Unadhi Guides

AI-output marking check sheet for software houses, web agencies, app makers and their accountants: Article 50(2) machine-readable marking, tools placed before 2 Aug 2026 due by 2 Dec 2026 (AI Act, Digital Omnibus Regulation (EU) 2026/1744)

Updated 2 Nov 2026 · Disclosure: the EURO AI Register mentioned below is an Unadhi product with a free edition. This guide is for information only, not legal advice.

What you get: a finished Article 50(2) AI-output marking check sheet for small software houses, app makers, web agencies and SaaS start-ups, and for the accountants and tax advisers who advise them. One row per AI feature that outputs text, images, audio or video: type in eleven short answers and the sheet returns the Article 50(2) timing (placed before 2 Aug 2026 and due by 2 Dec 2026, due from placing, not yet launched, exception claimed or not in scope), a status with the days left or overdue, and a to-do list citing the article behind each step. Also included: a 10-question checklist, a question letter for your AI model supplier, a customer questionnaire answer, a decision record and a client email for accountants. Formulas tested in LibreOffice on ten made-up rows.

Short answer: since 2 August 2026, any business that offers an AI tool under its own name that generates text, images, audio or video must make sure the outputs are marked in a machine-readable format and detectable as AI-generated (AI Act Article 50(2)). This sits with the provider of the tool, which includes small firms that wrap a bought-in model in their own app (Article 3(3), 3(68)). The Digital Omnibus on AI, Regulation (EU) 2026/1744, gave tools already on the market before 2 August 2026 until 2 December 2026 to comply (new Article 111(4)). That is 30 days from today. Tools launched since 2 August 2026 had no such transition. Features that only assist standard editing, or do not substantially alter the input or its meaning, are outside the duty. An open-source licence does not take a tool out of Article 50 (Article 2(12)).

Why this lands on the accountant's desk

Step 1: tab 1, Generators (one row per AI feature)

Create a sheet named Generators with headers in row 1 and the first feature in row 2. Create a second sheet named Settings and put =TODAY() in B1. Copy the formulas in L2 to N2 down as far as needed. In Excel 2019 or later, LibreOffice or Google Sheets, TEXTJOIN works as written.

