See it workFeaturesReviewsArticlesFAQLog inGet the app

ArticlesMessages

Emotional intelligence is real, and much smaller than advertised

Thirty years of research on emotional intelligence at work. What holds up, what quietly falls apart, and the one finding I changed my own habits over.

By Samet Durgun · Co-founder of Subtext · 11 min read

A few months ago I went looking for a number.

You’ve probably seen some version of it. It shows up in every second LinkedIn post about soft skills: emotional intelligence accounts for some enormous share of what separates outstanding leaders from average ones. Ninety percent. Eighty-five. The figure drifts depending on who’s selling the workshop.

The underlying interest is real, to be fair. When the World Economic Forum asked employers in 2025 which skills they consider core1, empathy and active listening landed in the top ten, and leadership and social influence came third, behind analytical thinking and resilience. Nobody manufactured this topic.

But I wanted to know where the number came from. I build a tool that reads the emotional tone of a message before you send it, so it seemed like a bad idea to keep repeating a statistic I’d never actually checked.

It took a while to track down, and it turned up somewhere strange. I’ll come back to that.

Four different things sharing one name

The first thing that surprised me is that researchers don’t agree on what emotional intelligence is. Not in a nitpicking way. In a “these are different constructs” way.

Peter Salovey and John Mayer defined it in 1990 as a set of mental abilities: noticing emotions, understanding how they blend and change, using them to think, and regulating them. That version gets measured with a performance test that has right and wrong answers, like an IQ test for feelings.

Then Daniel Goleman’s 1995 book made the term famous, and his version bundles in things that look a lot like personality and motivation. Reuven Bar-On’s EQ-i and Konstantinos Petrides’ trait model are self-report questionnaires. Petrides has been fairly explicit that what he’s measuring sits inside personality rather than alongside intelligence.

Researchers eventually sorted this mess into three streams2: performance tests, self-report measures of the ability model, and mixed competency models. The awkward part is that the ability tests and the self-report tests barely agree with each other. When someone tells you “the research shows emotional intelligence predicts X,” the first useful question is which of these four things they mean.

The real effect sizes are smaller than the pitch

The biggest recent synthesis I found pools 253 effect sizes from 78,159 people3 across three decades. Here’s what it reports, after statistical correction:

Job performance, about 0.30. Job satisfaction, 0.29. Organisational citizenship behaviour, 0.36. Job stress, negative 0.43. Organisational commitment, 0.26.

A correlation of 0.30 means emotional intelligence lines up with roughly nine percent of the variation in job performance. That’s a real relationship. It’s also nowhere near “accounts for ninety percent of what makes leaders great.” The earlier landmark meta-analysis in this area4 found the same general size across all three measurement streams.

Worth noting: that same 2022 paper found the correlation between EI and job performance was lower in published studies than in unpublished ones. Usually publication bias runs the other way. I don’t know what to make of it, and neither, as far as I can tell, does the author.

It depends enormously on what your job is

Before you file 0.30 away as a flat fact, there’s a moderator that almost never gets mentioned.

Joseph and Newman sorted 191 occupations5 by how much emotion regulation the work actually demands, then re-ran the analysis inside each group. In high emotional labour jobs, the ones with constant customer or interpersonal contact, ability-based emotional intelligence predicted performance positively. In low emotional labour jobs it predicted performance negatively.

That second half never gets quoted. If your work is mostly solitary and technical, being unusually tuned to what everyone in the room is feeling might be a tax on your attention rather than an advantage.

Their model has a shape I like, too: you have to notice an emotion before you can understand it, and understand it before you can do anything about it. Perception, then understanding, then regulation. Skip a step and the next one has nothing to stand on.

Where the number comes from, and what happened when someone checked it

Here’s the strange place I mentioned.

The paper that finally told me where the ninety percent figure originates is the same paper that takes it apart. Harms and Credé open their meta-analysis6 by laying out the claims they’re about to test. Two of them: a 1999 trade-press piece asserting that emotional intelligence explains close to ninety percent of what separates outstanding leaders from average ones, and an information package from the largest distributor of emotional intelligence tests, equating emotional intelligence with good leadership.

A trade citation and a sales document. That’s the pedigree of the statistic your last offsite was built on.

Then they pooled 62 independent samples on emotional intelligence and transformational leadership. When the same person rated both the leader’s emotional intelligence and the leader’s effectiveness, the correlation was 0.59. Impressive. Very quotable.

When those two ratings came from different sources, it dropped to 0.12.

