I use AI in my writing. Substack grades it. LinkedIn reduces its reach.
What the platforms measure now decides how you write.
TLDR: Substack has switched on AI detection and lets readers scan any post for an estimate of AI involvement. LinkedIn labels AI images with a Content Credentials badge but reduces the reach of AI-flagged text instead of labelling it. Both platforms measure the surface yet neither measures the judgement behind the work.
Here is my process, in the open, and why disclosure beats suppression.
I am Melanie Goodman, a lawyer turned LinkedIn consultant and the writer of The Link Tank. I teach time-poor professionals to use LinkedIn and Substack intelligently, without hacks or hype. If a platform announcement has changed how you feel about your own writing this month, you are in the right place.
Before I ran a consultancy, I practised law.
My first weeks as a junior associate at Allen & Overy taught me a lesson I have never shaken off. I sent a partner my first attempt at a drafting exercise and it came back with more red ink than black. Nobody apologised for that. Nobody expected a first draft to be the finished article. The mark-up was the method.
The habit that stayed with me was the blackline. Every revised document went out with a comparison version showing exactly what had changed since the last draft. The other side did not have to trust us. They could see the edits for themselves, line by line.
Disclosure was both professional courtesy and protection.
I spent sixteen years working across borders after that, and the rule held everywhere. You did not hide the changes. You showed your working.
Yes, this essay is about writing with AI.
It is also about the strangest fact of my working life: a lawyer, trained to trust nothing unverified, now writes with a tool famous for inventing things.
I use Artificial Intelligence in my research and my drafting. I have said so before and I will keep saying so. What I want to examine this week is what the two platforms I work on have decided to do about it. They have chosen opposite designs, and the difference matters more than either announcement admits.
How I actually use AI
My process has a fixed order. I decide the angle. I am interviewed about scope before a word is drafted, because structural recommendations made before questions are usually wrong. A draft follows. I reject the parts that do not sound like me or do not say what I think. Statistics are checked against a named source before they go in, never after. Claims that cannot be verified are cut or bracketed for me to resolve.
The tool does the volume; I make the judgement. On a good week the draft survives two rounds of my corrections. On a bad week it survives five.
An example from this month: While drafting my anniversary issue, a scene appeared in the draft that I did not recognise, a conversation with a wealth adviser in Geneva that had never happened in the form described. The machine had written something plausible. Plausible was the problem. The scene was bracketed, checked and rewritten before a reader ever saw it and the rule it produced now applies to everything I publish: anything invented gets flagged for me to verify or replace. My old profession would call that a mark-up. I call it the job.
What Substack decided to show
I read the Substack announcement on my phone last Tuesday evening, somewhere between finishing client work and being asked what was for dinner. My first feeling was recognition rather than worry. A platform had reached for the thing my old profession has always relied on: showing the reader what changed. Then I kept reading, thought about LinkedIn, and realised the two platforms I work on had just answered the same question in opposite ways.
On 21 July, Substack switched on AI detection through an integration with Pangram. Readers on the web and the iOS app can scan any post, note, reply or comment over 100 words and see an estimate of how much was human-written, AI-assisted or AI-generated (Axios, July 2026). Writers can scan their own drafts before publishing, report scans they believe are wrong, and add a statement called How I Make This that readers see alongside the work. Substack’s chief executive, Chris Best, framed the goal as stopping Substack becoming “like LinkedIn”.
LinkedIn faced the same problem two months earlier and chose a different answer, which is presumably why Best reached for the comparison.
Give Substack this much: the design is honest about its own uncertainty. The estimate sits in the open. The writer gets a right of reply beside it and a correction route when the scan is wrong. Readers decide what it all means.
What LinkedIn decided to do instead
In May 2026, LinkedIn announced new detection systems, trained with its editorial team, aimed at posts that appear machine-generated and carry no clear point of view (LinkedIn Editorial blog, May 2026, reported by The Decoder). Laura Lorenzetti, Vice President and Executive Editor of LinkedIn Global Editorial, set the standard plainly: “Your posts and comments need to represent your voice and your perspectives.”
The consequence is different from Substack’s.
Flagged content is not labelled but it is throttled. Posts judged generic lose distribution and stay largely within the author’s own network. LinkedIn reported a 94 per cent accuracy rate on generic content in early tests. That figure is self-reported, and the false positive rate has not been published, so treat it as a claim rather than a finding.
Images are treated differently: Since May 2024, LinkedIn has read Content Credentials, the metadata standard from the Coalition for Content Provenance and Authenticity (C2PA), and displays a small CR badge on images that carry it.
Clicking the badge opens a panel naming the tool that made the image, the date, and whether AI was involved. Two limits matter:
The system reads metadata rather than detecting anything, so a file with the metadata stripped out arrives unlabelled.
It also fires on a real photograph that has been through a single AI edit, which means a retouched headshot can carry the same badge as a fully generated scene.
The scale of the problem is significant.
A Pangram Labs study of more than a million social media posts, reported in July 2026, found that over 41 per cent of long-form LinkedIn content was fully machine-generated, the highest share of any platform studied.
The irony is also real. Microsoft, LinkedIn’s parent company, spent the same period promoting Copilot features that draft LinkedIn posts for you.
One hand fills the feed. The other hand filters it.
