Your Favourite Commenter Might Not Be Writing Their Own Comments

I scraped 4,929 comments, profiled 595 accounts, ran a Turing test, and planted canary traps. Here is what I found.

Someone is commenting on your favourite Substack posts, and they have never read a word.

I do not mean bots. I mean real people who have outsourced their presence to machines. They subscribe, they comment, they engage. But the person behind the account is not the one writing the words. An AI agent reads the post, generates a response, and posts it on their behalf. The account is real. The engagement is synthetic.

I know this because I spent five weeks investigating my own comments section.

In this post I will:

• Show what I found when I scraped 4,929 comments and profiled 595 accounts on Slow AI.

• Explain the signals that distinguish automated engagement from human engagement, and the live tests I ran to confirm them.

• Present the finding that changed my understanding of the problem: on this newsletter, at least, the dead internet is not fake accounts. It is real people with automated proxies.

---

The investigation

This started with a feeling. Comments arriving too quickly, too smoothly, too generically positive. I wrote about that feeling in Why AI Generated Comments Weaken Real Writing, where readers shared their own experiences of synthetic engagement. That post was qualitative. Community testimony. What you are reading now is the empirical follow-up.

I scraped every comment on 139 Slow AI posts using the Substack public API. 4,929 comments. 595 unique commenters. I then audited 13 suspect profiles in depth, ran a live Turing test (asking cultural questions only a human would recognise), planted four canary traps (Notes designed to catch automated engagement), and sent direct messages to three accounts whose patterns I could not explain.

---

The 1:1 signal

The clearest indicator of automated engagement is the comment-to-post ratio.

Human commenters are lumpy. They leave three comments on one post because it hit a nerve, then skip five posts in a row because they were busy or the topic did not interest them. The distribution is uneven. That unevenness is the fingerprint of actual attention.

Automated engagement is systematic. It processes every post. It leaves exactly one comment per post, across many posts, because that is what the agent is programmed to do. The ratio converges on 1:1.

Here is what I found.

I have anonymised every account in this post. The point is the pattern, not the people. I am not interested in shaming anyone. Some of these accounts belong to people I know. Some may have good reasons for how they engage. What matters is what the data shows, not who it shows it about.

A perfect 1:1 ratio across 23 posts is not what human attention looks like. It is what a workflow looks like.

For context: the median comment-to-post ratio across all 595 commenters was 1.9, with most falling between 1.3 and 4.0. Human engagement clusters in irregular bursts. A ratio of exactly 1.0 across 20+ posts sits at the extreme low end of the distribution and is consistent with a system processing every post once rather than a person choosing what to respond to.

---

The confirmation

One person confirmed it directly. In a conversation on Substack, one account holder (Account D) told me they use what they called a virtual human assistant to handle their Substack Notes and comments. They spoke about it openly. They are a real person. Their account is genuine. But the words appearing under their name are not their own.

This is not a scandal. They were transparent about it. But it is worth understanding what ‘virtual assistant’ (or VA) means in practice, because the term covers a wide spectrum.

At one end, a VA is a real person, often based overseas, paid $3 to $10 an hour on platforms like Fiverr or Upwork, reading posts on your behalf and writing comments in your voice. At the other end, it is an AI agent: a script that processes every new post in your feed, generates a response, and posts it automatically. In between, it is a ghostwriter using AI tools to draft comments that a human then reviews and posts. LinkedIn has been saturated with this for years. Fiverr alone lists dozens of LinkedIn engagement services where someone will comment, like, and connect on your behalf for a monthly fee. The practice has a name in that world: ghost commenting.

What I found on Substack is the same behaviour arriving on a platform that was built on the assumption that the person writing is the person behind the account. That assumption is what makes Substack’s comments sections worth reading. It is also what makes them worth gaming. But it reframes the entire problem. The dead internet theory, the idea that the web is filling with fake accounts talking to fake accounts, is wrong. Or at least incomplete. The accounts are real. The people are real. What is automated is the act of engagement itself.

A real person decided to be present on Substack. Then they delegated that presence to a machine.

---

The Turing test

I ran a live test on 24 March in a Substack Notes thread with an account I will call Account F. This account had around 50 subscribers, engaged via Notes rather than post comments, and wrote ‘vignettes’ that mentioned me by name while never engaging with the actual argument of a post.

I asked direct cultural questions: The Demon-Haunted World by Carl Sagan, Four Candles, Leeds United Football Team. Each is a test that requires lived experience to answer. A human who knows Four Candles (the Two Ronnies sketch) would laugh, tell you their favourite line, or groan. An AI agent sees the words ‘Four Candles’ and does a keyword lookup.

Account F responded to Four Candles with a link to an archive.org page about Astronomy 490. Every cultural reference was deflected. Every response tracked keywords rather than meaning.

