Executive Summary

The Federation of Korean Industries (FKI), the business group that speaks for the country's large companies, said on October 29 that it had gathered its members' views and sent the Ministry of Science and ICT 19 proposed changes to the AI Basic Act. The 19 fall into five areas, and the item that opens them is AI built for in-house work. A company that builds its own AI and lets only its employees or its affiliates use it has no way to tell whether it carries the same operator duties as a commercial service sold to outside customers.

FKI is not the only one asking. The AI Basic Act support desk, opened by the ministry when the law took effect, took 552 inquiries over the ten weeks from January 22 to March 30, and among the online inquiries alone, 19.6 percent asked whether the company's own AI counts as high-impact AI. At the end of March the ministry went on to publish a collection of worked cases. Seven months later, a business group has brought the same question back.

Sections 1 through 4 follow the submission FKI released in full and the text of the AI Basic Act and its Enforcement Decree. Section 5, which carries this gap over to the question of records, is this article's own reading.

Key Numbers

What companies asked most often once the law took effect is something the government counted itself. The question this submission brings is already on that tally.

Sources: Etoday · Newspim (ministry support desk tally) · Electronic Times (Enforcement Decree notice)

552

Support desk inquiries in the first ten weeks

January 22 to March 30; 262 by phone, 290 online

51%

Of online inquiries, on transparency labeling duties

The most asked subject of all

19.6%

Of online inquiries, on whether an AI is high-impact

That is, does this duty fall on us

1 year+

Grace period the ministry set before fines

Fact-finding continues; only the fine waits

1

Nine Months In, 19 Items Land at the Ministry

On October 29 FKI said it had collected its members' views and put 19 proposed improvements to the AI Basic Act before the Ministry of Science and ICT. The statute at issue is the Framework Act on the Development of Artificial Intelligence and Establishment of a Foundation for Trust, in force since January 22 of this year and nine months into its run.

FKI sorted the 19 into five areas.

  • Clarifying who and what the AI rules apply to
  • Making the duty to disclose and label AI use workable
  • Spelling out the rules for high-performance and high-impact AI
  • Tidying up the standards for using training data
  • Making any widening of the rules predictable

Open the contents page of the document FKI attached to its release and the weight of each area shows. Four items sit under operator status and scope of application, seven under transparency duties for generative AI, five under high-performance and high-impact AI, two under training data, and one under the predictability of any widening.

19 items piled into five areas Coverage led with the first item in the smallest area Scope & operator status 4 Disclosure & labeling duties 7 High-impact AI rules 5 Training data standards 2 Predictability of expansion 1
▲ FKI's 19 proposed AI Basic Act amendments by area — drawn by Pebblous from the submission's table of contents

By count the second area is the thickest, yet the press coverage led with the first item. More companies now build their own AI and run it for internal work or share it with an affiliate, and where an AI that runs that way sits in the law has not been settled.

Kwon Hyuk-min, who heads FKI's growth strategy office, said duties should be spelled out to match the real risk of the AI and the role the company plays in it, so that companies can put their effort into safe use and into innovation. He added that the AI Basic Act will be measured by how many companies on the ground come to trust AI enough to use it. The weight sits on getting a line drawn rather than on doing away with the rules.

That the same question had been piling up is visible in the government's own tally. The AI Basic Act support desk, run by the ministry with the Korea Artificial Intelligence and Software Industry Association, logged 552 inquiries between the law's start date of January 22 and March 30. Of those, 262 came by phone and 290 online. Sorted by subject, the online inquiries put transparency labeling duties first at 51 percent, with whether an AI counts as high-impact next at 19.6 percent. Around 68.9 percent of those who wrote in were companies, of which small and venture firms made up 36.2 percent and large firms 32.7 percent.

Put the two subjects together and roughly seven of every ten online inquiries were asking whether a duty lands on the company at all. That is not a question about how to comply but about whether the law is addressing you. When the ministry published its support desk casebook on March 31, it noted that companies were less puzzled by the provisions themselves than by how those provisions would apply to their own situation. This submission stands seven months downstream of that finding.

