What the draft does
The guideline is the work of a special subcommittee on AI legal technology within the Ministry of Justice, chaired by an academic. Its central move is definitional. It creates a category — judgment-type AI legal-tech services — covering services in which a user enters concrete facts, the system applies law to those facts, and the output is a legal conclusion about that user's situation.
Draft Article 2 provides that a lawyer may not supply such a service to general consumers for a fee. Services that stay on the other side of the line — explaining what a statute says, describing a procedure, summarising how a type of dispute generally works — remain available for paid supply.
The model is Japan's 2023 Ministry of Justice guideline on AI contract review, which addressed a structurally similar question: when does an automated service cross from tool into the practice of law reserved to qualified persons.
Why a ministry guideline is the instrument
The underlying statute is the Attorney-at-Law Act, which reserves the handling of legal affairs to qualified attorneys. The question of whether a software service handling legal affairs falls inside that reservation is genuinely unresolved, and it is not a question a ministry can answer by decree.
What a guideline can do is state how the Ministry reads the statute, which tells providers where prosecution risk sits. That is a weaker instrument than legislation and a stronger one than commentary, and it is the same mechanism that makes KVKK's guidance consequential in Türkiye without being binding.
Korea has particular reason to reach for a guideline rather than wait for legislation. The long dispute over the Lawtalk platform — a lawyer-matching service the bar association fought for years, with disciplinary action against participating lawyers eventually overturned — demonstrated that leaving the boundary to professional self-regulation produces years of conflict, inconsistent outcomes and no clarity. The Ministry appears to have concluded that stating a position early is better than another decade of that.
The fabricated-citation data
Follow-up reporting on 23 September supplied the evidentiary backdrop, and it is the part of this story that should travel furthest.
Two figures were reported from the Korean judiciary. In a case before the Patent Court, eight precedents cited in submissions were found not to exist. In a judicial training session, thirty-nine of forty precedents circulated were false.
The second figure is the one to sit with. That is not a litigant cutting corners; it is material in a professional training context, where the participants are judges, and where the base rate of fabrication was almost total.
Citation fabrication is by now a documented international phenomenon — courts in the United States, the United Kingdom, Canada, Australia and elsewhere have sanctioned practitioners for filing submissions containing invented authorities, and the running tally is well into the hundreds. What the Korean figures add is a sense of density. A single fabricated citation is an individual failure. Thirty-nine of forty is a systemic property of the tool as used.
It also explains the shape of the draft. If the failure mode were occasional error, the answer would be verification duties. If the failure mode is that the output is confidently wrong at scale and its wrongness is invisible to a non-expert, then the answer has to address who is allowed to receive the output unmediated. That is precisely what a consumer-facing restriction does.
The asymmetry problem
The objection from Korean practitioners is immediate and hard to answer: the rule binds domestic providers and lawyers, while general-purpose models operated from abroad are subject to none of it.
A Korean consumer who wants a legal conclusion about their own facts can obtain one in seconds from a general assistant that has no Korean regulatory relationship, no professional insurance, no disciplinary exposure and no obligation to be right. What the guideline restricts is the domestic service that would have been built by people subject to Korean professional regulation, supervised by a bar association, and reachable by a Korean court.
This is not a novel problem — it is the standard shape of platform regulation, and it recurs in every jurisdiction that regulates a domestic industry facing an unregulated foreign substitute. But it has a sharper edge here, because the regulated activity is one where the quality difference is exactly what is at stake. A Korean legal-tech service that is prohibited from giving judgment loses to a foreign chatbot that gives judgment badly. The consumer is worse off, the domestic industry is worse off, and the only party helped is the unregulated provider.
The defensible answer is that the restriction attaches to the lawyer rather than to the software, and that professional regulation has always applied to professionals rather than to everyone who offers an opinion. A person may give their neighbour bad legal advice; a lawyer may not sell it. That framing is coherent, and it is probably the right one, but it does not make the competitive asymmetry disappear.
Where the line actually falls
The information-versus-judgment distinction is intuitive and considerably harder to operate than it sounds.
Consider a gradient. A service that reproduces the text of a statute is plainly information. One that explains what the statute means in general terms is still information. One that explains what it means for a described type of situation is beginning to move. One that asks the user for their facts and tells them which provision applies has arguably crossed. One that tells the user whether their claim is time-barred has certainly crossed. One that drafts a document reflecting that conclusion has crossed twice.
The difficulty is that the transition is continuous, whereas the rule is binary — and providers will build precisely at the boundary, because that is where the product value is. Expect services phrased conditionally: not "your claim is time-barred" but "claims of this type are generally time-barred after this period; if the facts you have entered are accurate, this would ordinarily apply." That is judgment wearing the grammar of information, and it will be the central interpretive fight.
