AIB-C01 Exam Guide in Japanese vs. English: Finding, With AI, Where the Translation Changes the Meaning of 58 Skills

The AIB-C01 exam guide has a Japanese version, and AWS Certified AI Business Strategist (AIB-C01) can be taken in Japanese. But when the Japanese version we retrieved on November 7, 2026 is set side by side with the English original, “build-buy-partner decisions” becomes ビルドバイパートナーの決定 (birudo-bai-pātonā no kettei, “build-buy-partner decision” written as a single katakana word), and “appropriate cross-functional representation” becomes 部門横断的な適切な表現 (bumon ōdan-teki na tekisetsu na hyōgen, “appropriate cross-departmental expression”). The first makes a three-way choice between building, buying and partnering (our wording) look like one word; the second turns a point about having each department represented into a point about phrasing. If you take the exam in Japanese, which terms should you check, and how?
This article contains affiliate links. As an Amazon Associate, we earn from qualifying purchases.
This article is for anyone who plans to take AIB-C01 in Japanese, works with Japanese-speaking colleagues who will, or wants to use AI to check a localized exam guide against its source. We lined up all 58 skills in Japanese and English and built a table that rates every term as “changes meaning”, “hard to parse” or “no issue” (we call it the answer key in this article). We then ran two experiments with AI models as readers. Experiment 1 points at each problem term and asks the model to explain it, once with the Japanese alone and once with Japanese and English side by side. Experiment 2 hands the model all 58 skills in both languages at once and asks it to list the problem terms itself.
Here are the results up front. The answer key rates 5 of the 58 skills as “changes meaning”, 17 as “hard to parse” and 36 as “no issue”. In this run of Experiment 1, with the Japanese alone, claude -p gave an acceptable explanation 52 times out of 57 and gemma4:12b for 14 of 19 terms; with Japanese and English side by side, that rose to 57 out of 57 and 17 of 19.
In Experiment 2, claude -p explicitly flagged 4 of the 5 “changes meaning” skills as “changes meaning” in all three runs, and it flagged zero skills that the answer key rates “no issue” as “changes meaning”. But once “hard to parse” was included, it listed 23 to 25 of the 36 no-issue skills. The practical takeaway from this run: if you ask AI, ask only for terms whose meaning changes.
The ratings in the answer key are a proposal based on our own criteria, not the view of AWS. Claude built the answer key by reading the source text, and no human checked every row. We have not taken this exam, and everything said here about translation is limited to the Japanese version of the AIB-C01 exam guide. We cannot verify the wording of the actual exam questions, so we do not discuss them. For an overview of the certification and a preparation plan, see What Does the AWS AI Business Strategist Certification Prove? A 13-Week Plan to the February 15 Early-Adopter Badge (published 2026-11-16).
- What to check when you read the AIB-C01 exam guide in Japanese
- The rating scale: changes meaning, hard to parse, no issue
- Term-mapping table: Japanese terms to learn paired with the English
- Experiment 1: with the term pointed out, many explanations met the criteria even in Japanese alone
- Experiment 2: asking for “hard to parse” terms too pulled in no-issue skills
- A term the AI found: “ensure awareness of” in skill 1.1.6
- How to check the public AIB-C01 exam guide in both languages with AI
- Switching to English and extra time when you take the exam in Japanese
- What this test cannot tell you
- Summary: learn each translated term together with its English original
- Sources
What to check when you read the AIB-C01 exam guide in Japanese
The web version of the Japanese AIB-C01 exam guide is split into eight pages (the PDF has 18 pages): the overview, Content Domains 1 to 4, Technologies and concepts that might appear on the exam, and in-scope and out-of-scope services. The core of the exam scope is the four content-domain pages, which list 13 tasks and 58 skills. Skills are referred to by number, such as “2.1.2”, and this article uses the same numbers.
The guide itself carries no version date, so this article covers “the version retrieved on November 7, 2026”. The Japanese and English text of the 58 skills and 13 tasks we retrieved, 142 strings in total, also matched the pages when we fetched them again on the same day. Translations can be revised, so read everything below about Japanese wording as accurate as of November 2026.
The reason to check the wording is simple. If you start studying with a term misread, you may end up preparing for something other than the action the exam scope describes. Read a point about team structure as a point about phrasing, for example, and what you memorize changes. The checks come down to three things.
- Meaning of the term: read in its ordinary Japanese sense, does it point to the same object and action as the English?
- Word boundaries and what modifies what: can you see where one word ends and the next begins, and does each modifier attach to the same thing as in the English?
- How translations diverge: how is the same English term translated elsewhere in the guide, or in other AWS documents?
The goal of this article is not to judge the quality of the translation. It is to give readers what they need to decide “I will learn this skill through its English term.” The existence of an official Japanese version is itself a real help to people taking the exam in Japanese. With that in mind, we list the terms worth supplementing as you read.
How it was translated is unknown
Some Japanese AWS documents carry a machine-translation notice, but none appears on the eight web pages or in the PDF of this exam guide. The method is not stated officially, so we do not know it. In a 2022 blog post AWS explained that it uses machine translation for technical documentation and adds human post-editing for highly visible content, but that is a general description, not a statement about exam guides. So this article does not discuss how the guide was translated. It only looks at whether the finished wording reads with the same meaning as the English.
The exam guide uses the shared vocabulary of AWS AI certifications. If you want those core terms in English first, the Sybex study guide for AWS AIF-C01 covers them in print.
The rating scale: changes meaning, hard to parse, no issue
For each term, we asked one question: “Could a candidate who reads only the Japanese version answer the same thing as someone reading the English original?” We did not judge stiffness or naturalness of style. A skill takes the most severe rating among its terms. The three levels are defined as follows.