2

The Word the Law Uses Is 'Provide'

What is ambiguous becomes plain as soon as you open the text. Article 2, subparagraph 7 of the AI Basic Act splits the AI business operator into two kinds. One who "develops and provides" AI is a development operator, and one who uses that AI to "provide an AI product or AI service" is a deployment operator. The user in subparagraph 8 is the one who "is provided with" that product or service. Article 31, paragraph 1, which sets the transparency duty, says that where a company "intends to provide a product or service using high-impact AI or generative AI," it must tell "the user" so in advance.

The hinge on which the duty swings open or shut is providing, and the user. Yet neither word carries any qualifier pointing outward. When a company puts the AI it built on an internal portal and three thousand employees start using it, are those three thousand users who have been provided with a product, or simply people who work there? Read the text alone and it goes either way.

It is not that the rules ignore internal AI altogether. Article 23, paragraph 4, subparagraph 2 of the Enforcement Decree lets the advance disclosure and labeling duties be waived in whole or in part where the AI "is used only for the internal business purposes of an AI business operator." The support desk casebook gives the same guidance: AI bought from outside and used only for internal work leaves the company in the position of a user rather than an operator with transparency duties, and AI built in house for staff work can fall under this exception.

The exception, though, opens with the words "of an AI business operator." It assumes you are already an operator and lifts a disclosure duty; it does not decide whether you are one. FKI points at exactly that gap. Even with the exception in place, a company still has to work out its own status and the other duties that come attached to it. The casebook also notes that handing the AI to a third party can trigger transparency duties whether or not money changes hands, and that high-impact AI carries its obligations regardless of whether use stays internal.

The internal-use exception waives disclosure only It does not settle operator status — the Decree already assumes you are one Myth Internal use alone means no operator duties exist ≠ Reality Operator status is unchanged Only disclosure & labeling waived Still liable if high-impact/generative Operator status and the disclosure waiver are two separate questions
▲ What the Enforcement Decree's internal-use exception actually waives — drawn by Pebblous from the AI Basic Act

The item FKI placed first among the 19 is this boundary around providing — general internal-work AI that is not high-impact AI should be marked off from services provided outside. High-impact AI is defined separately in Article 2, subparagraph 4: AI that may have a serious effect on, or pose a risk to, human life, bodily safety and fundamental rights, used in a field the law lists, such as energy, healthcare, hiring assessment, credit assessment and transportation. Internal-work AI that lands in one of those fields is a different matter, and FKI assumes as much.

Trouble starts with everything outside that list. An AI that summarizes meeting minutes, looks up internal policy and helps review code appears nowhere among the high-impact fields. Operator duties, however, do not attach to high-impact AI alone. A product or service using generative AI picks up the disclosure and labeling requirements of Article 31. And the definition of generative AI in Article 2, subparagraph 5 covers a system that produces output "by imitating the structure and characteristics of input data," which can sweep in AI that was never built to make text or pictures. Hence FKI would narrow the definition to AI whose purpose is generating text, images, video and the like.

A complaint that the rules are heavy and a complaint that nobody can find the edge of the rules are two different complaints. The first asks for a duty to be lightened. The second only wants to be told which side of the line you are on, duty or no duty. The grace period on fines runs at least a year, so no money leaves the building today, but fact-finding goes ahead during that period all the same. The clock is running while companies still cannot say whether the law is addressing them.

3

If Affiliates Share the Model, Is That Inside?

FKI did not stop at saying a line is needed; it wrote down where to put it. Joint use across affiliates should count as internal work provided certain conditions hold, and there are three of them. The AI must not be provided outside, its users must be limited to internal ones, and access rights and security management must be in place.

All three ask how far the data reaches. Nothing in them touches the size or the capability of the model. The line FKI draws does not ask how clever an AI is. It asks only which fence the AI, and the data it was fed, sits inside.

That boundary is not a simple one, and FKI grants as much. Where a holding company and its affiliates build and run an AI together, it explains, the judgment is harder still because each affiliate is a separate legal person. Staff may share one building and one employee numbering scheme, and the legal identities stay separate all the same. If an employee of Company B uses a model Company A built, then on a literal reading of the text, Company A has provided its AI to another legal person.

