The false-fluency trap: when "good-enough" English is worse than no English

TL;DR — Knowing a language poorly is worse than not knowing it at all.

Not knowing it makes the gap visible: the room agrees on an interpreter and reorganizes around it.
Poorly knowing it hides the gap: fluent-sounding errors get minuted as agreed; the cost surfaces weeks later in the contract review.

Aviation regulators acted on this mechanism in 2008. Medical research has been counting it since 2003. The boardroom hasn't.

Every cross-border meeting today defaults to a shared language nobody fully owns — almost always English. The arrangement feels like it works. The rest of this post is the evidence that it works less than people think, drawn from three independent literatures (aviation safety, medical interpretation, international-business research), and an honest engagement with the strongest published counterargument (the Foreign-Language Effect).

How many people in the room are operating in a second language?

Of the roughly 1.5 billion English speakers in the world, about 400 million are native; the remaining 1.1 billion learned English as a second or additional language (David Crystal, English as a Global Language, Cambridge University Press, 2nd ed., 2003; Ethnologue speaker estimates). In any cross-border business meeting the ratio is usually worse than the global average, because English is over-represented in professional contexts.

EF Education First publishes an annual English Proficiency Index covering more than 100 countries. "Very high proficiency" is defined there as the ability to "use nuanced language in social, professional, and academic situations." That is the bar most measured populations do not clear. The median country sits in the "Moderate" band; "High" and "Very High" are concentrated in a small set of Northern European and a few East Asian economies.

In practice, the average global business meeting is staffed mostly by people operating somewhere between B1 and C1 on the Common European Framework — fluent enough to function in routine conversation, not fluent enough to argue contract language, regulatory nuance, or technical edge cases without loss.

That is the population the rest of this post is about.

Failure mode 1 — False precision

Fluent-sounding non-native speech is the most expensive failure mode in this space, because it removes the signal that anything needs checking.

Lev-Ari and Keysar showed that listeners judge information delivered in a non-native accent as less credible, even when the content is identical and the speaker is reading from a script (Why don't we believe non-native speakers?, Journal of Experimental Social Psychology, 2010). The bias is automatic, robust, and one-directional.

The opposite half of the same bias is the more expensive one in business. When a non-native speaker produces a sentence that sounds fluent — grammar correct, intonation natural — listeners assume comprehension is symmetric: that the speaker understood the surrounding discussion as well as they appear to have produced their own sentence, and that the listener understood the speaker as well as they appear to have spoken. Both assumptions are routinely wrong.

Tenzer, Pudelko, and Harzing surveyed employees across 15 multinational teams and found that perceived language proficiency systematically inflated trust attributions, with the imputed competence regularly turning out to be misplaced (The impact of language barriers on trust formation in multinational teams, Journal of International Business Studies, 2014).

The mechanism in the room:

• The speaker emits a sentence whose grammar is correct and whose meaning is off by 15%.
• Listeners hear the fluency, not the 15%.
• Nobody asks the clarifying question, because the sentence "sounded fine."
• The 15% gap is written into the minutes as agreed.

A confidently wrong sentence is more dangerous than an obvious gap, because the gap earns a follow-up question; the false clarity gets minuted. This is false precision — the load-bearing cost of operating in shared imperfect English.

Failure mode 2 — Self-censorship of nuance

The person most likely to know the answer is often the person least equipped to express it in the working language.

Volk, Köhler, and Pudelko reviewed the cognitive-neuroscience literature on L2 processing in multinational corporations and reported measurable additional load on working memory, processing speed, and emotional regulation when professionals operate in a non-native language (Brain drain: The cognitive neuroscience of foreign language processing in multinational corporations, Journal of International Business Studies, 2014). The first thing to disappear under that load is nuance — the qualifiers, hedges, conditional clauses, and counter-arguments that a native speaker deploys without thinking.

The observable consequence is documented in the international-business literature: senior subject-matter experts who would dominate a meeting in their native tongue become the quietest people in the room when forced to operate in English (Tsedal Neeley, Global Business Speaks English, Harvard Business Review, 2012; Neeley, The Language of Global Success, Princeton University Press, 2017). They are not silent because they have nothing to add. They are silent because the cost of expressing their actual view in English — finding the right verb tense, hedging without sounding evasive, qualifying without sounding weak — exceeds the cost of staying quiet.