- Changes meaning: read in its ordinary Japanese sense, the term makes the skill point to a different object or action. This covers cases where the ordinary meaning of the Japanese word does not include the sense the English has in context, and cases where a wrong word boundary or part of speech creates a meaning that is not in the original
- Hard to parse: the object and action do not change, but the Japanese alone does not pin down the meaning. This covers katakana loanwords whose established Japanese sense differs from the context, words whose boundaries are invisible, modifiers that shift so the sentence can be read two ways, and words whose range differs from the English so that context is needed
- No issue: the meaning of the original is preserved and the Japanese alone is clear. Katakana loanwords that are established general IT or business vocabulary (deploy, monitoring, baseline, guardrail and so on) are not treated as problems
We set four supplementary rules. When an example in parentheses fills in the meaning, the rating drops one level (“business functions” in 2.1.1 is such a case). When the same term appears several times, each skill it appears in gets the same rating, but the term is counted once when counting terms. Rows that could also be read as the neighboring level are marked “borderline”, since another reader might rate them differently, and we give counts both with and without them. For no-issue rows we listed only one pair of core terms per skill, but the rating was made after reading the whole sentence.
Where the ratings come from, and the counts
Claude built the answer key by reading the source text, and the ratings are a proposal based on our criteria. They are not the view of AWS, and no human checked every row. The table was made before the experiments, and the creation timestamp recorded in the file holding the table is also earlier than the first model call. However, change history exists only from after the experiments, so the records cannot show that the table was not edited during them. We call it the “answer key” because it is used to score the experiments, not because its ratings are settled. Judgments about the ordinary meaning of words are also our own reading, not the result of checking every dictionary.
Counting again from the answer key with a script gives the following.
| What is counted | Changes meaning | Hard to parse | No issue | Total |
|---|---|---|---|---|
| 58 skills (rated by the most severe term) | 5 | 17 | 36 | 58 |
| Term-pair rows (including skills with two pairs) | 5 | 21 | 39 | 65 |
| 13 tasks (counted separately, not part of the 58 skills) | 1 | 1 | 11 | 13 |
The five “changes meaning” skills are 1.1.3, 2.1.2, 2.1.3, 3.2.1 and 4.4.1. Of these, 1.1.3 and 2.1.3 are borderline rows, so excluding borderline leaves three skills: 2.1.2, 3.2.1 and 4.4.1. Among the “hard to parse” rows, four are borderline: 1.1.5, “leadership alignment” in 4.1.1, 4.1.4 and 4.4.2.
Move the line and the numbers move too. Six of the 17 “hard to parse” skills (2.2.1, 4.2.3, 4.3.2, 4.4.4, 4.4.5 and 4.4.6) are rated that way only because of the katakana word for “initiative”. It is widely used in AWS documents, so if you count it as no issue, “hard to parse” drops to 11 and “no issue” rises to 42. Treat the counts as a rough guide and look at the actual terms in the tables below.
Term-mapping table: Japanese terms to learn paired with the English
Below is a reader-facing mapping table, term by term. The explanation column is our own explanation, not an official AWS translation. We do not reproduce the full text of the 58 skills, so read the guide pages for the complete exam scope. All Japanese terms are from the version retrieved on November 7, 2026, and everything said about them is as of November 2026. Each Japanese term is followed by its romanization and a literal English back-translation.
Terms that change meaning
| Skill | English term | Japanese version (romanization, literal back-translation) | Our explanation |
|---|---|---|---|
| 2.1.2 | build-buy-partner decisions | ビルドバイパートナーの決定 (birudo-bai-pātonā no kettei, “build-buy-partner decision”, written as one word) | Deciding among three options: build it yourself, buy it, or partner |
| 3.2.1 | appropriate cross-functional representation | 部門横断的な適切な表現 (bumon ōdan-teki na tekisetsu na hyōgen, “appropriate cross-departmental expression”) | Representatives of each relevant department are part of the governance structure |
| 4.4.1 | envision, experiment, launch, scale | 構想、実験、起動、スケーリング (kōsō, jikken, kidō, sukēringu, “concept, experiment, start-up, scaling”) | “Launch” is the stage after experimenting, when the chosen solution goes into production |
| 2.1.3 (borderline) | strategic alignment | 戦略的連携 (senryaku-teki renkei, “strategic partnership”) | An initiative is in line with the company strategy |
| 1.1.3 (borderline) | the relevance of data types for AI | AI におけるデータ型の関連性 (AI ni okeru dēta-gata no kanren-sei, “the relatedness of data types in AI”) | How different kinds of data matter to AI |
| Task 3.3 (outside the 58 skills) | direct mitigation strategies | 直接的な軽減戦略 (chokusetsu-teki na keigen senryaku, “direct [adjective] mitigation strategies”) | Direct, as in lead, the mitigation strategies (“direct” is a verb) |
Build-buy-partner (2.1.2). The English joins three words with hyphens, but the Japanese drops the hyphens and the result looks like one word. The English refers to choosing among three options: build it in-house, buy it, or team up with another company. Because the katakana “bai” can also be read as “by”, the phrase can even look like “building by a partner” (that reading is ours). The same term appears in the guide overview as ビルドバイパートナーの意思決定 (birudo-bai-pātonā no ishi kettei, “build-buy-partner decision-making”) and on the Technologies and concepts page (Japanese) as “build-buy-partner decisions about AI solutions”.
We could not find an official AWS Japanese translation for this three-word set. “A three-way choice between building, buying and partnering” is our own explanation. At the level of single words, the Japanese PDF of AWS CAF-AI translates build as 構築 (kōchiku, “construct”) and buy as 購入 (kōnyū, “purchase”). AWS CAF-AI (CAF-AI below) is the AWS Cloud Adoption Framework for Artificial Intelligence, Machine Learning, and Generative AI. Its English version is marked as being for reference, and the English HTML redirects to another page. The Japanese pages opened on November 7, 2026. As a hypothetical example, when a small 3D printing shop brings in AI to monitor its prints, comparing whether to build its own model, buy an off-the-shelf product, or work with another company is exactly this kind of decision.