The Japanese guideline encountered the same problem in the contract-review setting, and the workable distinction that emerged there was about the role of the human professional: a tool whose output is reviewed by a lawyer before it reaches the client is a tool, while a tool whose output reaches the client directly is something else. That test is more administrable than the content-based one, because it asks about a process fact rather than about where a sentence sits on a gradient.
What it does to the business model
The restriction is on paid supply to general consumers, and each of those three elements is a design instruction that providers will follow.
"Paid" points toward indirect monetisation. A free conclusion that generates a referral to a lawyer, with the lawyer paying for the referral, delivers the same product without the prohibited transaction. Whether that survives scrutiny depends on how the guideline treats value flowing through a channel other than the user's payment — and referral arrangements are themselves contested territory in Korean professional regulation, which is what the Lawtalk dispute was substantially about.
"General consumers" points toward business customers. A service supplied to companies, in-house teams or law firms sits outside a consumer-facing restriction. This is where most legal-tech value has been anyway, and the likely effect is to accelerate a move that was already happening: away from the consumer market, which is large but hard to monetise and now regulated, toward the enterprise market, which is smaller, better funded and unrestricted.
"Judgment" points toward the reviewed-output model. A service where the system drafts and a qualified lawyer approves before delivery is a productivity tool rather than a regulated supply. This is the most likely destination, and it is a considerably better outcome than the debate suggests — it preserves the efficiency gain, keeps a professional accountable, and gives the consumer someone to sue.
The effect of the draft, if adopted, is therefore not to remove AI from Korean legal services. It is to require a human professional in the delivery path for consumer-facing conclusions, and to push everything else toward business customers. Whether that is the right allocation is a policy question. It is at least a coherent one.
The liability gap underneath
There is a question the draft does not address and that no jurisdiction has answered: who is responsible when an AI legal service is wrong.
For a lawyer the position is settled. Professional negligence, disciplinary jurisdiction, mandatory insurance in most systems, and a client relationship with defined duties. The client who receives bad advice has a defendant, a standard against which to measure the advice, and usually a fund to recover from.
For a consumer who receives a bad legal conclusion from a software service, essentially none of that exists. The terms will disclaim advice, disclaim reliance and cap liability at the subscription fee. There is no professional standard to apply, because the provider is not a professional. There is no insurance requirement. Where the provider is abroad, there is a jurisdiction problem before any of the substantive ones.
This is the strongest argument for the Korean approach, and it is one the draft makes only implicitly. The reservation of legal services to qualified persons is not fundamentally about qualification. It is about attaching consequences to advice. A regime that lets a service give legal conclusions without attaching those consequences has not deregulated the activity; it has removed the remedy while keeping the harm.
The honest counter-argument is that the alternative for many consumers is no advice at all, and that a flawed answer may beat silence for someone who cannot afford a lawyer. That argument deserves more respect than professional bodies usually give it. But it is an argument for making accurate advice cheaper, not for making inaccurate advice unaccountable — and the fabricated-citation figures suggest the current gap between those two is wider than the optimistic case assumes.
What this means for Türkiye
Türkiye has the same statutory architecture and has not yet had the argument.
The Attorneys' Act reserves the provision of legal services, and Article 35 in particular reserves categories of legal work to advocates. A software service that receives a Turkish consumer's facts and returns a conclusion about their legal position engages that reservation on its face, and there is no Turkish guideline, bar association position or reported decision that resolves whether it falls inside.
The Union of Turkish Bar Associations has contributed to KVKK's guide on data protection in legal practice, published the day after the Korean draft was reported. That guide addresses what a lawyer may do with client data. It does not address what a non-lawyer service may do with a consumer's legal question, which is the separate and unasked question.
Turkish legal-tech products are being built now. Some of them will sit on the judgment side of the Korean line. The sensible time to settle the boundary is before a disciplinary proceeding or a criminal complaint under the unauthorised practice provisions forces it to be settled on bad facts, in a single case, with no opportunity for the profession to consider the policy.
There is also a consumer dimension that the professional-monopoly framing tends to obscure. The reason to care about a service that gives legal conclusions to consumers is not principally that it competes with lawyers. It is that the recipient cannot evaluate the output. A consumer who is told their claim is time-barred has no way to detect that the authority supporting the answer does not exist. The thirty-nine-of-forty figure is what that looks like at scale, and it is a consumer protection fact before it is a professional one.
What to watch next
Whether the draft is adopted in its reported form, and in particular whether Article 2's fee condition survives — a restriction on paid supply invites the referral workaround, while a restriction on supply as such would be far broader and far more contested. Whether the Korean Bar Association aligns with the Ministry's position or seeks something stricter, given its history on Lawtalk. Whether the labelling obligations under Korea's AI Framework Act and its transparency guideline are extended to reach legal conclusions delivered by general-purpose assistants, which is the only route that touches the foreign providers. Whether any Korean court is asked to rule on the Attorney-at-Law Act boundary directly, which would settle far more than a guideline can. And whether the Union of Turkish Bar Associations, having just worked with KVKK on the data question, takes up the scope question while it is still a policy exercise rather than a disciplinary one.
