Executive Summary

Greece has set out to govern AI at work and platform work in a single bill. The Ministry of Labour and Social Security opened the bill for public comment on November 5, and the window closes at 11 p.m. on November 19. The clause that stands out is the one on decisions that cut off a person's ability to work. Suspending or terminating an account could not be done by an automated process alone, and a trained person would have to look at the decision again.

The schedule is tight because of a clock in Brussels. The deadline for writing the EU platform work directive (2024/2831) into national law is December 2, two weeks after comments close. Greece's current statute, Law 4808/2021, starts from the assumption that someone working through a platform is self-employed, so this bill also turns that starting point around. The government consultation page carrying the text of the bill cannot be opened from outside Greece, which means article numbers can only be confirmed for the EU directive and the current Greek law.

The facts and figures in sections 1 through 3 come from the labor ministry's announcement, Greek and German press coverage, a survey of transposition across Europe, the text of the EU directive, and commentary on the current Greek law; the account of Korean law in section 5 comes from the AI Framework Act and commentary on its enforcement decree. The second half of section 4, which strings the four records into a single chain, and all of section 6 are this article's interpretation.

Key Numbers

The first two numbers mark the time Greece has left. The last two are the interval the law asks platforms to keep and the line it tells them not to cross.

Sources: Greek Ministry of Labour and Social Security, Directive (EU) 2024/2831, borncity.

November 19

The day the comment window closes

11 p.m. local time. It opened on November 5

December 2

Deadline to write the EU directive into national law

Two weeks after comments close

Every 2 years

How often the algorithm impact assessment has to be redone

Workers' representatives take part (directive, Article 7)

Five

Uses of AI the bill rules out, as reported

Union profiling, emotion data, private conversations, biometric ID, off-hours monitoring

1

Two Weeks of Comments, and a Date 27 Countries Share

The Greek Ministry of Labour and Social Security presented the bill to the cabinet on November 2 and posted the full text three days later, on November 5, to the government consultation site opengov.gr. Its formal name is long: "Rules on the use of artificial intelligence in the labour market — protection of platform workers — transposition of Directive (EU) 2024/2831 of the European Parliament and of the Council — pension provisions and other provisions." As the name already says, one bill carries two bundles of a very different nature. One is a set of rules on how AI may be used at work. The other governs the status of people who take their work through a platform.

Greek outlets carried the same account of what Labour Minister Niki Kerameus said. Artificial intelligence is a highly useful tool, but its use has to be held in check by transparency and safeguards. That is why the bill builds human oversight into the algorithmic management that platforms run, and adds provisions that protect workers in practice. The remarks were made in Greek and reported by Greek outlets, so this article does not put quotation marks around any particular wording.

This is the first time Greece has written rules for algorithmic management of work. And those rules do not stop at platforms. Reporting says the bill also takes aim at systems used in hiring, evaluation and productivity monitoring. The EU directive sets its floor at platform work, so Greece has drawn the line wider, across employment in general.

Greece is not the only country under time pressure. December 2 falls on all 27 member states alike. A survey of transposition through the middle of 2026 reported that countries such as France, Germany, the Netherlands and Belgium had said they would meet the date, yet no published bill or consultation could be confirmed in them, and that Italy had gone as far as a draft. November 5, the day Greece opened its consultation, left less than four weeks before that date, and in the same survey few countries had reached the stage of publishing a document and taking comments on it.

A limit belongs here at the start. The opengov.gr page carrying the text of the bill is closed to outside access, so it could not be opened directly. The article numbers cited in this article therefore belong to Directive (EU) 2024/2831 and to the current Greek labor statute, Law 4808/2021; no article number from the Greek bill itself appears here. What the bill contains is reported only as far as the ministry's announcement and several press accounts could be laid over each other.

2

Greece Turns Its Self-Employment Presumption Around

Arguments about platform work tend to converge on one question. Is this person a worker employed by the platform, or a self-employed contractor selling services to it? The answer carries the minimum wage, working time, social insurance and protection against dismissal with it, all of them at once or none of them at all. So national laws differ less in how they weigh the substance of the relationship than in where they place the starting point. When a dispute arises, which side is the default, and which side has to rebut it?

The default Greece set in 2021 was self-employment. Law 4808/2021 defines digital platforms in Article 68, and in Article 69 it splits the relationship between a platform and a service provider into an independent services contract and a dependent employment contract. It lays self-employment underneath as the presumption, and to overturn it the law asks whether the person chooses their own hours and volume of work, whether the work can be passed to someone else, whether no non-compete clause binds them. Even within Europe that was an unusual design. Spain went the other way in that same year, 2021, and was first to legislate that a platform controlling the work is presumed to be an employer, in what became known as the riders law. The EU asked every member state to go in the same direction in 2024. Article 5 of the directive sets out that a contractual relationship with platform control is presumed to be an employment relationship, and it is the first piece of EU legislation to presume a labor relationship this way.