The decision is then made by whoever can express themselves fluently. Who are not, in general, the people who know most.

Hinds, Neeley, and Cramton called this language as a lightning rod: language proficiency becomes a proxy for status, and status determines who speaks (Language as a lightning rod, Journal of International Business Studies, 2014). The most knowledgeable person in the room becomes the least articulate.

Failure mode 3 — Fluency outranks authority

In any negotiation conducted in shared English, the native-English side holds an advantage before any substance is exchanged. The non-native side spends part of its cognitive budget on language production; the native side spends all of it on substance.

This is a measurable processing asymmetry, not a personality effect. The neuroscience review cited above reports working-memory penalties of roughly 20–30% on equivalent tasks when the same person operates in L2 versus L1. Translated into a live negotiation: the non-native side is running on perhaps 70% of its cognitive budget while the native side runs on all of it.

The visible consequence is not that the non-native side speaks less. The consequence is that the non-native side agrees more. Hedges drop. "I think we could probably consider" collapses into "OK." The native side gets the wording it wanted; the non-native side feels it got most of what it wanted; the gap only surfaces in the contract review weeks later.

This is the part of the language tax that is hardest to see in the moment and most expensive to fix afterwards.

The strongest evidence: where "good-enough" English literally kills

The three failure modes above are well-attested in research literatures, but skeptical readers can dismiss them as soft-science findings about feelings. The strongest evidence for the thesis comes from the two domains where second-language English has been measured against a hard outcome — and where regulators have already acted on what the data shows.

Aviation — when ICAO regulated the problem

Tenerife, 27 March 1977. A KLM 747 and a Pan Am 747 collided on the runway at Los Rodeos. 583 people died — still the worst accident in aviation history. The investigation identified several contributing causes, including a non-standard radio exchange. KLM Captain Van Zanten, operating in second-language English under heavy time pressure, told the tower "we are now, uh, at takeoff." The tower acknowledged with "OK." Van Zanten's phrasing was ambiguous between "we are in the takeoff position" and "we are in the process of taking off"; the Pan Am crew, also second-language English speakers, were still on the runway. Subsequent reforms by ICAO and national authorities introduced standardized phraseology specifically to remove this class of ambiguity (Spanish Aviation Authority final report; SKYbrary case study).

Avianca 052, 25 January 1990. A Boeing 707 ran out of fuel and crashed near Cove Neck, New York, killing 73. The flight crew, operating in second-language English with New York ATC, reported being "running out of fuel" — a phrase that does not exist in ICAO standard phraseology and that ATC did not interpret as a declared emergency. The standard term "fuel emergency" or "minimum fuel" was never used. The NTSB's final report identified the failure of the flight crew to use standard phraseology among the probable causes (NTSB Aircraft Accident Report AAR-91/04).

The regulatory response. In 2003, ICAO adopted Language Proficiency Requirements (LPRs) — Annex 1 (Personnel Licensing) and Annex 10 (Aeronautical Telecommunications) — making demonstrated English proficiency a licensing requirement for international flight crew and air traffic controllers worldwide. A six-level rating scale was defined; Level 4 ("Operational") is the minimum for international operations. Below Level 4, the licence is not valid for international flight. The implementation deadline was March 2008 (ICAO Doc 9835, Manual on the Implementation of ICAO Language Proficiency Requirements).

The regulator's reasoning here is the thesis of this post, written into binding international law: non-native English at "good-enough" level is unsafe in high-stakes communication; either everyone speaks the language at a defined operational standard, or they don't get to fly the airplane.

Medicine — when the comparison is measurable

The other domain with a hard outcome is medical interpretation, and here the comparison is even cleaner: same clinical encounter, same patient, with versus without a trained interpreter.

Glenn Flores and colleagues audio-recorded 57 emergency-department encounters with Spanish-speaking limited-English-proficient patients in two pediatric EDs and counted every interpretation error and its potential clinical consequence (Flores et al., Errors of medical interpretation and their potential clinical consequences: a comparison of professional versus ad hoc versus no interpreters, Annals of Emergency Medicine, 2012).

Headline findings:

• 1,884 interpretation errors identified across the 57 encounters.
• 18% of all errors had potential clinical consequences.
• The clinically consequential error rate dropped to 12% when a professional interpreter with ≥100 hours of training was used.
• The rate was 22% with professional interpreters with <100 hours of training, 20% with ad hoc interpreters (family members, untrained bilingual staff), and 20% with no interpreter at all.

