Substack writers are now regulated AI users

How the AI Act applies to Substack writers

Aug 14, 2026

Everyone has an opinion on AI slop these days.

Some think that any content that has been touched by AI, even a little bit, is automatically tainted and does not deserve to be consumed. If there is even a hint of interference from those evil black boxes, the whole thing becomes blasphemous puddle of hot, stinking slop.

This debate really started to heat up when Substack introduced an AI detection tool provided by Pangram. With this tool, users can scan whether a Substack newsletter or Note has been generated by AI.

My good friend @Carey Lening here on Substack has written a brilliant piece on the GDPR implications of the Pangram tool and the rights available to Substackers who want to opt out of it. I encourage you to read it and her piece has actually led to users being able to successfully exercise their ability to opt out of the tool.

Privacat Insights

Score one for the GDPR

So, it took a minute, but a little over a week after I posted my ‘One Weird Trick’ post on how to object to Substack’s ‘Scan For AI’ feature, under the GDPR’s Right to Object, I’m happy to report that it worked…

10 days ago · 50 likes · 19 comments · Carey Lening

In my view, AI slop is ultimately a choice. AI is a tool that you can use for brainstorming, researching and also writing. And like with any tool, if you don’t learn how to use it effectively, then whatever you produce with that tool is unlikely to be very good. AI is not different - if you just try to one-shot prompt everything, expecting it to magically do everything you want to a higher quality, then you are clearly using it in the least optimal way possible and you will only get slop out of it.

AI is not god-in-a-box. It is a tool that needs direction and you are in charge of giving it that direction.

AI slop is a choice

Mahdi Assan

Apr 17

Read full story

And this is what bothers me about the Pangram tool. The insinuation it reinforces is that any use of AI is bad. The behaviour I therefore think it will end up encouraging among Substack writers and readers is rather counter-productive. Readers will use the Pangram scoring as a unquestionable signal to determine quality without actually taking the time to read the text and judge for themselves whether it is good or not. And, accordingly, writers already using AI in their work will just use it even more to make it sound less like AI.

The frontier labs have also now introduced their own means for detecting AI text. Earlier this week Anthropic announced a mechanism for embedding machine-readable watermarks into text generated by its models. A method for detecting these watermarks is also “forthcoming.”

These measures come as the AI transparency obligations under the EU’s AI Act come into force. From 2 August 2026, users of AI systems are required, in certain circumstances, to disclose their use of AI if it has been used to generate or manipulate text that they publish.

I have seen a lot of confusion on Substack about this and about whether and how these obligations apply to writers. Do these obligations apply to the average Substacker? And if so what is required exactly? Does it depend on how they use AI? Does it depend on what they write about? And so on.

So I thought I would write a post that answers all of these questions and explains how the AI Act applies to Substack writers and the steps they can take to follow its AI transparency rules.

It should be noted though that the views expressed here are all my own and is provided for general information only and does not constitute legal advice on which you should rely. You must obtain professional or specialist advice before taking, or refraining from, any action on the basis of the content in this newsletter.

In any case, I hope it is a useful aid that clarifies some of the confusion floating around on this topic.

Does the AI Act apply to me?

The first important question to ask is whether the AI Act even applies to you as a writer. This therefore concerns the scope of the Act in terms of its substance and its territorial reach.

Regarding substance, Article 2(1)(b) states that the Act applies to deployers (i.e., users) of AI systems (e.g., Claude).1 One exception to this is where users are individuals “using AI systems in the course of a purely personal non-professional activity.”2

But further clarification on the scope of this exemption comes from the European Commission. Its guidelines state that “[a]ny activity through which natural persons gain an economic benefit on a regular basis or are otherwise involved in a professional, business, trade, occupational or freelance activity should be considered as a ‘professional’ activity.”3

So if you’re a Substack writer, how do you know if the AI Act applies to you?

For one, if you run a paid newsletter, then you are likely in scope. Paid newsletters would constitute an activity that gains an economic benefit on a regular basis, since subscribers would be paying on a monthly or yearly basis.

This would be the case even if the paid newsletter consists of ‘personal’ content. The way the provision is written suggests that the exemption is cumulative, meaning that it is not about whether the newsletter is a purely personal or non-professional activity. Therefore, to benefit from the exemption, you would need to be acting in both a personal and a non-professional capacity.4 Accepting money for your newsletter dis-applies that second condition regardless of how personal your content might be.

Now what about free newsletters? Are they exempt? Well, it depends.

