PERADI has issued a 48-page ethics guideline telling Indonesian advocates how generative AI may and may not be used in legal work

What this document is, and what it is not

PERADI, the Indonesian advocates association, published a 48-page guideline on the use of generative AI by advocates. The text was drafted by the association's Komite Tetap Teknologi Hukum dan Inovasi and is signed by PERADI's Ketua Umum, Ahmad Fikri Assegaf, and its Sekretaris Jenderal, Emir Z. Pohan.

It is worth being blunt about the register, because trade coverage of documents like this tends to drift. This is not a statute. It is not a government regulation, not a ministerial decree, and not a court rule. Nothing about it went through the Indonesian legislature or Komdigi. It is a professional association telling its own members how existing professional obligations apply to a new tool.

The document places itself under three existing instruments it names by number: Law 18/2003 on Advocates, the Kode Etik Advokat Indonesia, and Law 27/2022 on Personal Data Protection. Bagian I records that the drafters worked from the ABA's Formal Opinion 512, State Bar of California guidance, the CCBE guide, guidance from the Law Society of New South Wales and the Law Institute of Victoria, and Singapore's Ministry of Law guide, then adapted the result to Indonesian law.

It carries no serial number. We looked. The word Nomor appears in the text only in references to Law 18/2003 and Law 27/2022, never attached to the guideline itself, and the PDF cover page carries only a title and the month.

Three dates are in circulation. Here is the one we used and why

This document arrives with three plausible dates attached, which is exactly the situation that produces wrong bylines.

The PDF filename embeds 2026-08-07 at 18.22, which is the digital signing timestamp. The announcement page's article metadata reports a published time of 7 August 2026 and a modified time of 8 August 2026. Neither is the instrument's date.

Bagian XVI, the closing page of the guideline itself, reads: adopted in the Special Capital Region of Jakarta on 6 August 2026, followed by the two signature blocks. The same page states that the guideline takes effect from the date of adoption. The launch announcement independently carries a 6 August 2026 Jakarta dateline for the hybrid talkshow at which the guideline was released.

So we date this 6 August 2026: it is the date the document assigns to itself, in its own operative closing clause, and it is corroborated by the launch dateline. The 7 August signature file timestamp is when the PDF was cryptographically signed, which is a file event, not an adoption event.

The confidentiality rule says less, and more, than the headline version

The confidentiality provision at 2.5 is the one clause most likely to be quoted badly. In full, in the original: "Advokat tidak boleh memasukkan informasi yang bersifat rahasia dari klien ke dalam alat Gen AI yang menimbulkan risiko material terhadap kerahasiaan tanpa persetujuan klien."

That is not a ban on putting client information into AI tools. It has two conditions doing real work. The prohibition attaches where the tool creates a material risk to confidentiality, and it lifts where the client has consented. Read plainly, it is a consent-and-risk rule, not an abstinence rule.

The clauses around it fill in what a material risk looks like. Paragraph 2.6 treats entering confidential client information into a public Gen AI tool that trains on user input as equivalent to disclosing that information to a third party, and directs advocates who need such a tool to disable data retention and training, anonymise, or move to an enterprise product with adequate confidentiality assurances. Paragraph 2.7 requires cleansing or redaction of sensitive details before a document is processed, listing party names, addresses, fee and price figures, transaction types and values, and identity numbers.

Paragraph 2.8 then does something unusual for a document of this kind, and it is the paragraph we would put in front of a client. It states that even after cleansing and redaction, uploaded documents may still be used by the AI platform as training data or for other purposes outside the advocate's control, and that redaction is therefore an additional safeguard rather than a substitute for choosing an adequate tool and obtaining client consent. That is a drafter refusing to let a procedural step launder a substantive risk.

One more precision point. Paragraph 8.3 says the consent clause inserted into an engagement letter may not be boiler-plate and must be informed consent. A generic AI paragraph in a retainer will not satisfy the text as written.

Obliged, or merely encouraged: the distinction the text keeps and summaries drop

Indonesian drafting distinguishes wajib, an obligation, from dianjurkan, a recommendation. This guideline switches between them clause by clause, and the pattern is not random.

On the obligation side: an advocate is obliged to read and understand the terms of use and privacy policy of the platform in question, so as to know who can access the data entered (2.3). Confidential correspondence between counsel, including Sans Prejudice letters and settlement communications, may not be entered into a tool with leakage potential (6.3). Disclosure to the client is mandatory in defined circumstances (8.2). Time-based billing must reflect time actually spent, including prompt drafting and output review, and an advocate may not bill the client for time spent learning a tool as part of maintaining competence (8.6, 8.7). The incident chapter, Bagian XII, is written almost entirely in wajib: stop or isolate the affected tool, preserve logs and evidence, notify the vendor, and meet the UU PDP breach notification duty where personal data is involved. Trainee advocates are required to keep a usage logbook recording the tool, the purpose and the output used, and core legal reasoning exercises must be done without AI (14.3, 14.6).

On the recommended side, and this is where the widely repeated checklists overreach: risk classification of a matter before using Gen AI is dianjurkan (2.3, 2.15). The written internal firm AI policy, with its governance structure, approved tool list and incident reporting procedure, is dianjurkan (5.3). Vendor due diligence before adopting a tool is dianjurkan (5.4), as is preferring vendors certified to ISO/IEC 27001, 27701 or 42001 (11.2). Access controls are dianjurkan (5.2).