ColumnWhat to writeSource
AAI feature or product that outputs content (for example "product photo background generator")You
BOutput: Text, Image, Audio, Video or MixedArt 50(2)
COffered under your own name or trademark? Y or NArt 3(3) provider
DMade available or used in the EU? Y or NArt 2(1)(a), (c); Art 3(9)-(10)
EDate first placed on the EU market (blank if not yet)Art 111(4) as added by Reg (EU) 2026/1744
FDoes it only assist standard editing, or not substantially alter the input or its meaning? Y, N or UnsureArt 50(2), last sentence; recital 133
GReleased under a free and open-source licence? Y or NArt 2(12)
HMarking method: None, Upstream model marks, Metadata, Watermark, Metadata + watermark or OtherRecital 133
IMark tested and detectable on real outputs? Y or NArt 50(2), second sentence
JDoes it also talk to people, for example as a chat? Y or NArt 50(1), 50(5)
KCould customers publish deep fakes or public-interest text with it? Y or NArt 50(4)
LArt 50(2) timing: =IF(A2="","",IF(NOT(AND(C2="Y",D2="Y")),"NOT IN SCOPE OF ART 50(2) ON THESE ANSWERS",IF(F2="Y","EXCEPTION CLAIMED: ASSISTIVE EDITING OR NO SUBSTANTIAL CHANGE (CHECK IT)",IF(F2<>"N","CHECK BY HAND: EXCEPTION UNCLEAR",IF(E2="","NOT YET ON THE EU MARKET: MARK FROM LAUNCH",IF(E2<DATE(2026,8,2),"PLACED BEFORE 2 AUG 2026: MARK BY 2 DEC 2026","PLACED ON OR AFTER 2 AUG 2026: MARKING DUE FROM PLACING"))))))Art 50(2), 113; Art 111(4)
MStatus (uses Settings!B1 = today): =IF(A2="","",IF(AND(LEFT(L2,3)="NOT",LEFT(L2,6)<>"NOT YE"),"NO ART 50(2) ACTION ON THESE ANSWERS",IF(OR(LEFT(L2,5)="CHECK",LEFT(L2,9)="EXCEPTION"),"CHECK BY HAND",IF(AND(H2<>"None",I2="Y"),IF(H2="Upstream model marks","MARKED VIA UPSTREAM MODEL AND TESTED: KEEP THE EVIDENCE","MARKED AND TESTED: KEEP THE EVIDENCE"),IF(LEFT(L2,6)="NOT YE","FIX BEFORE LAUNCH",IF(LEFT(L2,13)="PLACED BEFORE",IF(Settings!$B$1>DATE(2026,12,2),"OVERDUE BY "&(Settings!$B$1-DATE(2026,12,2))&" DAYS",IF(DATE(2026,12,2)-Settings!$B$1<=30,"GAP: "&(DATE(2026,12,2)-Settings!$B$1)&" DAYS LEFT: URGENT","GAP: "&(DATE(2026,12,2)-Settings!$B$1)&" DAYS LEFT")),"OVERDUE: REQUIRED SINCE PLACING ("&(Settings!$B$1-E2)&" DAYS)"))))))As above
NTo-do list: =IF(A2="","",TEXTJOIN("; ",TRUE,IF(C2<>"Y","not offered under your name or trademark: you are probably not the provider (Art 3(3)); if you use it, check deployer duties in Art 50(4)",""),IF(AND(C2="Y",D2<>"Y"),"not made available or used in the EU on these answers: re-check if EU customers can use it (Art 2(1))",""),IF(AND(AND(C2="Y",D2="Y"),F2="Y"),"write down why the output only assists standard editing or does not substantially alter the input or its meaning (Art 50(2), last sentence; recital 133)",""),IF(AND(AND(C2="Y",D2="Y"),F2<>"Y",F2<>"N"),"decide whether the feature only assists standard editing or does not substantially alter the input; if unsure, mark it",""),IF(AND(AND(C2="Y",D2="Y"),F2="N",H2="None"),"choose a marking method: recital 133 names watermarks, metadata, cryptographic provenance, logging and fingerprints",""),IF(AND(AND(C2="Y",D2="Y"),F2="N",H2="Upstream model marks",I2<>"Y"),"get written confirmation from the model supplier that its outputs are marked, and test it on your own outputs (recital 133; Art 3(68) downstream provider)",""),IF(AND(AND(C2="Y",D2="Y"),F2="N",H2<>"None",I2<>"Y"),"test that the mark survives download, export and your own post-processing, and keep the test result",""),IF(AND(AND(C2="Y",D2="Y"),F2="N",B2="Text"),"text is the hardest output to mark: record what you tried and why it is the state of the art for you (Art 50(2) second sentence; our reading)",""),IF(AND(AND(C2="Y",D2="Y"),G2="Y"),"open-source release does not remove Art 50 duties (Art 2(12))",""),IF(AND(C2="Y",J2="Y"),"it also talks to people: tell them they are interacting with AI unless obvious (Art 50(1)), at the latest at first interaction (Art 50(5))",""),IF(AND(C2="Y",K2="Y"),"customers may publish deep fakes or public-interest text: tell them about their own disclosure duty (Art 50(4); our reading, not a stated provider duty)",""),"record who decided, why and the date"))As cited in each item

Step 2: what the test rows returned

Tested in LibreOffice Calc on 2 Nov 2026 with Settings!B1 fixed at 2 Nov 2026, on ten made-up rows (no real business). All ten gave the intended result; an empty row stays blank; no formula errors.