Appropriate cross-functional representation (3.2.1). The problem lies with “representation”; the Japanese word for “cross-functional” is fine. The English means that representatives of each relevant department are part of the governance structure, but reading the Japanese word 表現 (hyōgen, “expression”) turns a point about structure into a point about wording. Elsewhere in the same guide, 4.3.2 translates represent as を代表する部門横断的なチームを作り (o daihyō suru bumon ōdan-teki na chīmu o tsukuri, “form a cross-functional team that represents”). The Japanese version of CAF-AI also renders a team with representation as 複数のビジネスユニットの代表者でチームをまとめることです (fukusū no bijinesu yunitto no daihyōsha de chīmu o matomeru koto desu, “assembling a team from representatives of multiple business units”). When you memorize it, simply note that in 3.2.1 hyōgen means representatives.
Launch (4.4.1). Launch is the stage after experimenting, when the chosen solution goes into production. Level 3 of the English maturity model likewise describes it as the stage of moving from proof of concept to production deployment. The Japanese 起動 (kidō, “start-up”) is the word for starting a machine or a program (our reading), so it reads awkwardly as the name of a stage. The EC2 documentation (Japanese) translates launching an instance as kidō, but that is a different context from 4.4.1.
The four words in the exam guide are identical to the four words in the image alt text of the Japanese version of the generative AI maturity model, which carries a machine-translation notice. In that page’s headings, only Level 1 stays in English as Envision; the other three words match. Elsewhere in Japanese AWS material, however, the translation of launch varies. The Japanese version of the AWS Cloud Adoption Framework (AWS CAF), which the guide overview lists among the material candidates should know, uses ローンチ (rōnchi, the English “launch” in katakana), while the Japanese version of CAF-AI uses 着手 (chakushu, “getting started”). Note, though, that the set of stages is not the same across these documents either.
| Document | English stages | Japanese stages (romanization, literal back-translation) |
|---|---|---|
| AWS CAF | Envision / Align / Launch / Scale | 構想/連携/ローンチ/スケール (kōsō / renkei / rōnchi / sukēru, “concept / collaboration / launch / scale”) |
| CAF-AI | Envision / Align / Launch / Scale | 構想/調整/着手/スケーリング (kōsō / chōsei / chakushu / sukēringu, “concept / coordination / getting started / scaling”) |
| Generative AI maturity model | Envision / Experiment / Launch / Scale | 構想/実験/起動/スケーリング (kōsō / jikken / kidō / sukēringu, “concept / experiment / start-up / scaling”); this is the image alt text, and the Level 1 heading reads Envision |
| AIB-C01 exam guide 4.4.1 | envision, experiment, launch, scale | 構想、実験、起動、スケーリング (kōsō, jikken, kidō, sukēringu, “concept, experiment, start-up, scaling”) |
The second stage is Align in CAF and Experiment in the maturity model, so the four-word sets differ. When you memorize the stage names, be careful not to mix up which document a set comes from.
Strategic alignment (2.1.3, borderline). “Strategic alignment” appears as one of the criteria for setting priorities and means that an initiative is in line with the company strategy. To us, 戦略的連携 (senryaku-teki renkei, “strategic partnership”) is the vocabulary of alliances between companies, so it reads as if a counterpart were involved. In the same guide, “align with” in Task 2.1 becomes ビジネス目標に沿った (bijinesu mokuhyō ni sotta, “in line with business goals”), and “leadership alignment” in 4.3.1 becomes リーダーシップの足並みを揃えて (rīdāshippu no ashinami o soroete, “with leadership moving in step”). On the other hand, the Japanese AWS CAF also translates Align as 連携フェーズ (renkei fēzu, “collaboration phase”), so there is precedent within AWS documents. We counted it as a borderline row that could also be read at the neighboring level.
Relevance of data types (1.1.3, borderline). The English asks the candidate to explain how kinds of data, such as structured and unstructured data, matter to AI. In our reading, データ型 (dēta-gata, “data type”) is the programming term for integer or string types, and 関連性 (kanren-sei, “relatedness”) leaves unclear what is related to what. The same guide’s Technologies and concepts page, with a different word order, translates Types of data as データの種類 (dēta no shurui, “kinds of data”). The Japanese AIF-C01 exam guide also translates types of data as さまざまな種類のデータ (samazama na shurui no dēta, “various kinds of data”). (We covered the AIF-C01 scope in AWS AI Practitioner AIF-C01: Complete Guide — 5 Domains and a 4-Week Roadmap (published 2026-06-08).) That said, the phrase “structured and unstructured data” just before it supplies context, so it can also be read as merely hard to parse.
Direct mitigation strategies (Task 3.3, outside the 58 skills). The English of Task 3.3 is “direct mitigation strategies”, and from that sentence alone it is hard to tell what part of speech “direct” is. The guide overview, however, describes the same content as “direct appropriate mitigation strategies”, and because “appropriate” follows “direct”, it is clearly a verb. The Japanese overview also renders it as a verb: 適切な軽減戦略を指示する (tekisetsu na keigen senryaku o shiji suru, “direct appropriate mitigation strategies”). Only the Japanese of Task 3.3 takes it as an adjective, so the same guide contains two different translations. We did not include it in the 58-skill counts and counted it once in the separate task tally.
Terms that are hard to parse
Here are the main “hard to parse” terms. Terms marked borderline are rows that could also be read at the neighboring level.