The bill moves that default toward Spain and the EU. If a platform directs and controls the work, the relationship is presumed to be employment, and the platform has to prove that it is not. The presumption can be broken. What changes is that the burden of proof sits with the platform rather than with the person doing the work. To overturn the presumption, a platform now has to open up how it runs.

Status is not the only thing being reworked. According to press reports, the bill would make platforms register with ERGANI, the labor ministry's information system, and hold them jointly and severally liable for obligations arising along the subcontracting chain. Provisions that make parental and carer's leave easier to take and widen social security coverage for children went in alongside. The registration duty comes back later in this article. Oversight is only a word unless there is a roster of who is running what.

3

Who Oversees, and Which Decisions a Person Has to Make

Human oversight left as a declaration changes nothing. The EU directive writes it into two articles, and these are the two the Greek bill takes over. Article 7 sets conditions on the side doing the overseeing. Article 8 sets them on the decision itself.

3.1Conditions on the person doing the oversight

Article 7 requires platforms to monitor and evaluate, with sufficient human resources and on an ongoing basis, the individual decisions taken by automated monitoring and decision-making systems. Whoever does the overseeing must have the authority to override an automated decision along with the training and competence that go with it, and must be protected from dismissal or discipline for having done so. Where the oversight turns up a problem such as discrimination, it has to be corrected within two weeks. And at least once every two years the platform has to assess the effect of the system on working conditions, with workers' representatives in the room.

The Greek bill adds two things on top. The oversight has to be carried out by staff who have been specifically trained, and the platform has to consult workers' representatives, such as a union, before bringing a system in. A data protection impact assessment is also required. Once before adoption, continuously while the system runs, and again every two years.

3.2Conditions on major decisions

Article 8 splits decisions into two branches. Decisions that carry real weight — restricting, suspending or terminating an account, refusing payment, determining contractual status — cannot be taken by an algorithm on its own and have to be taken by a person. Written reasons are due by the day the decision takes effect, and the person affected may discuss the facts with a competent contact and ask for an explanation. Article 10 extends this transparency, oversight and right of review to platform workers who are self-employed rather than in an employment relationship. Articles 17 and 18 forbid adverse treatment for having used these rights and put the duty to explain the grounds for dismissal on the platform.

The diagram below sets out how decisions from the same automated system part ways according to their weight.

Decision from an automated system Routine decisions task order, how listings surface Major decisions suspension, refusal to pay, status May be made automatically. In exchange they fall under continuous oversight and a review at least every two years. A person has to make them. Written reasons are due by the day it takes effect, plus a right to an explanation and to a review.

▲ Original Pebblous diagram. The branches and the conditions are taken from Articles 7 and 8 of Directive (EU) 2024/2831.

There is also a side the bill shuts off entirely. The list of prohibitions reported by a German-language outlet runs to five: predicting union activity or the inclination to join one, collecting data on emotions and psychological states, analyzing private conversations, identifying people from biometric data, and monitoring activity outside working hours. The same report says this line applies not only to platforms but to hiring, evaluation and productivity monitoring systems. Prohibition is the device that sits opposite oversight. Some decisions have to be looked at again by a person; some inputs may not be collected in the first place.

4

Oversight Stops Where the Records Stop

Picture someone whose account has been suspended asking for an explanation, and it becomes clear why this bill is a story about records. To answer, the person handling the case has to be able to pull up what data the decision looked at and what standard it was measured against. If the model changed a few times in between, they need to know which version was running then. With nothing to pull up, a review becomes a second attempt at the same decision, and that is not oversight.

Lay out what the bill and the directive require and there are four devices, each leaving its record at a different moment.

Device When the record is left What it makes possible
Platform registry (ERGANI) Before operations begin A roster of who is running what
Data protection impact assessment Before the system is adopted What data is used, and for what purpose, written down in advance
Written reasons By the day the decision takes effect Why this one case was decided the way it was
Periodic impact assessment At least every two years Skew and discrimination visible once decisions have piled up

▲ The registry and the impact assessment are what reporting says the bill contains; the written reasons and the two-year cycle are written into Articles 7 and 8 of the directive.

Stand the four in order of time and they become a single chain. A roster goes up before operations start, the data to be used is written down before adoption, the grounds are left at the moment of decision, and the accumulated decisions are looked at again on a cycle. The right to ask for an explanation and to request a review is the door through which the person doing the work enters that chain. If a link in the middle is missing, opening the door shows them nothing.

From here on this is the article's interpretation. Put what the text requires into the vocabulary of data and it comes to three things. Preserve the inputs a decision used in the state they were in at that moment. Bind each decision to the standard in force at the time, the rule and the model version. And keep a separate record of the fact that a person overturned a decision. The last of the three is the one most often missing, even though that record is the only way to check afterward whether human oversight actually worked.