Made-up rowArt 50(2) timing (L)Status (M)
1 Product photo background generator in a web shop app, placed March 2026, no markingPLACED BEFORE 2 AUG 2026: MARK BY 2 DEC 2026GAP: 30 DAYS LEFT: URGENT
2 AI blog writer SaaS on a bought-in model, relies on the model supplier, not testedPLACED BEFORE 2 AUG 2026: MARK BY 2 DEC 2026GAP: 30 DAYS LEFT: URGENT
3 Voice clone for explainer videos, launched 10 Sept 2026, metadata not testedPLACED ON OR AFTER 2 AUG 2026: MARKING DUE FROM PLACINGOVERDUE: REQUIRED SINCE PLACING (53 DAYS)
4 Spell and grammar fixer in an invoicing toolEXCEPTION CLAIMED: ASSISTIVE EDITING OR NO SUBSTANTIAL CHANGE (CHECK IT)CHECK BY HAND
5 Agency uses a third-party image generator for adsNOT IN SCOPE OF ART 50(2) ON THESE ANSWERSNO ART 50(2) ACTION ON THESE ANSWERS
6 Avatar video tool, launch planned for JanuaryNOT YET ON THE EU MARKET: MARK FROM LAUNCHFIX BEFORE LAUNCH
7 Open-source image model wrapper, metadata + watermark testedPLACED BEFORE 2 AUG 2026: MARK BY 2 DEC 2026MARKED AND TESTED: KEEP THE EVIDENCE
8 Logo generator sold only to US customersNOT IN SCOPE OF ART 50(2) ON THESE ANSWERSNO ART 50(2) ACTION ON THESE ANSWERS
9 Chatbot that writes product descriptions, exception unclearCHECK BY HAND: EXCEPTION UNCLEARCHECK BY HAND
10 Image upscaler marked via the model supplier, testedPLACED BEFORE 2 AUG 2026: MARK BY 2 DEC 2026MARKED VIA UPSTREAM MODEL AND TESTED: KEEP THE EVIDENCE

Step 3: the 10-question checklist

  1. Have we listed every feature in our products that generates text, images, audio or video, including features built on a bought-in model?
  2. Is each one offered under our own name or trademark, free or paid? If yes, we are its provider (Art 3(3)).
  3. Can people in the EU use it, or is its output used in the EU (Art 2(1)(a), (c))?
  4. When was it first made available on the EU market? Before 2 August 2026 means 2 December 2026; later means the duty applied from launch (Art 111(4)).
  5. Does it only assist standard editing or leave the input and its meaning substantially unchanged? If so, write down why (Art 50(2), last sentence).
  6. What marking do we use: metadata, watermark, cryptographic provenance, logging, fingerprints or a mix (recital 133)?
  7. If we rely on the model supplier's marking, do we have it in writing, and have we checked it on our own outputs?
  8. Does the mark survive download, export, resizing and our own post-processing?
  9. If the tool also chats with people, does it tell them they are dealing with AI at the first interaction, unless obvious (Art 50(1), 50(5))?
  10. Do our terms or help pages tell customers about their own duty to disclose deep fakes and public-interest text (Art 50(4))?

Step 4: templates

Question letter to your AI model supplier

Dear [supplier], we integrate [model name/version] into [our product], offered in the EU under our name. To meet Article 50(2) of Regulation (EU) 2024/1689 (by 2 December 2026 for our product, placed on the market on [date]): 1. Are outputs of the model, as we receive them through your API, marked in a machine-readable format? 2. Which method (metadata, watermark, other) for each output type: text, image, audio, video? 3. Which tool or method can we use to detect the mark? 4. Does the mark survive [our processing, e.g. resizing, format conversion]? 5. Will you tell us before you change or remove it? Please reply by [date].

Answer for a customer AI questionnaire

[Product] generates [output types] using [model/supplier]. We are the provider of [product] under the AI Act. Outputs are marked in a machine-readable format using [method], tested on [date] with [tool]; the test record is available on request. [Feature X] only assists standard editing and is outside Article 50(2) for the reason recorded in our decision log. Our users are told when they interact with an AI system (Article 50(1)). [Product] was first placed on the EU market on [date].

Decision record (one per feature)

Feature: [name]. Output: [type]. Provider: [us / not us, because]. First placed on the EU market: [date]. Art 50(2) timing: [from sheet]. Exception claimed: [none / reason]. Marking method: [method]. Test: [date, tool, result]. Decided by: [name, role] on [date]. Next review: [date].