Two things to read out of this. First: ad-hoc interpretation by a fluent-sounding bilingual person who is not a trained interpreter performs no better than no interpreter at all on clinical-consequence rates. The fluency does not translate into accuracy on the parts that matter. Second: the protective effect only kicks in once interpretation crosses a defined training threshold — exactly the same regulatory pattern as ICAO Level 4. "Good enough" is not a category that exists in this data. Either you cross the threshold or you don't.

Earlier work by the same author had already shown the same pattern in primary care (Flores et al., Errors in medical interpretation and their potential clinical consequences in pediatric encounters, Pediatrics, 2003). The 2012 study quantified it against a defined competency standard.

Why this matters in the boardroom

The boardroom is not the cockpit and it is not the ED. The stakes per minute are lower; the consequences arrive in months, not seconds; the cost is in dollars and reputation rather than lives.

But the mechanism is identical. A non-native speaker operating in a shared imperfect language emits fluent-sounding utterances whose semantic content has shifted; listeners receive the fluency as a signal of accuracy; the gap is not detected at the moment of emission; the gap surfaces only when the artifact built on top of it (the procedure, the dispatch instruction, the contract clause) is executed against reality and turns out to mean something different than the room believed.

Aviation and medicine matter for the boardroom because they are the two domains where the cost of running this mechanism was quantified, and a regulator agreed it was unacceptable. They are the natural experiment for the thesis.

The cost in business — where there is data

Business doesn't have a Tenerife or a Flores study with the same scientific cleanliness. What it has is survey data — self-reported, mixed-mechanism, but at scale.

The Economist Intelligence Unit, 2012. A survey of more than 500 senior executives across 51 countries — Competing across borders: How cultural and communication barriers affect business. Among the findings:

• 49% of respondents reported that misunderstandings have stood in the way of major international transactions, incurring significant losses for their company.
• 64% reported that differences in language and culture make it difficult to gain a foothold in unfamiliar markets.
• 67% reported that miscommunication is interfering with their international business efforts.

The report does not split "language miscommunication" from "cultural miscommunication" cleanly, but for the half of respondents reporting failed major transactions, the cost line is real even if the mechanism inside is mixed. The business literature lacks the controlled "with vs without interpreter" comparison that medicine has, and the regulatory backstop that aviation has — but the underlying mechanism is the one the cleaner literatures have already pinned down.

"But doesn't operating in L2 make people better decision-makers?"

The strongest published counterargument to this post's thesis is the Foreign-Language Effect, demonstrated by Keysar, Hayakawa, and An at the University of Chicago (The Foreign-Language Effect: Thinking in a Foreign Tongue Reduces Decision Biases, Psychological Science, 2012). Their experiments showed that when people considered classic decision-theory problems — framing effects, loss aversion, the Asian disease problem — in their second language, they exhibited less of the standard cognitive bias than when considering the same problems in their L1. Subsequent work has replicated the effect across multiple language pairs and decision-bias paradigms (see also Costa et al., Cognition, 2014).

This is a real finding. So why doesn't it refute the thesis?

Three reasons:

1. The Foreign-Language Effect is about individual reasoning, not multi-party communication. Keysar et al. measured what happens when one person decides a problem alone in their head, in their L2, with no listener and no communication partner. The three failure modes in this post are all properties of interaction — false precision and self-censorship and the fluency-authority gradient only exist because there is more than one person in the room. The Foreign-Language Effect doesn't enter that domain at all.
2. The effect operates on emotional/heuristic biases, not on semantic accuracy. The mechanism is that L2 creates emotional distance from the framing of the problem, so the listener falls back on more analytical (System 2) processing. Useful for reducing loss aversion. Not useful for transmitting a contract clause without semantic drift, which is what the boardroom is doing.
3. Aviation and medical regulators have already weighed both effects against each other. ICAO and the medical-interpretation literature both know the L2 reasoning literature exists. Neither field decided that the bias-reduction benefit was worth the false-precision cost. Both went the other way: define a proficiency standard, enforce it, and require interpreters or standardized phraseology for everyone below the standard.

The honest summary: operating in L2 makes you — alone, in your head — slightly more rational on certain framing tasks. It makes the room you are in less able to transmit information accurately. The Foreign-Language Effect is a property of individual cognition; the three failure modes above are properties of multi-party communication. They don't cancel.