The risk tiers themselves are worth understanding because the consequence attached to them is mandatory even though the classification step is not. Paragraphs 2.15 to 2.18 set Low, Medium and High. Gen AI may be used only for Low and Medium risk work; High risk work is prohibited with Gen AI. High risk is defined by example as deepfakes or fabricated evidence, full delegation of advice or advocacy to AI, autonomous filing to a forum without advocate review, and predictive assessment substituting for professional judgement.

We also found an internal inconsistency readers should know about, because it changes what an engagement letter should say. Paragraph 2.4 lists the disclosure triggers as: the client asks, the engagement agreement requires it, or the use affects an important decision in the matter or affects fees. Paragraph 8.2, the detailed provision, lists the third trigger differently, as the advocate being about to enter confidential client information into a Gen AI tool with the client's consent. The two paragraphs do not carry the same third trigger. That is our reading of the text, not a position PERADI has stated.

How it binds, and what a US instructing team can do with it

The enforcement route is the bar's own ethics jurisdiction. Bagian X opens by stating that the listed acts are prohibited for any purpose, including efficiency, client interest, experimentation or urgency, and that breach may be qualified as a code of ethics violation or professional misconduct forming the basis for action. Paragraph 10.2 adds that the same conduct may draw forum sanctions including document rejection, costs and contempt of court, civil or criminal claims, and action by the Dewan Kehormatan Advokat.

That is different in kind from a statutory duty. There is no regulator issuing fines here and no private right of action created by the guideline. What exists is a written standard that a disciplinary body can measure a member against, plus a document a court can be pointed at.

The twelve prohibited acts in 10.1 are worth reading in full, but the shape is: creating or using deepfakes or misleading synthetic media, including as evidence, testimony or persona; fabricating documents, regulations, evidence, citations or official documents with AI; submitting AI-invented citations, regulations or judgments to a forum; fully delegating professional judgement, advice or advocacy to AI; letting an autonomous AI system file documents or communicate with a forum without advocate review; using an AI persona or avatar posing as a human or a real advocate without disclosure to the forum; manipulating electronic evidence; entering client secrets into AI tools without safeguards or consent, or uploading confidential documents to public tools; using AI for surveillance, interception or discriminatory profiling; using AI to draft intimidating, harassing or demeaning communications; and substantially reproducing third-party copyrighted work.

For a US firm or in-house team instructing Indonesian counsel, the practical value is narrow but real. There is now a named local standard to reference in an engagement letter rather than exporting your own outside counsel AI guidelines and hoping they translate. Conduct in Bagian X is not a judgement call your Indonesian counsel gets to weigh against deadline pressure; the text closes that door explicitly. And if you want a specific contractual hook, the mandatory items are the better ones to cite, since the firm policy and vendor due diligence provisions are drafted as recommendations and counsel can accurately say so.

One asymmetry to plan around. Paragraph 3.1 requires independent verification of Gen AI research output against official sources, confirming that a judgment or regulation actually exists, is quoted accurately and is still good law. Paragraph 3.2 discourages using one Gen AI tool to verify another, on the stated ground that models will confirm the existence of judgments that do not exist. Verification against Indonesian official databases is a manual step with a real time cost, and a fee estimate built on AI-accelerated research that ignores it will be wrong.

What we did not verify

What we opened: the signed 48-page Indonesian PDF at peradi.id in full, including Bagian I through XVI, the five annexes and the source list, and the PERADI launch announcement page with its metadata. The verbatim quotation at 2.5 reproduced above was character-matched against that extraction. Paragraph numbers, the wajib and dianjurkan classifications, the Bagian X list, the signature block and the 6 August 2026 adoption date all come from that text.

What we did not open: any English-language version, official or otherwise. We did not locate a PERADI decree, circular or numbered instrument adopting or transmitting the guideline, and we cannot rule out that one exists somewhere off the association's public site. We did not open the Dewan Kehormatan Advokat's procedural rules, so we did not confirm how a Bagian X allegation is actually docketed, heard or appealed. We did not check whether PERADI's rival advocate organisations have adopted, rejected or ignored this text, which matters because Indonesia's advocate organisation landscape is not unitary. We did not read Law 18/2003 or the KEAI directly for this piece; we report only what this guideline says about them.

What we refuse to claim: that this instrument has statutory force, or any force at all outside PERADI's own membership and ethics process. That any of it has been applied to a real advocate. That the recommended provisions, including the internal AI policy and vendor due diligence, are obligations, because the Indonesian text does not say so. That the document has a serial or gazette number, because none appears in it, which under this desk's own instrument-number rule makes it a borderline candidate for coverage and is why we have flagged it here rather than buried it. The reading of the 2.4 and 8.2 disclosure-trigger mismatch is ours, and PERADI has not addressed it. Any English rendering of the Indonesian in this piece is our translation and the Indonesian text governs.

Related briefings

• ABA Formal Opinion 512 on generative AI and lawyer ethics
• CCBE guide on generative AI for European lawyers
• Singapore MinLaw guide for generative AI in the legal sector
• NSW Law Society guide to responsible AI use by solicitors
• AI Regulation News hub