| Japanese version (romanization, literal back-translation) | Skill | English term | Our explanation |
|---|---|---|---|
| ディメンション (dimenshon, “dimension”), responsible AI | 3.1.1 | dimensions | The elements that make up responsible AI |
| ディメンション (dimenshon, “dimension”), evaluation | 4.1.1 | critical dimensions | Aspects to evaluate |
| イニシアチブ (inishiachibu, “initiative”) | 2.1.3 and 7 other skills (8 in all) | initiatives | AI efforts and projects |
| ワークフォース (wākufōsu, “workforce”) | 4.3.3, 4.3.5 | workforce | Employees and talent; workforce development means developing people |
| ボキャブラリー (bokyaburarī, “vocabulary”) | 1.1.6 | unified AI vocabulary | A shared system of terms everyone uses |
| ビジネス機能 (bijinesu kinō, “business function”) | 2.1.1 | business functions | Business functions or departments such as sales or R&D |
| ガバナンスバイデザイン (gabanansu-bai-dezain, “governance-by-design”, as one word) | 3.1.3 | governance by design | Building governance in from the start of planning |
| アクセシビリティ (akuseshibiriti, “accessibility”) | 4.2.1 | accessibility | The people who need data can reach it |
| 実装期限 (jissō kigen, “implementation deadline”) | 4.3.3 | implementation timelines | When and on what schedule something is rolled out |
| 開発投資 (kaihatsu tōshi, “development investment”), borderline | 4.1.4 | development investments | Investment that builds organizational capability |
| リーダーシップの連携 (rīdāshippu no renkei, “leadership collaboration”), borderline | 4.1.1 | leadership alignment | Leadership is in step on direction |
| 履歴データを使用して (rireki dēta o shiyō shite, “using historical data”), borderline | 1.1.5 | by using historical data | Modifies training a model on past data |
| 全社規模のデプロイに向けて構築を行う (zensha kibo no depuroi ni mukete kōchiku o okonau, “carry out building toward company-wide deployment”), borderline | 4.4.2 | build toward enterprise-wide deployments | Stack up small wins and expand to a company-wide rollout |
The three katakana loanwords for “dimension”, “workforce” and “initiative” are not unique to the AIB-C01 exam guide. The AWS responsible AI page (Japanese) and an AWS blog post (Japanese) use dimenshon, and the Japanese version of AWS CAF uses wākufōsu. The Japanese CAF-AI also contains パイロットイニシアチブ (pairotto inishiachibu, “pilot initiative”). People who read AWS documents are used to these words, but in Japanese alone they are hard to parse.
For “dimension”, keep the two uses apart. In 3.1.1 it is a responsible AI term, and the AWS responsible AI page translates the same “dimensions” both as dimenshon and as 要素 (yōso, “element”). In contrast, 重要なディメンション (jūyō na dimenshon, “important dimensions”) in 4.1.1 refers to aspects of evaluation, a different use from the responsible AI term. The examples in parentheses (data quality, governance frameworks and so on) show that it is a list of aspects.
In the guide, “initiative” is used in the sense of AI efforts or projects. However, the loanword paraphrase proposals of Japan’s National Institute for Japanese Language and Linguistics (NINJAL) (March 2006) paraphrase this word as 主導 (shudō, “taking the lead”) and 発議 (hatsugi, “proposing a motion”). The same entry adds that when citing a name of the form “X initiative”, it helps to explain it with words such as “X concept” or “X action plan” (its examples are political). NINJAL describes “accessibility” (spelled アクセシビリティー, akuseshibiritī, there) as “ease of use designed so that anyone can use it”, but in 4.2.1 it means that the people who need data can reach it. The same entry notes that depending on context it can be paraphrased as “ease of connection” or “ease of approach”, and the meaning in 4.2.1 is closer to those (our reading). On top of that, in the English “effects” attaches only to data silos, while the Japanese can be read as attaching to all three items.
Bokyaburarī is used for the size of an individual’s vocabulary (our reading), and the same guide translates “define terminology” in Task 1.1 as 用語を定義する (yōgo o teigi suru, “define terms”). For jissō kigen, “timeline” refers to the whole schedule, but 期限 (kigen) reads only as a deadline. For the historical-data phrase, the English attaches it to training, but the Japanese reads as if it attaches to “explain” at the end of the sentence. In gabanansu-bai-dezain the word boundaries are invisible, and we found no Japanese AWS page that explains the term.
Places where the same English term is translated differently
Terms translated differently within the same guide are useful clues for rereading. The main ones:
- alignment: 2.1.3 戦略的連携 (senryaku-teki renkei, “strategic partnership”), 4.1.1 リーダーシップの連携 (rīdāshippu no renkei, “leadership collaboration”), 4.3.1 リーダーシップの足並みを揃えて (rīdāshippu no ashinami o soroete, “with leadership moving in step”)
- represent / representation: 3.2.1 部門横断的な適切な表現 (bumon ōdan-teki na tekisetsu na hyōgen, “appropriate cross-departmental expression”), 4.3.2 を代表する部門横断的なチームを作り (o daihyō suru bumon ōdan-teki na chīmu o tsukuri, “form a cross-functional team that represents”)
- direct: Task 3.3 直接的な軽減戦略 (chokusetsu-teki na keigen senryaku, “direct [adjective] mitigation strategies”), overview 適切な軽減戦略を指示する (tekisetsu na keigen senryaku o shiji suru, “direct appropriate mitigation strategies”)
- data types / types of data: 1.1.3 データ型 (dēta-gata, “data type”), Technologies and concepts page データの種類 (dēta no shurui, “kinds of data”); the English word order differs
- human oversight: 3.1.4 人による監視 (hito ni yoru kanshi, “monitoring by humans”), 4.3.6 人による監督 (hito ni yoru kantoku, “supervision by humans”); this is an example where the meaning does not change
Experiment 1: with the term pointed out, many explanations met the criteria even in Japanese alone
Setup
We used two models as readers. One was claude -p (Claude Code 2.1.319, --model sonnet, no tools and no MCP); the recorded model ID was claude-sonnet-5-5. The other was gemma4:12b running locally on Ollama 0.40.15 (think: false, temperature 0, context length num_ctx of 4096 in Experiment 1 and 16384 in Experiment 2). claude -p does not let you set the temperature, so the temperature conditions are not matched. Both used the same system prompt (translated from Japanese): “You help people study for certification exams. Answer only in the format instructed.” claude -p was run 3 times and gemma4:12b once, all on November 7, 2026.
We used 25 terms. Nineteen were problem terms: the 5 “changes meaning” terms from the answer key, 13 “hard to parse” terms, and the Task 3.3 term. The other 6 were negative controls, terms that should be easy to understand, mixed in to see whether reading failures concentrate on the problem terms. They were the Japanese renderings of context window, baseline metrics, leading indicators, bias drift, AI champion and center of excellence.