Why "good-enough" machine translation does not fix any of it

A generic machine-translation layer slotted on top of an English meeting addresses none of the three failure modes, and in some configurations makes them worse:

• False precision compounds. A good-enough MT layer also emits fluent-sounding output, now layered on top of fluent-sounding non-native English. Two stacks of unverified fluency separate source intent from listener comprehension.
• Self-censorship persists. If the working language of the meeting is still English and translation only serves the listeners, the speakers still pay the L2 cost. They still drop the nuance. The translation pipeline preserves the loss faithfully.
• The fluency–authority gradient flips, but does not flatten. A poorly tuned translation layer just moves the advantage to whoever has the best engine in their corner, not to whoever knows most.

The fix is not "add translation on top of an English meeting." The fix is to remove the requirement that any participant operate in a language they do not fully own. That is a different architectural choice — and the one we're building toward.

What changes when every participant speaks their native language

The structural change is simple to state and hard to engineer:

1. Every participant speaks their own native language — no L2 cognitive load, no nuance drop, no self-censorship.
2. Every participant hears every other participant in their own native language, with sub-second latency and tone preserved.
3. The translation layer is auditable: source utterance, target utterance, per-language transcripts exported as a single bundle, with per-pair quality measured and published monthly on real traffic rather than asserted as a marketing number.

All three failure modes are linked to the same requirement — that someone in the room operate in a shared imperfect language. Removing that requirement removes them together. Adding translation on top of that requirement does not.

That is the difference the next class of cross-border meeting tooling is going to have to be measured against. Aviation regulators didn't accept "everyone speaks pretty good English" as the answer. Medical regulators didn't either. The boardroom should not be the last domain that does.

A reading list

The load-bearing sources for this post, in roughly the order they were used:

On the cognitive and social mechanism

• Lev-Ari, S., & Keysar, B. (2010). Why don't we believe non-native speakers? Journal of Experimental Social Psychology, 46(6), 1093–1096.
• Tenzer, H., Pudelko, M., & Harzing, A.-W. (2014). The impact of language barriers on trust formation in multinational teams. Journal of International Business Studies, 45(5), 508–535.
• Hinds, P. J., Neeley, T. B., & Cramton, C. D. (2014). Language as a lightning rod: Power contests, emotion regulation, and subgroup dynamics in global teams. Journal of International Business Studies, 45(5), 536–561.
• Volk, S., Köhler, T., & Pudelko, M. (2014). Brain drain: The cognitive neuroscience of foreign language processing in multinational corporations. Journal of International Business Studies, 45(7), 862–885.
• Neeley, T. (2012). Global Business Speaks English. Harvard Business Review.
• Neeley, T. (2017). The Language of Global Success. Princeton University Press.
• Keysar, B., Hayakawa, S. L., & An, S. G. (2012). The Foreign-Language Effect: Thinking in a Foreign Tongue Reduces Decision Biases. Psychological Science, 23(6), 661–668.

Aviation

• ICAO. Manual on the Implementation of ICAO Language Proficiency Requirements (Doc 9835). International Civil Aviation Organization.
• NTSB. Avianca, The Airline of Colombia, Boeing 707-321B, HK 2016, Fuel Exhaustion, Cove Neck, New York, January 25, 1990 (AAR-91/04).
• SKYbrary. Tenerife airport disaster, 1977 — accident case study.

Medicine

• Flores, G., Abreu, M., Barone, C. P., Bachur, R., & Lin, H. (2012). Errors of medical interpretation and their potential clinical consequences: a comparison of professional versus ad hoc versus no interpreters. Annals of Emergency Medicine, 60(5), 545–553.
• Flores, G., et al. (2003). Errors in medical interpretation and their potential clinical consequences in pediatric encounters. Pediatrics, 111(1), 6–14.

Business

• Economist Intelligence Unit. (2012). Competing across borders: How cultural and communication barriers affect business. Sponsored by EF Education First.

Population

• Crystal, D. (2003). English as a Global Language (2nd ed.). Cambridge University Press.
• EF Education First. English Proficiency Index (annual).

— The Mind.com Team

Aviation said it in 2008. Medicine has been counting it for twenty years. The boardroom is the last room still pretending that "good enough" is good enough.