The draft wording comes attached. A sentence would be added to Article 2, subparagraph 7 so that developing AI that is not high-impact AI "for the performance of work within the same legal person," or letting employees use it, is not regarded as providing. Affiliates come in the sentence after that: joint work among domestic and foreign legal persons standing in a controlling and subordinate relationship, as prescribed by Presidential Decree, should be treated the same way. The category in play here is not the affiliate but the controlling and subordinate relationship, and it is not confined to Korea. The fourth item follows that thread, adding that technology transfer or joint development inside a corporate group should not by itself bring a company within the law's extraterritorial reach.

The duty turns on providing, and where providing begins is left unwritten What settles the position is where the data came from and how far access runs Company staff Same entity, internal portal Affiliate staff Separate entity, joint build Outside customers Operator duties attach The line FKI proposes Where the law stays quiet today What you have to show to claim the left side No outside provision · internal users only · access and security management
▲ The dividing line FKI proposes and the conditions a company must meet to claim it. Drawn by Pebblous from the submission and Articles 2 and 31 of the AI Basic Act.

Even if the three conditions are granted as written, what follows is the company's own job. Conditions are not met by declaring them met. That the AI was not provided outside, that the people who used it were internal users and no one else, that access rights were in fact managed: each has to be shown. All three are facts about the past, so none can be manufactured now. They come only from records kept along the way.

4

Where the Builder's Duty Ends and the User's Begins

FKI's second axis is how responsibility is split. As the law stands, duties such as disclosure, labeling and risk management fall on the developing firm and the deploying firm alike. FKI wants them pulled apart. Risk that comes out of the development process and the technical makeup of the system belongs to the developing firm, and risk that arises in actual use and operation belongs to the deploying firm.

That fork shows most sharply at the moment the purpose changes. Where an AI built for ordinary business work is turned by another company into a high-impact AI service, FKI argues, the sensible holder of the risk management duty is the company that changed the purpose. It cites Article 25 of the European Union's AI Act, which treats a party that converts an existing AI into a high-risk AI as the provider of that AI and places the duties there. In that case the original provider ceases to hold provider status. Under that structure a duty moves rather than multiplies.

Duty does not add up, it moves Same principle as EU AI Act Article 25 — whoever changes the purpose carries the risk Developing firm Builds AI for general work Converting firm Turns it into high-impact AI Purpose changes → Exits provider status → Takes on risk duty The original developer stays the same; where the model stands has moved
▲ How FKI's proposal moves the risk-management duty when purpose changes — modeled on EU AI Act Article 25, drawn by Pebblous

The principle comes straight back to any company running AI internally. A model built for in-house use crosses over to an affiliate, the affiliate attaches it to hiring or credit assessment, and while the company that built it has changed nothing, the model itself moves into high-impact territory. For a rule that the party who changed the purpose carries the risk to work at all, who changed what, and when, has to be on record.

There is a request about the path information travels, too. The current law has operators publish on a website what they have done to meet their high-impact AI obligations, and much of what would be published is core technical information, so the range that can actually be disclosed is narrow, FKI points out. The other direction disappoints as well: a company deploying someone else's AI finds that the developing firm's duty to hand over material is only a best-efforts duty, which makes the information hard to get. So FKI's answer is to drop the general publication duty and write into the law that the developing firm must pass to the deploying firm the substance of its risk management and explanation plans, its user protection measures, and the name and contact details of the person who manages and supervises that AI.

The move is to narrow what goes out to the world and to raise what goes across to the firm that uses it into a duty. This is less about the volume of information than about the direction it flows. As things stand, with only a best-efforts duty, a company that picks up a model and asks what data went into it may get no answer back. With no answer, the deploying firm is stuck before it can even begin on the duties that have landed on it.