There were two conditions: giving the Japanese sentence alone, and giving the Japanese and English sentences side by side. In both, we pointed at the term and asked in the following form (translated from the Japanese prompt). We did not say it was an AWS exam.
The following is a Japanese sentence describing the scope of a certification exam.
Sentence: (one Japanese sentence from the guide)
What, concretely, does "(term)" in the Japanese sentence refer to? Explain in two sentences or fewer so that a first-time reader can understand.
Return only the following JSON.
{"meaning": "explanation"}
In the side-by-side condition, we added the English sentence and noted that the Japanese sentence was a translation of it. Each answer was judged pass or fail against “criteria for an acceptable explanation” and “examples of unacceptable explanations” set for each term before the experiment. For example, build-buy-partner had to name all three options, and launch had to be explained as the stage after experimenting when the chosen solution goes into production. After a first sort by keywords, Claude Opus 5.5, which we use in producing this site, read all 200 answers and made the final pass/fail calls.
Results
| Reader | Condition | Problem terms (including borderline) | Problem terms (excluding borderline) | Negative controls |
|---|---|---|---|---|
claude -p (3 runs) | Japanese only | 52 of 57 | 34 of 39 | 18 of 18 |
claude -p (3 runs) | Japanese and English | 57 of 57 | 39 of 39 | 18 of 18 |
| gemma4:12b (1 run) | Japanese only | 14 of 19 terms | 10 of 13 terms | 6 of 6 terms |
| gemma4:12b (1 run) | Japanese and English | 17 of 19 terms | 12 of 13 terms | 6 of 6 terms |
We calculated the figures excluding borderline terms ourselves. The six borderline terms are strategic alignment, relevance of data types, development investments, leadership alignment, “using historical data” and “build toward enterprise-wide deployments”, so that leaves 19 − 6 = 13 terms, or 13 × 3 = 39 answers for claude -p. From those we subtracted the failed answers on non-borderline terms. For claude -p with Japanese only, all 5 failures were outside the borderline set, giving 39 − 5 = 34. For gemma4:12b, 3 of the 5 terms it failed with Japanese only were outside the borderline set, giving 13 − 3 = 10 terms; side by side, 1 of its 2 failures was, giving 13 − 1 = 12 terms.
The negative controls passed in every condition, so the failures with Japanese alone were concentrated on the problem terms. Putting the English alongside also increased the number of acceptable explanations. Still, the fact that many terms were explained correctly from Japanese alone should not be taken to mean “the Japanese version is readable”. An AI being able to read it does not mean a person can. The reason lies in the content of the answers, covered next.
What the failed answers said
The interesting part is the failed explanations. For build-buy-partner, claude -p explained it as a two-way choice in 2 of its 3 Japanese-only runs. Here is the first answer (translated from Japanese):
“Build-bai-partner” is a translation of the English “build vs. partner” and refers to the decision of whether to develop AI capabilities in-house (build) or to achieve them by partnering or collaborating with outside companies or vendors (partner).
The option to buy has disappeared. gemma4:12b explained it as “the collaborating party when partnering with an outside company or specialist organization” (translated). In both cases the model failed to pull three options out of what looks like a single word. With the English alongside, all three claude -p runs and gemma4:12b described the three-way choice of building in-house, buying, or teaming up with another company.
For “appropriate cross-functional representation”, gemma4:12b, given Japanese only, explained it as “standardizing shared terms and definitions” (translated). That reads hyōgen as a matter of wording, the same direction as the phrasing-related “example of an unacceptable explanation” we had set for this term. In contrast, claude -p explained it in all three Japanese-only runs as, in effect, “a state in which representatives are participating”, which met the criteria. The word “representation” does not appear in its answers, but the substance of the explanation is the English meaning.
This is why AI results cannot be applied to people. A model can guess the English source term. In fact, one Japanese-only answer wrote out “build vs. partner” in English. There is no basis for assuming that a candidate reading only the Japanese could read it the same way.
For “direct mitigation strategies”, both models read the Japanese alone as measures that act directly. In all three runs claude -p gave explanations along the lines of “lower it directly” or “reduce it by acting on it directly”, and gemma4:12b said “directly suppress the impact” (both translated). With the English alongside, claude -p read “direct” as a verb in all three runs and pointed out the translation problem on its own. The first half of its first answer (translated from Japanese):
In the English “direct mitigation strategies”, “direct” does not mean “direct” as an adjective; it is a verb meaning “to command and lead (the implementation of mitigation measures)”.
gemma4:12b, on the other hand, explained it as “taking measures directly” (translated) even with the English alongside, and failed the criteria. We gave it only the one Task 3.3 sentence, not the overview sentence that shows “direct” is a verb.
Among the borderline terms, gemma4:12b explained “relevance of data types” as structured and unstructured data complementing each other, and explained “development investments” as investment in software or systems both with Japanese only and side by side. The English “development” alone also has a broad range, and reading it as investment that builds organizational capability is our judgment. Treat this as a row where our judgment and another reader’s split.
Conversely, some terms passed. For “initiative”, all three claude -p runs and gemma4:12b explained it from the Japanese alone as efforts or projects. For “accessibility”, both models also described it as reaching data. Even so, in our counts “initiative” stays rated as hard to parse.
Experiment 2: asking for “hard to parse” terms too pulled in no-issue skills
Setup
In Experiment 2 we did not point at any term. We gave the model the English and Japanese of all 58 skills and the definitions of the three rating levels at once, and asked it to list only the skills containing a “changes meaning” or “hard to parse” term, in the form term, rating, reason. Readers and number of runs were the same as in Experiment 1, and we did not give the answer key. We compared the AI answers with the answer key at the skill level. In some runs the model returned terms labeled “no issue” against instructions; those were not counted.
