We asked ChatGPT to explain when an employer has to report under the Pay Transparency Directive. The answer sounded convincing. It was wrong on four points. Pay transparency is not a language task. It is a legal-statistical domain with 36 articles in Richtlijn (EU) 2023/970 and a Dutch implementation that is still moving. The calculation side has nothing to do with language.
Last update: May 18, 2026 · Reading time: 6 minutes
A generic chatbot such as ChatGPT, Gemini or Copilot works with patterns in text. That collides hard with a domain where one wrong article number or one wrong threshold pulls the floor out from under your compliance.
The Dutch implementation of Directive (EU) 2023/970 is still moving at the time of writing. The explanatory memorandum to the draft bill clarifies definitions and scope, but is not consulted as an authoritative source by any public chatbot. What you get back is, by definition, a guess based on whatever the chatbot has seen somewhere in its training data.
Asked "when do I have to report?", we see the same four errors repeatedly.
"Article 11 WGBMV" does not exist, but is cited with confidence. The same goes for references to directive articles that have nothing to do with reporting. The chatbot picks a plausible-sounding number and presents it as a source.
The directive must be transposed by 7 June 2026. The Netherlands will not make that date. The target date in the draft bill of 26 March 2025 is 1 January 2027. Almost no chatbot makes that distinction; most stay stuck on the EU deadline.
The law sets a three-yearly reporting obligation for employers with 100–249 employees and an annual obligation from 250 upwards. Chatbots regularly throw in 150, or drop the three-yearly category entirely.
Article 4 of the directive defines work of equal value on four factors: skills, effort, responsibility and working conditions. Chatbots collapse this into "equal work" and lose the entire job-evaluation side.
One wrong article number in a memo to your works council is enough to put the whole policy document in doubt.
Invented case law is not a theoretical risk. Judges call it out explicitly.
A lawyer lost credibility because the ECLI numbers in his court filings did not exist. The court spotted the hallucination pattern within minutes.
The Raad van State ruled on fraud in which ChatGPT output had been used as a source. The ruling sets a line on the value of unverified AI output in legal documents.
An employer who bases his report or his Joint Pay Assessment on a chatbot hallucination ends up in the same category. Where a pay gap of 5% or more cannot be objectively explained, Article 10 of the directive requires a Joint Pay Assessment. Article 16 reverses the burden of proof: the employer must show that there is no discrimination. An invented article number or a wrong threshold will not survive in court.
Beyond the factual errors there is a structural gap. These are the five things a language model cannot deliver, however well crafted the prompt.
The explanatory memorandum to the Dutch bill clarifies definitions and scope. A chatbot does not hold the text as an authoritative source and will not consult it.
Belgium implements the directive differently from the Netherlands. Chatbots blend the regimes and present a Belgian threshold or definition as Dutch law.
The difference between the gross pay gap and the pay gap after correction for role, experience and hours is the core of the analysis. Chatbots use the terms interchangeably.
Clustering employees into groups of equal-value work and significance testing on pay differences are not language tasks. A language model cannot perform them. It can at most talk about them.
Pasting salary data into a public chatbot is a data breach. The data leaves your processor environment and lands in training or logging infrastructure you have no grip on.
Payqual is built on the directive text and the explanatory memorandum, not on patterns in arbitrary internet text. The logic is domain-built: thresholds, deadlines, definitions and reporting frequencies sit as rules in the system, not as a probability distribution in a model.
Thresholds, deadlines, definitions and reporting frequencies are encoded as rules. No probability distribution, no training on arbitrary internet text.
The analysis runs without personal data. What you analyse does not leave your processor environment and does not end up in logging or training infrastructure.
Job evaluation, clustering and significance testing are real statistical operations. Not language predictions that happen to look like numbers.
The same input produces the same outcome, with traceable references to the article on which the conclusion rests. Ready to hand to the Labour Inspectorate without ad-hoc substantiation.
A chatbot is a fine sparring partner for a draft email. For the WGBMV it is a liability risk. Pay transparency deserves tools built for the problem.
For tone and structure, that is reasonably safe. For substantive obligations, the risk is too high: which salary information is required, which pay level is expected and which WGBMV article underlies it. Verify every legal claim against the directive text or a specialised source before publishing.
It is a data breach. Salary and role data are personal data under the GDPR. Public chatbots use input for further training, storage and logging. You have no processing agreement, no data-retention policy and no audit trail. For a pay transparency analysis, use tooling that keeps the data inside your processor environment.
Open Directive (EU) 2023/970 or the Wgbmv and look up the article. Does the article exist? Does it cover the topic the chatbot claims? Two minutes of verification saves you from a memo, policy document or inspection response that leans on an invented source.
Employer organisations are sounding the alarm at BNR: Dutch SMEs are going to miss the pay transparency deadline. The bottleneck isn't the law — it's the absence of tooling.
Back pay, fines, exclusion from procurement and talent loss. The full risk landscape of the EU Pay Transparency Directive.
Under the new pay transparency directive, the burden of proof is reversed. If an employee suspects discrimination, the employer must now prove it did not occur.
Payqual does not interpret the WGBMV; it applies it. Job evaluation, clustering, 5% checks and reporting come from one system with traceable source references, so that your answer to an information request or an inspection does not lean on a chatbot hallucination.