The Commission guidelines mention activity that is part of a professional, business, trade or occupational or freelance undertaking. So even a free newsletter that builds a professional profile, generates client work, or operates as an extension of your practice is plausibly caught even if the newsletter itself does not generate any revenue itself. A genuinely hobbyist newsletter with no commercial dimension is probably the only way to benefit from the exemption and not have the Act apply to you.

But then there is the territorial scope of the legislation. Clearly the Act applies to writers based in the EU,5 but it also applies to those based outside the EU, as per Article 2(1)(c), where “the output produced by the AI system is used in the Union.” In essence, if you have EU readers of your newsletter, the Act applies to your use of AI in your newsletter, though this depends on exactly how you use AI, which is addressed in the next section.

What transparency obligations apply to me?

The short answer here is that even if you are a deployer by virtue of your professional status as per the Act, its transparency obligations only apply to you if your newsletter:

Includes AI-generated or manipulated text
Is published
Has the purpose of informing the public
Is on matters of public interest
Is not subject to human review or editorial control and editorial responsibility

Let’s break down each of these conditions.

Includes AI-generated or manipulated text

Article 50(4) mentions “Deployers of an AI system that generates or manipulates text…”, which means it applies to writers using systems like Claude or ChatGPT to generate or manipulate their newsletters.

Under the Commission guidelines, “generation” means the AI produces synthetic material, typically from a prompt.6 “Manipulation” means the AI alters content that already exists, whether synthetic or human-written.7

There are three further points to mention here.

Firstly, the content in your newsletter does not need to be solely AI-generated.8 Newsletters that are a mixture of human-written and AI-written text counts, so there is majority-human safe harbour. Additionally, even if you generate sections of your newsletter with AI and rewrite it, this also still counts.

Secondly, the threshold here is AI generation or manipulation that goes beyond “standard editing.” In the Commission guidelines, “standard editing” means “preparing existing content for publication or distribution” and “does not involve generating new content”. This includes things like checking grammar, spelling, readability, format, and accessibility. But editing crosses this line if AI is used to change text “in a material way (substantive modifications, structural changes etc.) that affect its meaning, style or intent”.9

Thirdly, manipulation covers AI alteration of your own human-written draft, not just AI drafting from scratch. So writing a draft newsletter yourself first and then manipulating it with AI is still in scope, as long as the AI use goes beyond standard editing.10

So if you use AI to help with research for your newsletter, this is not caught by the transparency obligations. This is also the case if you only use AI to:11

Generate outputs that merely reproduce, select, arrange or rank existing content
Conduct internal analytical processes that “extract and structure data, but do not summarise it”
Generate short sequences of “numbers, symbols or letters (e.g. single words, image captions, alt-text, UI labels)”

Published

The Commission guidelines clarify that “published” means the text is accessible by “an indeterminate, fairly large number of unrelated, potential readers simultaneously and/or successively, whether or not against payment (e.g. subscriptions)”.12

This means that a paywall does not impact the public element. Only genuinely closed groups like private correspondence and internal corporate networks escape the published condition.13 The Commission does not provide specific quantities in terms of how many subscribers one would need to have for their work to be considered published when a newsletter is sent out, but newsletters with more than a handful of subscribers probably count here.

Has the purpose of informing the public

The Commission guidelines state that the text must “intend to communicate knowledge, opinions or facts”.14 This excludes “short texts which do not materially communicate knowledge, opinions or facts”.

This is a pretty low bar. Even if your pieces are more commentary than fully-polished reports, they are caught here unless the text is less than the typical length of a newsletter.

Is on matters of public interest

The Commission guidelines states that public interest matters include matters “relevant to society at large, whether at a local, national, Union or international level, and meriting public debate or scrutiny”.15 This includes:

Politics and democratic processes, public administration and services,
The administration of justice and law enforcement
The protection of fundamental rights, public security, public health, environmental protection, consumer safety
Any economic, financial, political, scientific, or cultural development that may be relevant subject of public debate

That final category is a very broad catch-all that would cover a lot of different newsletters on Substack. One example given in the guidelines is “AI-manipulated parts of a lifestyle-website article on diets and disease.” But an example that would not be a matter of public interest is “AI-generated fantasy novels.”16

Is not subject to human review or editorial control and editorial responsibility

Article 50(4) states that transparency obligations do not apply “where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.”

The Commission guidelines clarify that there are two separate conditions here: (i) human review or editorial control and (ii) editorial responsibility.

Human review or editorial control
Under the Commission guidelines, “human review” means “the deliberate examination of the substance of the content by one or more natural persons possessing relevant knowledge and professional judgement pertaining to the subject matter”, with “[f]act-checking the accuracy of the content” as “a minimum requirement”.17 Examples include academic peer review and professional validation chains.