Client email from the accountant or tax adviser

Hello [name], your [product] has an AI feature that writes or creates [content]. Under the EU AI Act, tools like this must mark their output as AI-generated in a machine-readable way, and for tools launched before 2 August 2026 the deadline is 2 December 2026. The attached one-row-per-feature sheet (11 answers each, about 10 minutes) shows whether each feature is affected and what is still open. It is not legal advice; for anything marked OVERDUE or CHECK BY HAND we suggest speaking to a technology lawyer.

Related free guides

Keep each AI feature, its supplier and its evidence in one place

Record each AI feature, its model supplier, the Article 50 answer and the evidence (supplier reply, marking test, decision record) in the EURO AI Register free edition. It runs in your browser with no account and no upload, covers your first 10 AI systems and 10 evidence records, the screening questions, duties by role with dates, and a printable Article 4 AI-literacy record.

When a client has more than 10 systems, the EURO AI Register membership starts at $19 a month and adds unlimited systems, rule updates and email help. Neither tool is legal advice or proof of compliance.

Partner programme opening soon

Partner programme opening soon: join the free waitlist. Planned terms are 50% of each referred sale (75% once you have 5+ paying customers), single level, paid through Gumroad only from revenue received.

Limits and open points

Read for this guide (2 Nov 2026, EUR-Lex): AI Act (EU) 2024/1689 Article 2(1) and 2(12), Article 3(1), 3(3), 3(4), 3(9), 3(10) and 3(68), Article 25(1), Article 50(1)-(7), Article 99(3), (4)(g) and (6), Article 113, recitals 133 and 134; Regulation (EU) 2026/1744 (Digital Omnibus on AI) amendment of Article 111 (new paragraph 4) and its recital on the four-month transition. We searched the Omnibus text for other mentions of Article 50 and found only those two (our check).

Not read for this guide: any Commission guidelines or FAQ on Article 50, any code of practice on marking and labelling AI-generated content under Article 50(7) or an implementing act approving one, technical standards and industry provenance specifications for watermarks or metadata, Article 53 duties of general-purpose AI model providers, and national penalty rules under Article 99(1)-(2).

Our reading, not the law's words: the column list and grades, "mark from launch" for tools not yet placed, treating an untested mark as a gap, the 30-day "URGENT" window, the note that text is the hardest output to mark, the advice to tell customers about their Article 50(4) duty, and the sample wording in the templates. Article 25(1), which turns a deployer into a provider when it rebrands or modifies a system, is written for high-risk AI systems; whether a white-label generator is "yours" for Article 50 follows from the Article 3(3) definition (our reading).

Frequently asked questions

What does Article 50(2) of the AI Act require?

Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content must ensure that the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The technical solutions must be effective, interoperable, robust and reliable as far as technically feasible. The duty does not apply to the extent the system performs an assistive function for standard editing or does not substantially alter the input data provided by the deployer or its semantics (Regulation (EU) 2024/1689, Article 50(2)).

Is there extra time for generative AI tools already on the market?

Yes, a short one. Regulation (EU) 2026/1744 (the Digital Omnibus on AI) added Article 111(4): providers of AI systems generating synthetic audio, image, video or text content that were placed on the market before 2 August 2026 must take the necessary steps to comply with Article 50(2) by 2 December 2026. Systems placed on the market from 2 August 2026 get no transition under that paragraph.

We build on a bought-in AI model. Are we the provider?

If you develop an AI system, or have it developed, and place it on the market or put it into service under your own name or trademark, whether for payment or free of charge, you are a provider under Article 3(3). Article 3(68) calls a provider that integrates someone else's AI model a downstream provider. Recital 133 says marking can be implemented at the model level, which helps the downstream provider, but the Article 50(2) duty itself sits with the provider of the AI system.

Does an open-source licence exempt us?

No. Article 2(12) says the AI Act does not apply to AI systems released under free and open-source licences unless they are placed on the market or put into service as high-risk AI systems or as an AI system that falls under Article 5 or 50.

Not legal advice. The sheet helps sort AI features one by one; it does not decide legal questions. Spotted an error? Write to mininglife7@gmail.com and we correct it.

Sources

← All guides

Unadhi Guides are free, independent explainers. We correct errors as soon as we find them: write to mininglife7@gmail.com.