What happened
It was reported by JMITU’s IBM Japan branch that IBM Japan told group employees in August 2019 that it had introduced its own AI, Watson, into pay assessments. The union says it asked IBM Japan to explain what the AI considered and what it showed managers.
It was reported by the union that IBM Japan refused those requests. On 3 April 2020, the union says it applied to the Tokyo Metropolitan Labour Relations Commission, alleging unfair labour practices under Article 7 of the Labour Union Act: bargaining in bad faith and interference.
According to the union, IBM Japan said the AI was only a tool to support managers’ judgement. The union reports that IBM Japan also said the information shown to managers was not intended to be disclosed to employees as it stood.
What an auditable version would have shown
It was reported by the union that company material described about 40 kinds of data considered by the AI and pay proposals shown to managers, including by what percentage to raise pay. The union also says IBM’s sales material said managers tend to follow its recommendations.
The union reports that IBM Japan’s position was that managers made the judgement. The sources reviewed do not show how managers actually used the AI’s proposals or whether anyone’s pay was set differently because of them.
Where the gap was
It was reported by the union that the parties reached a settlement before the commission on 1 August 2024. Under the settlement published by the union, IBM Japan agreed to disclose to the unions the titles of all items the AI was made to consider in the 2019 and future pay adjustments. The unions agreed not to disclose those titles to third parties indiscriminately.
The settlement records that when a member’s pay review is disputed for specific reasons, such as a cut, zero raise or low rating, IBM Japan will present and explain the AI’s proposal. It will also explain how the items considered by the AI relate to its grading rules and discuss future questions about the AI’s pay evaluation method in good faith.
The settlement contains no finding that IBM Japan acted unlawfully. It was reported by the union that Kyndryl Japan, which took over part of IBM Japan’s business after the complaint, does not currently use AI in pay assessments.
A ConductRecord keeps, for each pay review, what the AI considered, what it proposed and what the manager decided.
A MetricRecord is designed to produce signed figures, such as how often managers followed the AI’s proposal, that a union or regulator can verify.
ConductRecord and MetricRecord are Headlights designs.
What governance should have looked like
It was reported by the union that the settlement gave it a way to see the AI’s proposal when a member’s pay review was disputed, about five years after IBM Japan announced it had introduced the AI. No statement by IBM Japan on the settlement was found as of 3 October 2026.
Where an AI proposes pay, best practice would be to tell employees what it considers, keep its proposal and the manager’s decision for each review, and show both to an employee who disputes the outcome.
Failure Pattern: according to the union, an employer used its own AI to propose pay changes to managers and refused to disclose what the AI considered until a settlement required it.
Governance Principle: when an AI proposes pay for a person, the employer should be able to show what it considered and what it proposed for that person.
The reference implementation of ConductRecord and MetricRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- ＡＩ不当労働行為事件勝利和解, statement on the settlement (JMITU IBM Japan branch, 1 August 2024, posted 17 August 2024)
- ＡＩ不当労働行為事件都労委で勝利和解, continued, with the settlement terms (JMITU IBM Japan branch, 2 September 2024)
- 日本ＩＢＭ・ＡＩ不当労働行為申立 (JMITU IBM Japan branch, 19 April 2020)
- ＡＩ不当労働行為事件 (JMITU IBM Japan branch, 17 April 2022)
- Bengo4.com news, 2 August 2024