Cultural And Linguistic Blind Spots
It is fluent in your language. It is not fluent in your market.
This page covers one specific way AI gets things wrong at work, and what to do about it.
It runs in order. What goes wrong, why it happens, where you'd notice it on an ordinary day, who takes the blame, roughly what it costs, and the check that catches it. Then one thing to try this week.
The dollar figures are estimates, not measurements. The assumptions behind each one are printed right there, so you can swap in numbers that fit your job. Anything actually measured carries an OBSERVED tag.
What is cultural and linguistic blind spots?
Training data is unevenly distributed across languages and cultures, weighted heavily toward English-language and Western-internet content. Capability follows the data.
Two failures:
1. Uneven language quality. Performance in lower-resource languages is materially weaker, and the output still reads confidently to a non-speaker. You cannot assess quality in a language you don't have.
2. Cultural defaults. Where norms differ — directness, formality, hierarchy, how apology and obligation are expressed, what constitutes a complete answer — the model applies the defaults of its dominant training distribution. Output can be linguistically perfect and pragmatically wrong: too direct, insufficiently deferential, missing an expected acknowledgment.
The second failure is the expensive one in business contexts, and it is completely invisible in review unless a native speaker with domain context reads it.
What do people assume?
That multilingual capability implies cultural competence — that a system producing correct Japanese produces appropriate Japanese for a formal customer apology from a vendor to an enterprise buyer.
Grammatical correctness and situational appropriateness are different competencies, and the gap between them is invisible to a reader who doesn't speak the language.
Where does it show up at work?
A support organization deploys AI-drafted responses across eight markets, reviewed by a team that reads three of them. Responses in the other five are grammatically fine.
In two markets, the register is wrong for a vendor addressing a customer after a service failure — direct where deference is expected, and missing the acknowledgment the situation requires. Customers don't complain about tone. They escalate, or they don't renew, and the reason never enters the CRM.
Who carries the downside?
Vendor: none. Executive: sees a retention number with no explanation. Manager: owns the region. You: deployed the workflow. The cause is invisible in every dashboard the company has.
What does it cost?
[MODELED — not reported]
ASSUMPTIONS Customer interactions in markets without native review: 4,000 / year Rate w/ material register / cultural mismatch: 8% (320) Rate affecting relationship: 15% (~48) Cost per damaged relationship: $1,500 – $15,000 (churn risk, expansion forgone)
Annualized exposure: ~$72,000 – $720,000
Note the structural problem: this exposure is invisible in standard reporting. There is no ticket category for "the tone was wrong." Which means it is systematically under-prioritized relative to its size.
How do you control for it?
Native-speaker review with domain context — not translation checking. The reviewer must know both the language and the business situation, and the question is not "is this correct" but "would a vendor say this to a customer here, in this situation?"
Where full review isn't feasible: native review of templates and registers rather than every instance, plus a sampling program.
CONTROL COST Template / register review: 40 templates × 30 min Ongoing sampling: 4 hrs / month Annual: 68 hours Fully loaded rate: $70 / hour
Annualized control cost: $4,760
What should you do this week?
RECOMMENDATION
Identify every market where your organization sends AI-assisted communication without a native speaker in the review path. List them.
That list is a risk register nobody has written, in a category nobody is measuring, and the argument for closing it is a revenue argument rather than a safety argument — which is why it will actually get funded. Bring the list, not the concern.
Evidence
RESEARCH Performance disparity across languages by training-data representation.
RESEARCH Cultural value defaults in model output reflecting training distribution.
ANALYSIS The "invisible in standard reporting" argument is the author's and explains why this mode is under-prioritized. Retain the label.
This is one of 24 failure modes. The book gives you all of them — plus the controls that catch each one and a 90-day plan to prove you ran them.
Preorder the Book