Results
First, the number of skills the AI said “change meaning”. The number on the left is the sum of the three columns to its right.
| Reader and run | Skills flagged “changes meaning” | Of which the answer key also says changes meaning (out of 5) | Of which the answer key says hard to parse | Of which the answer key says no issue |
|---|---|---|---|---|
claude -p run 1 | 9 | 4 | 5 | 0 |
claude -p run 2 | 7 | 4 | 3 | 0 |
claude -p run 3 | 7 | 4 | 3 | 0 |
| gemma4:12b, 1 run | 8 | 3 | 3 | 2 |
Next, the number of skills flagged as either “changes meaning” or “hard to parse”.
| Reader and run | Skills flagged (out of 58) | Answer key: changes meaning (out of 5) | Answer key: hard to parse (out of 17) | Answer key: no issue (out of 36) |
|---|---|---|---|---|
claude -p run 1 | 47 | 5 | 17 | 25 |
claude -p run 2 | 43 | 5 | 15 | 23 |
claude -p run 3 | 44 | 5 | 16 | 23 |
| gemma4:12b, 1 run | 13 | 4 | 5 | 4 |
The 4 skills claude -p explicitly flagged as “changes meaning” in all three runs were 2.1.2, 2.1.3, 3.2.1 and 4.4.1. The remaining one, 1.1.3, it rated “hard to parse” all three times. Since 1.1.3 is a borderline row, if you limit it to the 3 non-borderline skills, it flagged all of them as “changes meaning” every time. It flagged zero no-issue skills as “changes meaning” in all three runs.
That zero, however, is counted at the skill level. In run 1 the model flagged “tangible benefits” and “intangible benefits” in 2.2.1 as “changes meaning” (the Japanese renders them literally as “visible” and “invisible” benefits). The answer key’s 2.2.1 row rates “intangible benefits” as no issue and has no row for “tangible benefits”. Skill 2.2.1 as a whole is hard to parse because of “initiative”, so in the tally it falls under “answer key: hard to parse”. Looked at term by term, it includes terms that disagree with the answer key.
The model flagged 5, 3 and 3 “hard to parse” skills as “changes meaning”. Broken down: 4.3.3 (implementation timelines) in all three runs, 4.1.1 (leadership alignment) and 4.1.4 (development investments) twice each, and 4.4.2 once. The other 3 were 2.2.1 above, once, and 1.1.6, covered in a later section, twice. The 4.1.1 term, 4.1.4 and 4.4.2 are all borderline rows, terms we noted could also be read at the neighboring level.
Widening to “hard to parse” puts 43 to 47 of the 58 skills on the list
The problem is in the second table. claude -p listed 43 to 47 of the 58 skills, including 23 to 25 of the 36 skills the answer key rates no issue. With that many on the list, checking them one by one takes about as much effort as reading the whole guide.
Moreover, the listed terms included ones the model had explained without trouble in Experiment 1. In Experiment 2, claude -p flagged bias drift, context window and center of excellence as “hard to parse” in all three runs. These three are among the six negative controls in Experiment 1, where the same claude -p explained all six from Japanese alone and passed all 18 times. Because it flags terms it can explain itself, the number of flagged terms does not directly reflect how hard they are to read.
That said, the AI calling a term hard to parse when the answer key says no issue is not necessarily an AI error. How easy something is to read depends on the reader’s experience. What we can say is that even when given the rating criteria, the AI and the answer key did not draw the “hard to parse” line in the same place.
gemma4:12b listed 13 skills in its one run. Of the 5 “changes meaning” skills in the answer key, it explicitly flagged 2.1.3, 3.2.1 and 4.4.1 as “changes meaning”, rated 2.1.2 “hard to parse”, and did not list 1.1.3. It flagged two no-issue skills as “changes meaning”: “appropriately balance trade-offs” in 3.1.2 and リーダーシップの足並みを揃えて (rīdāshippu no ashinami o soroete, “with leadership moving in step”) in 4.3.1. The latter is a rendering of alignment that the answer key rates no issue. With a single run and different temperature settings, model differences cannot be separated from run-to-run variation.
What this means for readers: if you ask AI, narrow it to “changes meaning”
Within the scope of this run, our recommendation is that if you ask AI, look only at the terms whose meaning changes. With claude -p, that narrows 58 skills to 7 to 9, includes 4 of the answer key’s 5 every time, and includes no no-issue skills. That is a number a reader can check by hand.
But the AI not flagging something does not mean there is no problem. 1.1.3 never appeared in the “changes meaning” column, and gemma4:12b’s column included two no-issue skills. An AI’s list is a set of candidate terms to check, not a conclusion.
A term the AI found: “ensure awareness of” in skill 1.1.6
Experiment 2 also surfaced a term outside the answer key: “Ensure awareness of” in skill 1.1.6, rendered in Japanese as 確実に認識する (kakujitsu ni ninshiki suru, “reliably recognize”). The answer key looked only at the vocabulary term in 1.1.6 and missed this phrase at the end of the sentence. claude -p flagged it in all three runs, twice as “changes meaning” and once as “hard to parse”. gemma4:12b also flagged it as “changes meaning” in its one run. Here is the reason claude -p gave in run 1 (translated from Japanese):
Ensure awareness means making sure that others are informed and understand, but in Japanese it reads as the person themselves making sure to recognize something, so the object changes.
gemma4:12b also explained it as meaning “make sure people are aware (make it widely known)” (translated), pointing in the same direction. claude -p in run 3, on the other hand, wrote that “it becomes unclear whether the action is knowing it yourself or making the organization aware” (translated) and rated it “hard to parse”.
The AI’s point has a basis, but the English alone does not settle which reading is right. “Ensure awareness of” can mean making others aware, or the candidate knowing it themselves. Our view is “hard to parse”. We did not add this term as its own row in the answer key. It is, however, included in the skill-level counts of Experiment 2, as a skill the AI flagged.
As a study tip, simply note that 1.1.6 has two possible readings: being able to explain the frameworks yourself, and informing people in your organization. That the AI caught a term the Claude that built the answer key had missed is also one reason to use AI in the process of combing through terms. But the Claude used as a reader is from the same model family as the Claude that built the answer key, so the validity of what it flags needs to be checked separately.
