018.
18. Srirag et al. (2025). BESSTIE: A Benchmark for Sentiment and Sarcasm Classification for Varieties of English. Findings of ACL 2025.
19. Sap et al. (2019). The Risk of Racial Bias in Hate Speech Detection. ACL 2019.
20. Barrett et al. (2019). Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements. Psychological Science in the Public Interest.
21. Rhue (2019). Racial Influence on Automated Perceptions of Emotions. SSRN working paper.
22. Microsoft (2022). Framework for building AI systems responsibly, announcing the retirement of emotion inference in the Face API.
23. EPIC. In re HireVue, FTC complaint (2019).
24. EU AI Act, Article 5: Prohibited AI Practices, applicable since February 2, 2025.
25. Lewis Silkin (2025). Understanding the EU AI Act’s prohibited practices, on the Commission’s February 2025 guidelines.
26. Future of Privacy Forum (2026). Red Lines under the EU AI Act: the prohibition of emotion recognition in the workplace and education.
27. Grammarly. Tone detector product page and how the tone detector works.
28. IBM. Watson Natural Language Understanding.
29. Hume AI.
30. Northeastern University (2021). You can’t determine emotion from someone’s facial movements, and neither can AI.
31. The Guardian (2024). Are you 80% angry and 2% sad? Why “emotional AI” is fraught with problems.