Same construct, same literature, one methodological change. Rater agreement was 0.16 for emotional intelligence and 0.14 for leadership, which is to say people barely agree with each other about either thing. The authors also note that trait questionnaires produced higher validities than the ability tests, and that some researchers remain sceptical about whether the construct is valid at all.

In 2015 a team went back and asked why self-reported emotional intelligence predicts job performance so well7. Their answer: because those questionnaires are quietly measuring a bundle of things we already had names for.

Specifically, the content of mixed EI measures overlaps with ability EI, general self-efficacy, self-rated performance, conscientiousness, emotional stability, extraversion and cognitive ability, at a multiple R of 0.79. Their updated estimate of the correlation with supervisor-rated performance came in at 0.29, down from the 0.47 that had been circulating.

So the tests that give the most flattering numbers are also the ones most entangled with personality traits we could already assess. That doesn’t make them useless. It does make “we screened for EQ” a less impressive sentence than it sounds.

Career effects are real, slow, and mostly about other people

The best synthesis on careers pools 150 samples and 50,894 people8. Emotional intelligence correlated 0.39 with career satisfaction, 0.21 with salary, and negative 0.21 with wanting to quit. It correlated 0.03 with how employable people felt, which is not distinguishable from nothing.

The salary figure deserves a footnote nobody gives it: it rests on five samples and 1,446 people. That’s thin. The authors also flagged possible publication bias for career satisfaction and turnover, and every study in the pool used self-report measures because there weren’t enough performance-based ones to analyse.

The most interesting career study I found followed 126 people from college into their careers9 and looked at salary ten to twelve years later, controlling for personality and cognitive ability. Emotional intelligence did predict pay. But the effect ran through having a mentor, and it was stronger higher up the org chart. An earlier study by the same lead author had found nothing at the two-year mark.

So the mechanism runs through other people. They build relationships, those relationships turn into someone senior taking an interest, and that pays off a decade later. Nothing in there about being better at the actual work. Small sample, one study, treat it gently. But it matches how careers seem to go.

The finding I changed my own behaviour over

Everything above is interesting. This one is useful.

In 2005 Kruger, Epley, Parker and Ng ran five experiments10 on how well people convey tone over email. Participants wrote messages meant to be sarcastic, serious, angry or sad, then predicted how many the recipient would decode correctly.

Across 154 pairs, senders predicted 88.8 percent accuracy. Actual accuracy was 70.4 percent. Narrow it to email alone and it gets worse: 89.3 percent predicted, 62.8 percent actual. Face-to-face landed at 73.9 percent, voice-only at 73.3 percent. In an earlier experiment, email accuracy sat at roughly chance.

The confidence never moved. People’s actual ability changed a lot depending on the channel. Their belief in that ability stayed flat.

Two details make this stick for me. First, friends were no better calibrated than strangers. Knowing someone well did not help. Second, and this is the practical bit: in one experiment the researchers asked people to read their own message aloud in the opposite tone to what they meant, sarcastic lines read straight, straight lines read sarcastic. The overconfidence disappeared completely. Once the sentence stopped sounding the way they’d intended it, they could finally hear how ambiguous it was.

There’s a related finding I like: when people try to convey specific emotions in text11, senders’ confidence has no relationship at all to whether they succeeded.

The study that argues back

I’d rather show you the disagreement than pretend the field is settled.

In 2025, Pollmann and Roos took a different approach12. Instead of lab tasks with strangers and constructed sarcasm, they collected real messages people had actually received, then went and asked the actual sender what they’d meant. Study one covered informal texts, 347 receivers and 171 senders. Study two covered work and educational emails, 361 receivers and 61 senders.

Receivers got the tone right. The paper is titled “I get u” for a reason, and the authors argue it challenges the assumption that text is an emotionally impoverished medium.

I don’t think this cancels the older work. My read is that both are true about different messages. Most of what we send is unremarkable and lands fine. The misreading concentrates in the small pile of messages that are ambiguous, emotionally loaded, or trying to do something subtle. Kruger’s participants were forced to write exactly those. Pollmann’s participants forwarded whatever was in their phone.

Which happens to be the pile people ask for help with.

Neutral reads as negative, and the period is not innocent

Kristin Byron’s theoretical framework proposes two effects worth knowing by name13: receivers read positive emails as neutral, and neutral emails as negative. Emotion drains one notch downward in transit.

There’s empirical support. A study of real negatively perceived work emails found a negative intensification bias14: receivers rated the emails more harshly than neutral observers did, and their ratings had only weak relationships to what was actually in the message, including how much negative language it contained. The bias was stronger in bad communication climates and among people lower in the hierarchy. Which means the same email can be objectively fine and still land as an attack, depending on who’s reading it and how their week is going.