The request to be left out of high-impact AI comes with specific uses attached. FKI wants four of them named in the guidelines as falling outside: AI supporting preclinical research, literature and patent search and summarization, research data analysis, and general-purpose coding AI. Drug candidate discovery, which the press release put up front when describing how pharmaceutical and biotech firms use AI, is in fact absent from that list. The ministry's April guideline on judging high-impact AI advises that AI used as a tool at the research and development stage is not treated as subject to the obligations, but FKI says that explanation alone leaves research-support AI, which comes in many shapes, hard to judge. The same item would also set up a one-stop desk where the views of the Ministry of Health and Welfare, the Ministry of Food and Drug Safety and other relevant bodies could be confirmed quickly in one place.

Neither of the two remaining areas stops at its name. Training data accounts for only two of the nineteen items, yet for anyone who works with data it is the most familiar passage of the lot. A single AI takes in purchased data, public and open data, internal business data, overseas data, pseudonymized information and licensed databases, while the standards for whether all that is lawful lie scattered across the Personal Information Protection Act, the Copyright Act and contract terms. The permitted range shifts again depending on whether the work is training, fine-tuning, retrieval-augmented generation or embedding. So the first of the two calls on the ministry to weave the guidance of the relevant agencies into one set of guidelines organized by data type and manner of use, and to hand companies a standard checklist along with record items for data provenance. The other item would write research to develop, validate and improve AI into the AI Basic Act as scientific research within the meaning of the Personal Information Protection Act.

The last area holds a single item, which would have principles of advance notice and transitional relief settled before the high-impact AI fields are widened by Presidential Decree. Two implementing notices are to set the detailed compliance standards, and both have been stalled since they went out for public comment: the notice on high-impact AI operator obligations on March 12, and the notice on how to meet the safety assurance duty on July 8.

5

Why Pebblous Is Watching This Submission

Whichever way the legislation and the notices go, the question that follows a company saying "we are internal" is the same one. What will you show for it? The three conditions FKI set out, the principle that the party who changed the purpose carries the risk, the request that developing firms pass material to deploying firms: every one of them presumes a record. On the day the standard is fixed, what a company needs in hand is not an interpretation but that record.

This article is not alone in asking for records. Article 34, paragraph 1 of the law lists among the obligations of a high-impact AI operator the preparation and retention of documents by which its safety and reliability measures can be checked, and Article 27, paragraph 2 of the Enforcement Decree sets the retention period for those supporting documents at five years. Record items for data provenance are among the things FKI puts to the government: a call for it to draw up the form that says what goes in which box. Yet even if you wait for the form to arrive, the boxes can only be filled from what was kept in the meantime.

If you run AI inside your own company, four things are worth checking.

  • Is it written down what this model was fed on? Unless internal documents, data received from an affiliate and material bought from outside are told apart, the first half of any claim that the AI "only runs inside" is empty
  • Are the people who use it distinguished by legal entity? FKI's own point is that affiliates being separate legal persons makes the judgment hard, and an access log that records nothing but employee numbers puts your own staff and affiliate staff on the same line
  • Is there a history of access rights changing? The question will be who could open it last quarter. A screen that shows the current settings has no answer to that
  • Is the day the model's purpose changed recorded? If a model built as a summarizing tool was attached one day to assessment work, that day is the day responsibility moved

None of the four has anything to do with model performance. They are items whose answers would read the same after swapping in a better model, and none can be put together in a hurry. A company that has not been noting the provenance of its training data and the reach of its access cannot say which side of the line it falls on, wherever the line ends up. If FKI gets its way and internal-work AI is marked off from outside services, the benefit of the distinction goes only to companies able to show they are distinguished.

A grace period of at least a year before fines reads like breathing room. What has been deferred, though, is the fine, and records cannot be made in arrears. Once the standard is set, what remains is whatever was accumulated before it.

Thank you for reading this far. The contents of the submission were taken from the full document FKI attached to its October 29 release, and the statutory provisions were checked against the table in that document setting current law beside the proposed amendments, together with the text of the AI Basic Act. Not all 19 items are covered here. Those far from internal AI, such as where generative AI output must be labeled or how deepfakes are judged, were left out. If you run AI inside your own company, it is worth checking whether you could write down on one page today what went into that model and who uses it.

Pebblous Data Communication Team
October 30, 2026

R

References

Official Documents & Statutes

Press Coverage