How to manage this on LinkedIn today
1. Check what your images disclose before you post. AI tools from OpenAI, Adobe and Microsoft embed Content Credentials into the file at the moment of creation. Run the file through the C2PA Verify tool first, so you see what your reader will see.
2. Expect the badge on edited photographs. A single generative edit in Photoshop is enough to attach credentials to a real photo, so a retouched headshot can carry the same CR badge as a fully generated image. Decide whether that suits the post before you upload.
3. Know that there is no setting to remove the badge. LinkedIn reads the file, not your preferences. The only control point is the file itself, before it is uploaded.
4. Verify your profile. LinkedIn’s filtering approach favours verified members, and identity verification is free within the app under your profile settings.
5. Write text with a stated point of view. The detection systems introduced in May 2026 target posts that read as generic and machine-made. Context, expertise and a clear position are the qualities LinkedIn says its systems are built to protect.
6. Disclose on your own terms. LinkedIn has no equivalent of Substack’s How I make this statement, so add your own line when AI helped with a piece. The writers who disclose first will not be the ones explaining themselves later.
Two designs, one wrong measurement
Here is the difference that is of consequence
Here is the difference that matters.
Substack tells the reader and lets the reader decide.
LinkedIn splits the answer by format.
Your image gets a public label. Your words get a private penalty. Disclosure for pictures, suppression for prose, and no explanation for why the reader deserves the truth about one and not the other.
There is a further inversion worth noticing:
Substack lets a writer disable detection on their own work, and readers can draw their own conclusions from that choice.
LinkedIn’s image badge has no off switch at all. The platform that grades you trusts you with a setting. The platform that throttles you does not.
As a lawyer, I notice that this distinction is about to stop being philosophical.
Article 50 of the European Union’s AI Act is scheduled to take effect in August 2026. It requires disclosure to EU users when content has been machine-generated, with penalties of up to fifteen million euros or three per cent of worldwide annual turnover. Reducing the reach of a post is a moderation choice. Telling the reader what they are reading is a transparency duty.
The first does not obviously satisfy the second. A proposal to delay parts of the Act was tabled in late 2025, so watch the implementation date rather than assume it.
Both platforms share one weakness: Detection reads the finished language.
It cannot read the rejected drafts or the sentence that was cut because it could not be verified. A percentage describes the surface of a text. Whether a person stood behind the words is a different question and no detector answers it.
Interfaces create norms through what they measure and what they display. Substack’s norm will be disclosure with context. LinkedIn’s norm will be caution, because when the penalty for detection is invisible reach, writers will not experiment with the tools in the open. They will hide the use or stop the writing but neither outcome makes the feed more honest.
The image labels teach a different lesson. Because the badge reads metadata rather than the picture itself, the label is optional for anyone willing to strip a file before uploading. A rule that only binds the honest is a strange foundation for trust.
How I make this
I research with AI and I draft the first draft with it and this essay is its own example. AI gathered the platform coverage and produced the drafts. I set the argument, directed the corrections through several rounds and rewrote until the piece said what I think in the way I say it. When Substack’s scanner reads it, it will find the machine’s fingerprints on the sentences and mine on the personal experiences. Both readings are accurate. Consider this the blackline for the document you have just read.
Law worked out the answer to this decades ago. The blackline did not ask anyone to trust the drafter; it showed the changes and let the reader judge them. A percentage beside my writing does not trouble me for the same reason.
Scrutiny in the open improves the thing being scrutinised. Scrutiny in the dark teaches people to hide.
This essay was made the way I describe above. Substack can grade it. LinkedIn can decide what happens to the post that links to it. You can decide whether it earned your attention. Of those three judgements, only the last one was ever the point.
Would you keep using AI in your writing if every reader could see a percentage beside your name? I read every comment.
New here? Start with these:
The irony of platforms penalizing the surface of text while ignoring the human judgment and editorial standards behind it proves we need transparent disclosure models far more than quiet algorithm suppression.
Great post Melanie. Transparency is always the best way to approach anything.
On this part you said about LinkedIn, "posts that appear machine-generated and carry no clear point of view".
It's the second part about writing content that has no clear point of view that I wanted to comment on.
There's already a lot of things that people hide in their writing besides AI.
People don't discuss politics because I don't want to alienate half there audience.
People rarely truly disclose what their expertise or experience is and make claims that are embellishments at best.
Someone who uses sourced content as the basis of the article that they're writing don't always give any attribution to whatever they read that influence them.
And many of them write to try and please everyone or at least as many people as possible. They want to provide information that appeals to a wider audience.
You know the old saying. If you write for everyone, you write for no one.
That's where that last part comes in "and carry no clear point of view".
We live in an age where search is AI powered. Just writing to provide information without a clear opinion or point of view means that your content is going to get copied and regurgitated by AI without sending you any traffic and maybe even without giving you any attribution.
And people absorb information but they don't always remember where they got the information.
They do remember when you have a strong opinion about something and you don't mind sharing that opinion. They can agree or disagree but it's much more engaging when you actually say what you think.
Everybody keeps trying to "polish" their writing because they want it to look "professional". But in many cases that also means dry, informative, unopinionated.
AI can already do that. What I like to ask writers is *What is it about your content that someone can't read almost anywhere on the web or have AI generate?*
One of the first things is your opinion, your point of view, how you arrived at that point of view, the story behind it, how you experienced it personally, how you used that information to deliver an outcome.
I think it's ironic that it's machines that are going to end up forcing humans to write like humans.