The punctuation research is smaller but real. A one-word reply ending in a period gets rated as less sincere15 than the same reply without one, and the effect shows up in text messages but not in handwritten notes. A follow-up study replicated it with longer, more natural exchanges16.

“Yeah” and “Yeah.” are not the same message, and the difference is created entirely by the reader.

Can any of this be trained

Reasonably, yes, with caveats.

A 2024 meta-analysis of workplace training in emotional competencies17 found a pre-post effect of 0.44, and 0.46 when comparing against control groups across 27 controlled trials. Effects held up more than three months out. Teachers, health professionals and managers all benefited about equally, and it didn’t matter much whether the programme was branded as emotional intelligence, empathy or emotion regulation.

The authors are refreshingly blunt about the limits: high heterogeneity and low methodological quality across the studies they pooled. Earlier meta-analyses landed in similar territory18 [19]. So the honest summary is that these programmes move the measures, most of which are self-reported, and we know much less about whether they change what people actually do at work.

What happens when AI writes it for you

Subtext is an AI tool, so I have an obvious interest here. I’ll give you the uncomfortable findings first.

People cannot tell. Across six experiments with 4,600 participants20 judging 7,600 pieces of writing in professional, hospitality and dating contexts, accuracy at spotting AI-generated text was 50 to 52 percent. Chance is 50. Worse, the researchers showed you could deliberately optimise AI text against people’s flawed intuitions and get it rated as human more often than actual human writing.

And yet suspicion carries a cost. Earlier work from the same group found that when people believed a profile was AI-written21, they trusted its author less.

Put those together and you get something genuinely awkward. Detection barely works, so the penalty attaches to suspicion rather than to anyone actually catching you. Which means the question of what’s acceptable here gets settled socially, by norms, and not by evidence.

My own line, for whatever it’s worth: the tool should help you say what you actually meant, not manufacture a sentiment you don’t have. That’s a real distinction to me, even if it’s invisible from the outside.

What I actually do now

Four things survived the reading and made it into how I write.

I assume my tone lands about twenty points worse than I think it does. Not as a mood. As arithmetic. 88.8 predicted, 70.4 actual.

When something’s important, I read it in the wrong tone. Out loud, sarcastically, if the message is sincere. It’s mildly humiliating and it works better than any rule about exclamation marks. It’s the only intervention in that whole paper that fully eliminated the bias.

For anything genuinely negative, I get on a call. Face-to-face and voice both beat email by roughly ten points on accuracy, and more importantly they let you repair in real time when something lands wrong. Written follow-up afterwards, for the record.

I stopped treating “they’re overreacting” as a settled fact. The negative intensification research shows receiver perception is only loosely tied to what’s in the message. That cuts both ways. Sometimes it’s them. Sometimes the message really was ambiguous and I was the last person capable of noticing.

The thing I’d want someone to take from all this: the big claims about emotional intelligence and career success are softer than they sound, built on self-report and often on the same person rating both sides of the equation. The small claims about written communication are sturdier and considerably more actionable. Nobody’s going to sell you a certification in “read it back in the wrong voice before you hit send.”

It’s free, it takes eleven seconds, and it has better evidence behind it than most things I’ve paid for.

Think I’ve undersold emotional intelligence, or oversold the email research? 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.