“Editorial control” means control by “a responsible editorial entity (e.g. an editor-in-chief)… having the authority to approve, alter or reject the substance of the text based on substantive grounds”.18

The guidelines clarify that human review or editorial control does not include “[s]uperficial, solely formal or procedural checks (e.g. spell-checking or grammatical correction), the mere existence of an editorial policy, automated review processes or cursory editorial approval without substantive engagement”.19

Also, if AI is “used to modify, supplement, or reformulate content following editorial sign-off, the resulting content must be treated as AI-generated or manipulated.” This means that any “substantive AI intervention occurring after the human review or editorial control process has taken place will therefore cause the exception to become void.” An example includes an AI-generated article reviewed by another AI where “a human editor performs a mere superficial, grammatical check before publication”.

So if you run a finished draft through Claude and then publish it right after with no further checks, then this would not constitute exercising human review or editorial control.

Editorial responsibility
A person must “hold the ultimate legal responsibility over the publication of the content, including the human review or editorial control”.20 It can be an individual, an editorial board or a publishing company. And the identify and contact details of this person “should be made publicly available on an easily findable location”, for example a site’s terms and conditions or other user-facing legal information.

For Substack writers, your real name will need to be visible somewhere on your profile or newsletter and it is possible that the ability for users to DM you on the platform may qualifies as a means of contact though this is not clear.

Okay so what should I do?

There are basically two routes depending on your approach to AI use. One route focuses on securing the editorial exception, while the other focuses on AI newsletters you want to publish without review.

However, regardless of the route you choose, there are two things you should know.

Firstly, AI-generated text published before 2 August 2026 does not need any retroactive labelling. If you have drafts that have been AI generated or manipulated before this date that are then published on or after that date, then they will need to be labelled (if you meet all the aforementioned conditions).21 So check anything that might be sitting in your drafts.

Secondly, fines for non-compliance can be up to €15,000,000 or 3% of worldwide turnover, whichever is higher, but for SMEs including start-ups it is whichever is lower and regulators must weigh proportionality and economic viability when issuing fines.22

But the risk of a fine for the average Substacker is quite low realistically. Enforcement sits with national market surveillance authorities, who can act “on their on initiative or following a complaint, which every affected person or any other natural or legal person having grounds to consider such violations has the right to lodge”.23 So users or subscribers could complain about a newsletter, but authorities will likely provide chances to fix any compliance issues long before fines are considered.

Route 1 - Securing the editorial exception

This would remove the need to label your content as AI:

Publish your editorial responsibility statement. Make sure your About page has your name, contact details (perhaps enable users to DM you) and specify that you hold editorial responsibility for everything you publish.
Do substantive reviews and fact-check your work. Do not just check for spelling mistakes. You need to do a deliberate and substantive examination of your AI text before you publish it.
Do not run your draft through AI just before you publish. Do this before you do your final, human-only review.
Document how you use AI for your writing. This is good practice as well as something that is technically required under the Commission’s guidelines.24

Route 2 - publishing without review

This is if you want to label your content as AI:

Disclose that the text has been artificially generated or manipulated, ideally at the start of the newsletter. This cannot just exist in your About page only.
If your pieces are longer, it would be better to mark the specific sections where AI has been used.

Do either the watermarking or detection tools from AI providers change anything?

Neither the watermarking being introduced by the labs nor the Pangram tool on Substack change your responsibilities as a Substack writer under the AI Act.

The watermarking helps the labs comply with their obligations under Act as AI providers, but does not change what you need to do. Also, relying on the editorial exception (i.e., Route 1 above) does not depend on whether AI can be detected in your text, as the exception applies even if the text is obviously AI-generated.

Under Article 3(66) of the AI Act, systems like Claude and ChatGPT fall in the category of ‘general-purpose AI system’, which is a type of AI system that the Act regulates.

Article 2(10).

European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act (20 July 2026), para. 19 and Guidelines on prohibited artificial intelligence practices (29 July 2025), para. 34.

European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act (20 July 2026), para. 19.

Article 2(1)(b).

European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act (20 July 2026), para. 58.

European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act (20 July 2026), para. 59.

European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act (20 July 2026), paras. 90-92.

European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act (20 July 2026), para. 59.

European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act (20 July 2026), paras. 64-68.

European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act (20 July 2026), para. 131(i).

European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act (20 July 2026), para. 131(ii).

European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act (20 July 2026), para. 131(iii).