Employment Security Act
Screened before it is spoken
Every question the AI generates is screened by machine, before speech synthesis, against the 14 topics the Ministry of Health, Labour and Welfare lists as liable to lead to discrimination in hiring. A match is discarded before the candidate hears it, and the discard is recorded. This is not a setting; it works the same way on every contract.
The list
The 14 topics the screen covers
The Ministry's grouping, in our translation. The Japanese page carries the Ministry's own wording.
Group
Topic
01 — 14
Matters the person is not responsible for
Things no intention or effort can change.
4 topicsScreened: In the interview and at authoring
- Registered domicile or place of birth
- Family (occupation, relationship, health, medical history, status, education, income, assets)
- Housing (layout, number of rooms, type of dwelling, nearby facilities)
- Living or home environment
Matters that ought to be free
Thought and belief.
7 topicsScreened: In the interview and at authoring
- Religion
- Political party supported
- Outlook on life or personal creed
- People admired
- Ideology
- Involvement in labour unions, student movements or other social movements
- Newspapers, magazines or books read
Selection methods
Not the content of a question but the way selection is run. These are never generated as spoken questions, so they are screened on the criteria-editing screen.
3 topicsScreened: At authoring
- Background investigations
- Application forms that ask for matters unrelated to aptitude or ability
- Medical examinations with no reasonable, objective necessity
Where it runs
In the interview, and when criteria are authored
The same topics, checked in two places. In the interview, immediately before a question is spoken. At authoring, every time a criterion is saved.
In the interview
Discarded before it is spoken
Where an answer is thin, the interviewer follows up on the spot. The whole follow-up is screened immediately before speech synthesis. A match never reaches the candidate; the interview simply moves to the next question. If no verdict comes back, the question is discarded unspoken all the same.
When criteria are authored
Not publishable until each warning has a recorded reason
Every save of the criteria runs the same check and warns on any criterion that may touch one of the topics. A warning is cleared by recording why that criterion concerns aptitude or ability. A job with an unanswered warning cannot be published, and the reason and its author go to the audit log.
The record
What was not asked is recorded too
A discarded question is recorded as a discard. Being able to show afterwards that something was not asked is what makes the difference when a Labour Bureau or your own legal team asks.
- Questions the candidate never heard are recorded
- A question discarded before it was spoken did not reach anyone. The fact that the model generated it remains, so it is recorded. Logging only what was heard would make the screen look cleaner than it is.
- One row says what it matched
- The discarded text, the topic it matched, the reason for the verdict, and the model version that gave it. A discard with no verdict is recorded distinctly, so "many questions matched" and "the screen was down" can be told apart later.
- Authoring-time answers land in the same log
- A reason recorded against a warning in the criteria editor goes to the same audit log with its author, and can be read back reason by reason.
Design decisions
Fixed in the design, not in a setting
Neither can be changed on a settings screen. Two points your legal team can take as they stand.
Basis: Act on the Protection of Personal Information
No facial images are handled
A facial image is a personal identifier code under the APPI, which raises the bar for handling it — and an evaluation report does not need it. So there is no facial or emotion analysis, video is off by default, and only audio is handled.
What the candidate sees is our AI interviewer
The candidate's camera is never used.
- Camera use
- Storing facial images
- Facial or emotion analysis
Basis: design — an override requires a reason
A person makes the final decision
Hitoshiru does not decide outcomes. Its role ends at a score per criterion and the utterance behind it. A recruiter can overwrite any score with a reason; the model's original number stays, and the override, the reason and its author are written to the audit log.
Reason: incident handling at the previous employer confirmed in a follow-up conversation.
Sample. Not a real candidate.
Scope
What we can guarantee, and what we cannot
The mechanical screen covers the questions Hitoshiru generates and speaks to the candidate, and the content saved in the criteria editor. Interviews your own staff conduct, and the design of your application forms, are outside it.
Screened
- Follow-up questions the AI generates and speaks to the candidate during an interview
- Evaluation criteria saved in the criteria editor
Not screened
- Questions your own staff ask in interviews they conduct themselves
- The design of your application forms and what they collect
- Background checks or medical examinations run outside the interview
What Hitoshiru guarantees is that none of the 14 topics gets into a question we generate — not the lawfulness of your hiring process as a whole. We say so early in every conversation.
Written for your legal review
This page is written at the level of detail a legal or IT team can check as it stands. Security questionnaires and questions about the screen itself are answered individually.