5

Korea's AI Framework Act Names Hiring and Leaves the Rest Open

Korea has a law covering the same ground: the Framework Act on the Development of Artificial Intelligence and the Establishment of a Foundation of Trust, the AI Framework Act for short. It took effect as Act No. 21311 on January 22, 2026, some provisions took effect on July 21, and the enforcement decree has been running since August 20. Article 2(4)(g) classifies AI used for judgments or evaluations that have a significant effect on an individual's rights and obligations, such as hiring, as high-impact AI. Once a system falls into that class, the operator has to put together a risk management plan and an explanation plan, maintain a structure for human management and oversight, and give advance notice of decisions that significantly affect workers.

The design points the same way as the Greek bill. Where they part is scope. The stage the Korean text names clearly is hiring. That is a narrower span than Annex III of the EU AI Act, which treats the whole arc of employment as high risk, from hiring through task allocation and evaluation to termination of the contract. Whether automated decisions used after someone joins — performance reviews, promotion, reassignment, dismissal — fall under the high-impact class is read differently by different people. The part of the arc the Greek bill aims at head-on is exactly that later stretch. Suspending an account and terminating a contract happen at the end of the relationship, not at hiring.

For the places where readings diverge, the enforcement decree set up a way to ask. An operator may request confirmation of whether its own system counts as high-impact AI, and the government replies within 30 days, with one further 30-day extension if needed. Within 10 days of receiving the reply, the operator may ask for a second look. A company that has hooked automated decisions into the stages after hiring, and is not confident in its own reading, has that path available.

Movement in the National Assembly has been toward filling the gap. Bills amending the Fair Hiring Procedure Act, filed by lawmakers including Kim Wi-sang and Lim Lee-ja, would make advance notice of AI use in hiring and bias checks mandatory. Some go further and would bar dismissal, suspension, disciplinary leave, transfer or pay cuts decided by AI alone, and would have a person who was not involved in the decision look at it again when the affected party asks. That is in effect the same design the Greek bill writes into its text. The difference is which stage each has reached. The Greek bill sits in consultation with a deadline attached; the Korean bills have only been introduced.

6

Why Pebblous Is Watching This Bill

News like this is usually read as labor news. Seen from the data side, something else is left over. The moment a law requires human oversight, that requirement turns into a logging requirement inside the organization. What will be shown matters before who will do the overseeing. Seating someone in that chair is an org chart problem; putting the grounds for the decision in front of them is a systems problem.

Records of automated decisions behave unlike other data quality problems. A wrong value is caught by validation, but a value that was never written down is not caught even as an error. If the inputs behind a decision were overwritten, or the model version was never recorded, the fact surfaces on the day an explanation is demanded. That day usually comes well past the point where it could have been fixed.

So for an organization running automated decisions, what can be checked right now narrows to three things.

  • Are the inputs a decision used still held as they were at that moment? If later values overwrote them, the same decision cannot be reproduced.
  • Is each decision bound to the standard that applied to it? Rules and models change, and a decision log on its own does not restore why the call went the way it did.
  • Is human intervention recorded separately? Without who overturned what and when, the fact that oversight happened cannot itself be shown.

When Pebblous puts data into a state that AI can use, one of the quality items is whether there is a field to write these things in, for the same reason. Greece's consultation closes on November 19 and the EU deadline falls on December 2, but the question outlives that calendar. What is our own automated decision-making leaving behind right now?

R

References

Official documents & legislation

  • 1.Greek Ministry of Labour and Social Security. (2026). Public consultation notice for the AI-at-work and platform-workers bill. ypergasias.gov.gr
  • 2.Greek government public consultation platform. (2026). Bill text consultation page (HTTP 403 at time of access). opengov.gr
  • 3.European Parliament and Council of the European Union. (2024). Directive (EU) 2024/2831 on improving working conditions in platform work. EUR-Lex
  • 4.Eurofound. (2021). Greek Law no. 4808/2021. Eurofound Platform Economy Database

Greek & EU coverage

  • 5.RemoteWorkEurope. (2026). "Greece puts its AI-at-work and platform-workers bill out for consultation until 19 November." remoteworkeurope.eu
  • 6.RemoteWorkEurope. (2026). "Platform Work Directive enforcement divergence across EU member states." remoteworkeurope.eu
  • 7.Ta Nea. (2026). Coverage of the bill's public consultation. tanea.gr
  • 8.The Opinion. (2026). Coverage of the algorithmic-management and platform provisions. theopinion.gr
  • 9.borncity.com. (2026). "Greece's draft bans predicting union sympathies" (German-language coverage). borncity.com

Korea AI regulation sources

  • 10.Kim & Chang. (2026). Draft enforcement decree of Korea's AI Basic Act. kimchang.com
  • 11.datalaw.kr. (2026). The 10 domains of high-impact AI. datalaw.kr
  • 12.Digital Daily. (2026). "Are AI-based hiring and evaluation rules enforceable?" ddaily.co.kr