  1. World Economic Forum (2025). The Future of Jobs Report 2025, Skills Outlook. https://www.weforum.org/publications/the-future-of-jobs-report-2025/in-full/3-skills-outlook/
  2. Ashkanasy, N. M., & Daus, C. S. (2005). Rumors of the death of emotional intelligence in organizational behavior are vastly exaggerated. Journal of Organizational Behavior, 26, 441-452. https://doi.org/10.1002/job.320
  3. Doğru, Ç. (2022). A meta-analysis of the relationships between emotional intelligence and employee outcomes. Frontiers in Psychology, 13, 611348. https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2022.611348/full
  4. O’Boyle, E. H., Humphrey, R. H., Pollack, J. M., Hawver, T. H., & Story, P. A. (2011). The relation between emotional intelligence and job performance: a meta-analysis. Journal of Organizational Behavior, 32, 788-818. https://doi.org/10.1002/job.714
  5. Joseph, D. L., & Newman, D. A. (2010). Emotional intelligence: an integrative meta-analysis and cascading model. Journal of Applied Psychology, 95(1), 54-78. https://pubmed.ncbi.nlm.nih.gov/20085406/
  6. Harms, P. D., & Credé, M. (2010). Emotional intelligence and transformational and transactional leadership: a meta-analysis. Journal of Leadership & Organizational Studies, 17(1), 5-17. https://digitalcommons.unl.edu/cgi/viewcontent.cgi?article=1013&context=leadershipfacpub
  7. Joseph, D. L., Jin, J., Newman, D. A., & O’Boyle, E. H. (2015). Why does self-reported emotional intelligence predict job performance? A meta-analytic investigation of mixed EI. Journal of Applied Psychology, 100(2), 298-342. https://pubmed.ncbi.nlm.nih.gov/25243996/
  8. Pirsoul, T., Parmentier, M., Sovet, L., & Nils, F. (2023). Emotional intelligence and career-related outcomes: a meta-analysis. Human Resource Management Review. https://orbi.uliege.be/bitstream/2268/310903/1/Preprint_Pirsouletal.pdf
  9. Rode, J. C., Arthaud-Day, M., Ramaswami, A., & Howes, S. (2017). A time-lagged study of emotional intelligence and salary. Journal of Vocational Behavior, 101, 77-89. https://www.sciencedirect.com/science/article/abs/pii/S0001879117300416
  10. Kruger, J., Epley, N., Parker, J., & Ng, Z. (2005). Egocentrism over e-mail: can we communicate as well as we think? Journal of Personality and Social Psychology, 89(6), 925-936. https://web-docs.stern.nyu.edu/pa/kruger_email_ego.pdf
  11. Holtgraves, T. (2022). Implicit communication of emotions via written text messages. Computers in Human Behavior Reports, 7, 100219. https://doi.org/10.1016/j.chbr.2022.100219
  12. Pollmann, M. M. H., & Roos, C. A. (2025). “I get u”. People correctly interpret the tone of text messages and emails. Computers in Human Behavior Reports, 18, 100689. https://www.sciencedirect.com/science/article/pii/S2451958825001046
  13. Byron, K. (2008). Carrying too heavy a load? The communication and miscommunication of emotion by email. Academy of Management Review, 33(2), 309-327. https://journals.aom.org/doi/10.5465/amr.2008.31193163
  14. Sillars, A., & Zorn, T. E. (2021). Hypernegative interpretation of negatively perceived email at work. Management Communication Quarterly, 35(2), 171-200. https://journals.sagepub.com/doi/abs/10.1177/0893318920979828
  15. Gunraj, D. N., Drumm-Hewitt, A. M., Dashow, E. M., Upadhyay, S. S. N., & Klin, C. M. (2016). Texting insincerely: the role of the period in text messaging. Computers in Human Behavior, 55, 1067-1075. https://www.sciencedirect.com/science/article/abs/pii/S0747563215302181
  16. Houghton, K. J., Upadhyay, S. S. N., & Klin, C. M. (2018). Punctuation in text messages may convey abruptness. Period. Computers in Human Behavior, 80, 112-121. https://doi.org/10.1016/j.chb.2017.10.044
  17. Mehler, M., Balint, E., Gralla, M., Pößnecker, T., Gast, M., Hölzer, M., Kösters, M., & Gündel, H. (2024). Training emotional competencies at the workplace: a systematic review and metaanalysis. BMC Psychology, 12, 718. https://bmcpsychology.biomedcentral.com/articles/10.1186/s40359-024-02198-3
  18. Hodzic, S., Scharfen, J., Ripoll, P., Holling, H., & Zenasni, F. (2018). How efficient are emotional intelligence trainings: a meta-analysis. Emotion Review, 10(2), 138-148. https://doi.org/10.1177/1754073917708613
  19. Mattingly, V., & Kraiger, K. (2019). Can emotional intelligence be trained? A meta-analytical investigation. Human Resource Management Review, 29(2), 140-155. https://doi.org/10.1016/j.hrmr.2018.03.002
  20. Jakesch, M., Hancock, J. T., & Naaman, M. (2023). Human heuristics for AI-generated language are flawed. Proceedings of the National Academy of Sciences, 120(11), e2208839120. https://www.pnas.org/doi/10.1073/pnas.2208839120
  21. Jakesch, M., French, M., Ma, X., Hancock, J. T., & Naaman, M. (2019). AI-mediated communication: how the perception that profile text was written by AI affects trustworthiness. CHI Conference on Human Factors in Computing Systems. https://www.researchgate.net/publication/332747673_AI-Mediated_Communication_How_the_Perception_that_Profile_Text_was_Written_by_AI_Affects_Trustworthiness