What the JDLA G-Test Proves: Pass Rates, Exam Format, and a 40-Day Plan Built on AI

Your employer asked you to sit the G-test. Or you kept reading about AI until you noticed that your own understanding was a pile of fragments. Either way, the first questions are the same. What does the G-test actually certify, how long does it take to earn, and can the AI you already use every day be turned on the studying itself.

This article answers those three using only primary material from the JDLA, the Japan Deep Learning Association. Candidate counts and pass rates come from the association announcements, the exam format and calendar from the official pages, and the study-hour figure from the JDLA own survey of people who passed. Every number was taken on 21 September 2026, while applications for the November sitting are open.

The second half maps the 40 days between publication and 7 November onto six ways of studying with AI. Each of the six has its own article with procedures and measurements; this one is the map.

What the G-test is: the fork between generalist and engineer

The G-test is formally the Generalist Test, run by the JDLA. It is paired with the E-qualification, the engineer credential from the same association, and the difference is exactly what the names suggest: the side that uses AI against the side that builds it.

The JDLA introductory deck positions the G-test as a test for acquiring literacy in applying AI and deep learning, and states plainly that it is recommended for people on the business side. So what the G-test certifies is not the ability to implement a model. It is a systematic grasp of the technical methods and the business applications. There is no exclusive practice attached to it, the way there is for physicians or lawyers. It proves knowledge; no job is closed to you without it.

The E-qualification, by contrast, gates entry. You must have completed a JDLA-accredited programme within the two years before the exam date. The exam runs 120 minutes with roughly 100 questions, and the general fee is 33,000 yen including tax. In 2026 it is held twice, in February and August. The G-test has no eligibility restriction at all and runs six times a year.

Settling this fork early makes it easier to decide how deep to study. What the G-test asks is that you can match a method to its name and role, and that you understand the arguments in law and ethics. It does not ask you to derive backpropagation.

Concretely: one official sample question asks which recommendation method can be used from day one on an e-commerce site launching next month. You only need to see that collaborative filtering cannot work before there is any purchase history. The faculty under test is the ability to tie a technique to the conditions under which it does and does not work. That same faculty is used when a salesperson pitches AI to a customer, and when a planner signs a contract with a development firm.

The numbers: candidates and pass rates

The JDLA publishes candidates, passes and the pass rate for every sitting. Here are the most recent five.

Cumulatively, as of the fourth sitting of 2026, 219,761 people have taken it and 156,562 have passed. Results for the fifth sitting, held on 5 September, had not been published as of 21 September.

The trajectory is worth noting too. The cumulative candidate count was about 47,000 at the end of 2020, about 89,000 at the end of 2022 and about 140,000 at the end of 2024, passing 210,000 in 2026. The annual increments were roughly 23,000 in 2023, 28,000 in 2024 and 42,000 in 2025. Growth accelerated after generative AI became a public conversation in 2023 rather than tapering off, which suggests the credential is settling in as a shared vocabulary for the people who use AI rather than riding a fad.

Pass rates sit in a band of 77 to 83 percent. It would be hasty to read that as an easy exam. The JDLA does not publish the passing score, so nobody outside knows what fraction of correct answers is enough. And 85 percent of those who pass are working adults in their twenties to forties. In the JDLA survey, the leading reasons for sitting were that it seemed a skill that would be needed, 36.7 percent, interest in the field, 19.6 percent, and a recommendation from an employer or school, 15.8 percent. A high pass rate shows that most people who prepare get through. It does not show that people get through without preparing.

Who sits it: reading the credential from the demographics

The breakdown for the fourth sitting makes it clear who this credential is for. What the JDLA publishes per sitting is the profile of the people who passed, so the figures below are shares of the 7,677 who passed.

The second entry under job function is the striking one. Sales accounts for 15 percent, more than research and development. By industry, finance, insurance and real estate at 18 percent outranks software. This is not an engineering credential. It is one taken by people whose job is to explain AI to customers and colleagues.

That is also an evidence-based answer to the question of what is worth certifying in an age of AI. Building AI is certified by the E-qualification and by shipped work. What the G-test certifies is the ability to use the vocabulary correctly, and to judge what AI can and cannot do and where the legal lines fall, in the situations where AI is proposed, procured, planned and managed. That people in their twenties make up 38 percent suggests the value of holding that judgement early in a career is already recognised.