How to check the public AIB-C01 exam guide in both languages with AI
Here are the experiment results turned into a procedure for readers. Giving the AI the official material first and then asking is the same order we followed in How Much Does AI Know About the Agentic Architect Exam Scope? 28 Tools Checked Against Official Docs (published 2026-10-20). That article measured product knowledge, while this one checks translated terms, so what is being measured differs.
- Open the guide in both languages: the Sources section at the end lists the English and Japanese pages for the overview, Content Domains 1 to 4 and Technologies and concepts, plus the English and Japanese PDFs of the whole guide. For reading, the pages or the PDF are fine. For feeding an AI, the Markdown version you get by changing the
.htmlat the end of a page URL to.mdis easier to work with - Pair them by skill number: list number, English, then Japanese. In Experiment 2 we gave all 58 skills at once
- Ask the AI: use the prompt below to have it list the “changes meaning” terms
- Repeat the same request a few times: separate terms that come up every time from those that come up once. In this run, the skills
claude -pflagged as “changes meaning” were 2.1.2, 2.1.3, 3.2.1, 4.3.3 and 4.4.1 in all three runs, and 2.2.1 and 4.4.2 in only one (4.1.1, 4.1.4 and 1.1.6 came up twice) - Check them yourself: read each flagged term against the English original, and look at how it is translated in other skills of the same guide and in other AWS documents
- Build a term table: write the English term, the Japanese term and your own explanation as a set, and keep it at hand while you study
Here is an example prompt (translated from Japanese; use it in the language you study in). It is a trimmed version of the Experiment 2 prompt: we removed the definitions of “hard to parse” and “no issue” (including the rule that established katakana loanwords are not problems) and the notes on the two cases of “changes meaning”, and changed the answer format from JSON to one sentence. We have not run it in this exact form. If you want something closer to what we ran, ask for both levels as in Experiment 2 and read only the “changes meaning” column.
Below are the English original and the Japanese version of the skills in a certification exam's scope.
For each term, judge: "Could a candidate who reads only the Japanese version answer the same thing as someone reading the English original?" Do not judge stiffness or naturalness of style.
Changes meaning: read in its ordinary Japanese sense, the term makes the skill point to a different object or action.
(Paste the skills from the public exam guide here, paired as number, English, Japanese.)
List only the skills that contain a "changes meaning" term. For each, give the skill number, the Japanese term, the English term, and the reason in one sentence.
Do not paste official questions or study materials into AI
Give the AI only the public exam guide and material you made yourself. Section 1.6 of the AWS Certification Program Agreement treats exam materials as AWS confidential information and prohibits copying or transmitting them. The definitions in section 6 include study materials in exam materials. We do not judge whether the official practice questions fall under that definition, but the procedure in this article does not paste the text of official questions or study materials into AI.
Section 2.2 contains the exam rules, and (a) prohibits, among other things, using unauthorized AI, machine learning, language models or automation tools. Use AI only for study before the exam. This article does not offer an interpretation of the agreement. If in doubt, read the agreement itself and confirm with AWS.
Checking a guide with AI works best when the request is precise. This O’Reilly book is a practical guide to writing prompts for large language models.
Switching to English and extra time when you take the exam in Japanese
Separately from checking terms, here is how things work when you take the exam in Japanese. As of November 2026, the AIB-C01 certification page lists English and Japanese as exam languages. We do not know whether the terms in the exam questions match those in the guide, so this section covers only mechanisms set out in AWS policy.
Switching to view questions in English during the exam
The AWS Before Testing policy says that candidates who register for a localized exam can also use a toggle to view questions in English during the exam. In other words, if you register in Japanese, you can switch during the exam to see questions in English. The same sentence appears on the exam scheduling page.
However, both sentences sit in the section on online proctoring. We found no sentence that directly addresses whether you can switch when testing at a test center. The Before Testing policy says you can try the toggle in the AWS Exam Demo, and the demo is reached from the Pearson VUE AWS page. We have not checked what the demo contains.
The official condition for 30 extra minutes is “taking it in English”
Separate from the toggle, there is an accommodation for English as a second language (ESL), called “ESL +30”. The official condition for the 30-minute extension is that a candidate whose first language is not English takes the exam in English. The AWS Certification FAQ also splits the choice of exam language into two paths and places ESL only on the English side. You request it once before registering for an exam, and it applies to later registrations. To do so, choose ESL +30 MINUTES in the accommodation request in your AWS Certification account.
We found no sentence stating directly whether the extension applies when you take the exam in Japanese. The official condition is taking it in English. Nor is there any sentence saying that switching to read English earns an extension. Laid side by side, the reader’s options compare as follows.
| Take it in English | Take it in Japanese | |
|---|---|---|
| Question language | English | Japanese; you can switch during the exam to view questions in English |
| 30 extra minutes | Candidates whose first language is not English get the extension on request | The official condition is taking it in English |
| Check beforehand | Request the accommodation before registering for the exam | No sentence found that directly addresses switching at a test center |
We do not say which is better. Whether to take it in English with 30 extra minutes, or in Japanese with the English view as a backup, depends on your English reading and your preparation. Building a term table helps your study either way.
About the beta exam
As of November 2026, AIB-C01 is a beta exam. The Before Testing policy lists the differences between beta and standard exams, but neither the toggle nor ESL is on that list. The beta rules state no special treatment. Since no sentence explicitly applies these to beta exams or excludes them, confirm with AWS before you register.
Proctor languages and comments
Separate from the language of the questions is the language of the proctor for online proctoring. Japanese-speaking proctors are available Monday to Saturday, 9:00 to 16:00 Japan time, while English is available 24/7. Booking an online-proctored exam involves interacting with a proctor, so keep this in mind when choosing a time slot.
According to the During Testing policy, you can also leave comments on translation accuracy during the exam. Time spent commenting counts against your exam time, however. AWS recommends flagging questions and commenting at the end of the exam. Individual replies are not provided.
What this test cannot tell you
A summary of the limits to keep in mind when reading the results.