Frequently asked questions
Does the draft ban AI in Korean legal practice?
No. It addresses paid, consumer-facing services that deliver legal conclusions on the user's own facts. Lawyers using AI internally, and services supplying legal information without applying law to a specific situation, are outside the restriction as drafted.
Is it in force?
No. It is a draft guideline reported in September 2026, and guidelines are statements of the Ministry's position rather than binding rules. It signals enforcement posture; it does not itself change the Attorney-at-Law Act.
Why does it bind Korean providers and not foreign chatbots?
Because it operates through professional regulation, which attaches to lawyers and to services supplied by them. A foreign general-purpose assistant has no Korean professional relationship to regulate. This is the principal criticism of the draft and it has not been answered.
Are the fabricated-citation figures unusual?
The phenomenon is international and well documented, with sanctions imposed on practitioners in many jurisdictions. The Korean figures are notable for density rather than novelty — thirty-nine false precedents out of forty in a judicial training context is a different kind of number from isolated incidents.
Would a Turkish service face the same restriction?
There is no Turkish equivalent of this guideline, and the question is unresolved. The Attorneys' Act reserves legal services, and a service delivering legal conclusions on a consumer's facts engages that reservation — but nobody has yet stated where the boundary falls, and no reported decision settles it.
What is the practical test a provider should apply?
Ask whether a qualified professional reviews the output before it reaches the user. A reviewed output is a tool assisting a lawyer. An unreviewed conclusion delivered to a consumer is the thing every regime discussed here treats as the regulated activity.
Burhan Doğuş Ayparlar's View
This section sets out my personal assessment as the founder of this site and an AI ethics & compliance counsel.
I am instinctively suspicious of professional bodies regulating technologies that compete with their members, and I want to declare that before saying that I think the Korean Ministry is substantially right.
The reason is the thirty-nine out of forty. Strip away the professional-monopoly framing and what remains is a consumer receiving a confident legal conclusion, supported by authorities they cannot check, from a system with no obligation to be correct and no one to answer for it when it is not. That is not a competition question. It is the ordinary case for regulating the supply of advice that the recipient cannot evaluate — the same case that underlies the regulation of medical advice, investment advice and structural engineering. Nobody argues that a licence requirement for surgeons is protectionism, and the argument is not stronger here merely because the tool is cheap.
Where I think the draft is weak is the place its critics have identified, and the weakness is real rather than rhetorical. A rule that binds the domestic, insured, disciplinable provider while leaving the consumer free to obtain a worse answer from an unregulated foreign service does not protect the consumer. It relocates them. Korean consumers will not stop seeking legal conclusions from machines because Korean lawyers may not sell them; they will get them from a general assistant that was not built for the purpose, has no Korean law in its evaluation set, and will never be sued. If the objective is consumer protection, the instrument has to reach the service the consumer actually uses, and professional regulation cannot do that. That is an argument for addressing it in the AI or consumer-protection track rather than the professional one — Korea has an AI Framework Act and a transparency guideline, and a labelling or disclosure obligation on AI-generated legal conclusions would reach the foreign provider in a way the Attorney-at-Law Act never will.
For Türkiye my concern is that we will have this argument later and worse. The pattern I expect is the one we have seen in every adjacent area: products get built, they sit on the judgment side of a line nobody has drawn, and the boundary is eventually settled by a disciplinary file or a criminal complaint against one unlucky provider, on facts chosen by whoever complained. That is the worst possible way to make policy, because the resulting rule is shaped by an individual case rather than by any considered view of what consumers need. The Union of Turkish Bar Associations has just demonstrated, by contributing to KVKK's guide, that it can work with a regulator to produce something practical before a dispute forces it. The same exercise on the scope question — what a legal-tech service may tell a consumer without an advocate in the loop — is the obvious next piece of work, and it should be done now, while it can still be done calmly.
One last observation, because it connects to the KVKK guide published the following day. Korea is asking what an AI service may tell a consumer. Türkiye has just asked what a lawyer may put into an AI service. Those are the two ends of the same pipe, and a profession that answers only one of them has a policy on neither. The Turkish answer to the first question is currently silence, and silence is not a neutral position — it is a decision to let the market settle it.
This article is for information only and does not constitute legal advice. The facts concerning Korea are based on press reporting of 21 and 23 September 2026; the full text of the draft guideline was not available at the time of writing and may change before adoption. Statements about Turkish law are general in nature; specific cases require individual assessment. The analysis and assessments are the author's own.