The shape of the exam: 145 questions in 100 minutes

The format changed with the first sitting of 2026. The online exam runs 100 minutes with roughly 145 short knowledge questions in multiple-choice form, taken at home. The venue exam runs 120 minutes over the same roughly 145 questions, but the November sitting has no venue option.

Dividing 6,000 seconds by 145 leaves 41 seconds per question. That number decides how to study. This is not an exam you reason your way through; it is one where you pull the mapping between a term and a concept out of memory instantly. Look at the official samples and you will see most stems take the form of choose the single most appropriate option or choose the single least appropriate option, or fill the blank, always from four choices.

The 41-second constraint feeds back into how you build your materials. Rather than reading long explanations, you want to manufacture, in volume, the state where you can say in one line what category a term belongs to and what it is easily confused with. Mathematics and statistics is one of the 37 mid-level items on the technical side, and the official samples include one question mapping definitions across correlation, partial correlation, covariance and variance, plus one simple expected-value calculation with dice. Nobody is hammering a calculator. So the ratio of memorisation to understanding leans towards understanding-flavoured memorisation, and most of the work is building a network of terms in your head.

The fee is 13,200 yen for the general public and 5,500 yen for students, both including tax. Eligibility is listed as none: no educational or professional requirement. If you sat the exam within the previous two years you can buy a ticket at half price, 6,600 and 2,750 yen, so a second attempt is cheap. There is also a 30 percent discount for people who completed the Coursera course AI For Everyone, at 9,240 and 3,850 yen.

The 2026 calendar and the November deadline

In 2026 there are six online sittings and three venue sittings.

Applications for the sixth sitting run from 13:00 on 11 September to 23:59 on 29 October for individuals, and to 23:59 on 22 October for groups. You choose between the Friday 16:00 start and the Saturday 13:00 start. The 2027 calendar was announced on 11 September: the first sitting is online on Saturday 16 January 2027 at 13:00, with venue sittings from 15 to 17 January, and applications open from 13 November 2026 to 8 January 2027. Missing November costs you two months, not a year. Counted from the publication date of this article, 28 September, there is one month to the application deadline and 40 days to the exam.

From application to result

The process has three steps, described on the official page as buying a ticket, booking the sitting you want, and taking the exam on the day through the exam page. A ticket is valid for one year, so you can buy first and choose the sitting later.

The recommended environment for the online exam is published on the exam portal with version numbers. What the portal showed as of 17 July 2026 at 10:00 was Windows 11 25H2 or macOS 26.6.1, specified versions of Chrome, Firefox, Safari and Edge, a screen resolution of 1024 by 768 or better, a connection of 1 Mbps or better, JavaScript and cookies enabled, and pop-up blocking turned off. Taking the exam on a tablet or smartphone is explicitly not recommended, on the grounds that it causes malfunctions. Version numbers move, so confirm on the portal when you apply.

Results arrive by email within two to three weeks of the exam. The certificate is downloaded from your account page; for the fifth sitting the announced window opened at 10:00 on 13 October. Passing also issues an open badge conforming to the 1EdTech Consortium standard, shareable over social media and email, and gives access to CDLE, the community of people who have passed.

The map of Syllabus 2024: ten major items, fifty-five mid-level items

The scope is defined by Syllabus 2024, applied from the November 2024 sitting onwards. The current revision is 1.4, dated 11 May 2026. The technical domain holds 8 major items across 37 mid-level items, and the law and ethics domain holds 2 major items across 18 mid-level items.

Each mid-level item lists an objective and a set of keywords. The keyword list is the backbone of any glossary you build, and it is also the axis along which you should file your mistakes.

Do the twenty official sample questions first

The only past questions the JDLA publishes are twenty samples, drawn from sittings between the first of 2022 and the sixth of 2024, labelled with the sitting and question number. Four choices each, answers included.

Do these twenty before you start studying. The point is not to score well; it is to get the shape of the stems and the construction of the distractors into your body. The statistics questions, for instance, arrive as a matching exercise: descriptions such as an index expressing the degree of correlation and an index expressing the degree of correlation with the influence of other variables removed, mapped onto correlation, partial correlation, covariance and variance. Knowing that shape changes how you build your glossary.

Study hours and materials: 30 to 50 hours, built around the official text

According to the survey of people who passed, published by the JDLA, the most common amount of study time was 30 to 50 hours at 25.3 percent, followed by 15 to 30 hours at 21.5 percent and 50 to 70 hours at 18.4 percent. The same material notes that most people who passed worked from two or three sources centred on the official textbook.