- Claude built the answer key: the ratings are a proposal based on our criteria, and no human checked every row. The only terms where our judgment of ordinary meaning is backed by a dictionary or survey are those covered by the NINJAL proposals. Another reader might rate the borderline rows differently
- Limited record that the answer key came before the experiments: the creation timestamps on the files holding the answer key and the per-term pass criteria are earlier than the first model call. Change history exists only from after the experiments, so the records cannot show that the content was the same before and after
- The Claude reader is from the same family: the model that built the answer key and the
claude -preader are from the same model family, and Claude Opus 5.5 also made the pass/fail calls in Experiment 1. In Experiment 2, the AI was given the same rating criteria as the answer key. Agreement between the answer key and the AI may partly reflect the same family reading the same criteria - AI reading the Japanese alone does not mean people can: models can guess the English source terms. This article did not study how candidates read the guide. How easy something is to read depends on the reader’s experience
- The translation of the actual exam is unknown: this article covers only the Japanese version of the exam guide, and we do not know whether the exam questions use the same wording. Nor does it indicate any trend in exam questions
- Translations may be revised: everything here concerns the version retrieved on November 7, 2026, and the guide has no version date. Compare with the current pages when you read
- Few runs and unmatched conditions:
claude -pran 3 times and gemma4:12b once, so the numbers are values from this run. gemma4:12b used temperature 0 whileclaude -pcannot set temperature, so the two models differ in more than run count. We did not try other models or other ways of asking - Experiment 1 pointed at the terms: the AI did not find the terms itself. Experiment 2 looks at the ability to find terms, but its prompt only took the form of asking for both “changes meaning” and “hard to parse”
- The translation method is unknown: nothing can be said about the method from these results
- Cost is what this run recorded: the recorded cost of 153
claude -pcalls was $0.58, and how it is billed depends on your plan
Summary: learn each translated term together with its English original
- Checking the Japanese AIB-C01 exam guide against the English original across 58 skills, by our criteria 5 skills “change meaning” (3 excluding borderline), 17 are “hard to parse” and 36 have “no issue”. The ratings are a proposal, not the view of AWS
- In Experiment 1, where models explained terms we pointed at, Japanese alone gave acceptable explanations 52 times out of 57 for
claude -pand for 14 of 19 terms for gemma4:12b; side by side with English, that became 57 out of 57 and 17 of 19. But an AI being able to read it does not mean a person can - In Experiment 2, where models combed through the guide, limiting
claude -pto “changes meaning” narrowed 58 skills to 7 to 9 and included no no-issue skills. Widening to “hard to parse” put 43 to 47 on the list - The AI also flagged a term outside the answer key (“ensure awareness of” in 1.1.6), but the English alone does not settle which reading is right
- If you register in Japanese, you can switch during the exam to view questions in English. The official condition for the 30-minute extension is taking the exam in English; confirm with AWS before registering how switching works at test centers and whether these rules apply to the beta
If you take the exam in Japanese, start by pairing three terms with their English: birudo-bai-pātonā (build-buy-partner), bumon ōdan-teki na tekisetsu na hyōgen (appropriate cross-functional representation) and kidō (launch). Then give an AI the public AIB-C01 exam guide in both languages, have it list the “changes meaning” terms, and read each one against the original yourself. The AI’s list is a set of candidates to check; the final judgment is yours. Keep a term-mapping table at hand before you start studying.
Sources
- AIB-C01 exam guide: overview (English)
- AIB-C01 exam guide: overview (Japanese)
- AIB-C01 exam guide: Content Domain 1 (English)
- AIB-C01 exam guide: Content Domain 1 (Japanese)
- AIB-C01 exam guide: Content Domain 2 (English)
- AIB-C01 exam guide: Content Domain 2 (Japanese)
- AIB-C01 exam guide: Content Domain 3 (English)
- AIB-C01 exam guide: Content Domain 3 (Japanese)
- AIB-C01 exam guide: Content Domain 4 (English)
- AIB-C01 exam guide: Content Domain 4 (Japanese)
- AIB-C01 exam guide: Technologies and concepts that might appear on the exam (English)
- AIB-C01 exam guide: Technologies and concepts that might appear on the exam (Japanese)
- AIB-C01 exam guide PDF (English)
- AIB-C01 exam guide PDF (Japanese)
- AWS Certified AI Business Strategist certification page (English)
- AWS Certified AI Business Strategist certification page (Japanese)
- AWS CAF: Your cloud transformation journey (English)
- AWS CAF: Your cloud transformation journey (Japanese)
- AWS CAF: People perspective (Japanese)
- CAF-AI: Your AI transformation journey (Japanese)
- CAF-AI: Governance perspective (Japanese)
- CAF-AI PDF (English; marked as being for reference)
- CAF-AI PDF (Japanese)
- Generative AI maturity model: levels (English)
- Generative AI maturity model: levels (Japanese; carries a machine-translation notice)
- Generative AI maturity model: Level 3: Launch (English)
- EC2 User Guide: Launch an EC2 instance (Japanese)
- AIF-C01 exam guide: Domain 1 (English)
- AIF-C01 exam guide: Domain 1 (Japanese)
- AWS responsible AI page (English)
- AWS responsible AI page (Japanese)
- Considerations for addressing the core dimensions of responsible AI for Amazon Bedrock applications (AWS blog, Japanese)
- AWS KMS cryptographic details reference (Japanese; an example of the machine-translation notice)
- AWS Localization uses Amazon Translate to scale localization (AWS Machine Learning Blog, 2022-08-17)
- AWS Certification policies: Before Testing (English)
- AWS Certification policies: Before Testing (Japanese)
- AWS Certification policies: During Testing (English)
- AWS Certification policies: During Testing (Japanese)
- AWS Certification exam scheduling (Japanese)
- AWS Certification FAQ (English)
- AWS Certification FAQ (Japanese)
- AWS Certification Program Agreement
- Pearson VUE AWS Certification page (English)
- NINJAL Loanword Committee, Loanword paraphrase proposals, rounds 1 to 4, compiled edition (March 2006, PDF, Japanese)