The official textbook is the third edition of the Deep Learning G-Test Official Text from Shoeisha, published on 27 May 2024, 3,080 yen including tax, A5 format, 448 pages, supervised by the JDLA and written by Takayoshi Yamashita and others. It conforms fully to Syllabus 2024, adds foundation models and language models, and runs to eight chapters plus a case-study appendix, with chapter 8 given over to the law and ethics of AI. Choosing problem sets and courses is covered source by source in a separate article, but if you are restricting yourself to a single book, this is currently the one.

Securing 30 to 50 hours across 40 days means 45 to 75 minutes a day. Forty-five minutes on weekdays plus two hours at the weekend gets you there.

Three traps when studying with AI

• The AI does not know the scope. Ask for important terms in a major item without handing over the syllabus and it will cheerfully produce terms that are not in the scope at all. When this site measured it, a glossary built without the syllabus covered 169 of the 495 keywords and added 438 terms that were not on the list.
• The AI is fluent when it is wrong. Where correctness depends on proper nouns, on which ministry owns a guideline and on which revision is current, generated questions go wrong. Errors clustered in law and guidelines, not in technique.
• Reading an AI summary feels like understanding. It produces recognition, which the exam does not reward. The exam rewards being able to tell two neighbouring terms apart.

The 40-day design: six methods and where each one goes

The order matters. Ingest comes first because whether you hand over the syllabus changes the quality of everything downstream. In the measurement above, the glossary built with the syllabus covered 490 of 495 keywords; the one built without it covered 169 and invented 438 out-of-scope terms. The AI builds plausible-looking material without knowing the boundary. Ingest fixes the boundary first.

Generate sits in the middle for a related reason. Generated questions contain errors. In the measurement, one of 50 questions produced with the syllabus needed a correction, while 12 of 30 produced without it had a problem, two of them where the answer or a factual claim in the stem contradicted a primary source.

On the day: how to spend 41 seconds

One hundred minutes for 145 questions means that dwelling on one question costs you the tail of the paper. Three rules are enough.

1. Skip what you do not know. If you do not hold the mapping, thinking will not produce it. Flag it, move on, come back at the end.
2. Read the options first. The shape of the right answer is often visible among the choices, and you can narrow to two before reading the stem closely.
3. Do not reserve time for review. One pass through 145 questions fills almost all of the 100 minutes, so the time you spend returning to flagged questions is your review.

On the environment side, confirm the version of your browser the day before, and silence notifications from other applications on the day. Because the online exam is taken at home, reading the recovery procedure for a dropped connection in advance keeps you calm if it happens.

Summary: what to do today

The G-test certifies, systematically, the knowledge of the side that uses AI rather than the side that builds it, and the demographics bear that out: 15 percent of those who pass work in sales and 18 percent in finance, insurance and real estate. Pass rates over the last five sittings ran 77 to 83 percent, and the most common study time in the JDLA survey of successful candidates was 30 to 50 hours. The November sitting is online only, applications close at 23:59 on 29 October, and the exam is on 6 or 7 November. One hundred minutes for roughly 145 questions, 41 seconds each, is the premise that decides how to study.

Three things to do today: buy a ticket and book the November sitting, work the twenty official samples to learn the shape of the questions, and put the PDF of Syllabus 2024 revision 1.4 within reach. That third one is the raw material for the ingest step you start tomorrow.

Sources

• JDLA, About the G-test (format, fees, 2026 calendar, application periods, recommended environment)
• JDLA, About the E-qualification
• JDLA, 2027 annual G-test schedule announcement
• JDLA, results of the fourth 2026 sitting (candidates, pass rate, cumulative totals, demographics)
• JDLA, results of the second 2026 sitting
• JDLA, results of the first 2026 sitting
• JDLA, results of the sixth 2025 sitting
• JDLA, How to study for the G-test (study-hours and motivation survey)
• JDLA, official sample questions
• JDLA, G-test FAQ
• G-test exam site (certificate download dates)
• Shoeisha, Deep Learning G-Test Official Text, third edition

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

swiftwand

X

AIを使って、毎日の生活をもっと快適にするアイデアや将来像を発信しています。 初心者にもわかりやすく、すぐに取り入れられる実践的な情報をお届けします。 Sharing ideas and visions for a better daily life with AI. Practical tips that anyone can start using right away.