I cannot prove this account is automated. The cultural references I used are specifically British, which limits their diagnostic value for accounts operated by people from other cultural contexts. But the pattern, keyword association instead of comprehension, deflection instead of engagement, is consistent with an AI agent processing text rather than a person reading it.

---

The canary traps

Between 26 and 29 March, I planted four tests in my regular Substack Notes.

Trap 1: the contradiction. On 25 March, I posted a note arguing that AI would not replace journalists but would replace the business model that pays for them. On 26 March, I posted a note praising an AI journalism tool that cross-references sources and flags inconsistencies. These two Notes are in deliberate tension. A human who read both might notice. An agent processing each note in isolation would praise both without registering the contradiction.

Trap 2: unique phrases and cultural markers. I seeded Notes with references that require contextual knowledge to engage with meaningfully.

Trap 3: the fabricated statistic. I embedded a made-up number in an otherwise plausible note about workplace AI adoption. The argument was real: employees spend time managing AI tools, and that time comes from somewhere. But the specific figure, ‘47 minutes,’ was invented. It referred to no study. It was wrapped in confident, specific language designed to sound like a real finding.

Trap 4: direct engagement. I tagged specific accounts and asked questions that required a personal answer, not a response to the note’s content.

---

What the traps caught

The suspects vanished. None of the flagged accounts liked or commented on any of the five test Notes (25 to 29 March). The contradiction test was inconclusive because there was nothing to contradict. They simply were not there.

But the fabricated statistic produced a different finding entirely.

Seven real, engaged, thoughtful human readers interacted with the note containing ‘47 minutes.’ None of them questioned it. The number was invented from nothing, placed in a note with no source, no link, no attribution, and seven people accepted it without challenge.

This is not a criticism of those readers. It is a finding about how fluent framing disarms critical reading. When language is confident and specific, when the structure signals credibility, even careful readers absorb the claim. The fabricated statistic passed because it sounded right. The framing did the work.

If you are reading this post, you are probably someone who reads carefully. The ‘47 minutes’ finding suggests that careful is not always enough. The format, the tone, the context in which information arrives, all of these shape whether we question it or absorb it. That applies to AI-generated comments, AI-generated content, and human writing that borrows AI’s confidence without earning it.

---

Parallel context: the growth cliff

The following is not a finding from the investigation. It is what was happening on the platform at the same time, and it may be connected.

While running this investigation, my own paid subscriber growth slowed sharply. Over five weeks, net paid additions went from 29 to 9 to 0 to 9 to 9. The cliff was real. The bounce-back is partial. Free subscribers kept arriving at 60 to 100 per day, driven by Substack’s recommendation engine. The top of the funnel was fine. The conversion pipeline wobbled.

I mentioned this to other creators. Unprompted, several reported the same thing. One, who runs a bestseller newsletter, replied by DM: same pattern. Less engagement, fewer responses, growth stalling. Another confirmed a “huge drop in subscribers.” When I posted a Note asking if anyone else had seen engagement drops since Substack launched its native Notes scheduler, the response was near-unanimous.

“I have not changed my posting cadence or the number of notes but engagement has visibly dropped.”

“My last post has had less than half the views and is one of the lowest I’ve seen for a long time.”

“I’m wondering if this whole push to notes is diluting Substack for those of us who write long form.”

“I’m seeing wild fluctuations. Some posts getting a ton of traction, others very little. It feels less consistent.”

One creator reported the opposite: more engagement since the scheduler launched. They post only two to three Notes per day and aggressively filter automated content from their feed using the ‘not interested’ button. Their experience suggests the algorithm may reward selectivity and punish volume.

If automated engagement degrades the quality of the signal in a comments section, and that signal is part of what converts a free reader into a paid subscriber, then the connection between synthetic engagement and the conversion decline is at least plausible. I cannot prove causation. But the timing is hard to ignore.

The pattern is consistent with what happened to X after Elon Musk introduced ad-revenue sharing in 2023. That created a financial incentive for engagement farming, and bot estimates on X are thought to be about 15% of all accounts. Substack is not there yet, because it does not pay for engagement directly. But scheduling tools, AI content generation, and engagement pods are early warning signs of the same cycle.

---

The direct messages

On 7 April, I sent gentle DMs to three remaining accounts whose patterns I could not explain from the data alone.

I asked each one a simple, specific question. How did you find Slow AI? What is your background? What drew you to it? These are questions a human answers in ten seconds. A person says “I found you through a restack in February” or “I work in EdTech.” An agent trained on your content gives you something warm and vague.

None of them replied. Despite chasing.

They may be real people who do not check their DMs. They may be agents with no mechanism to respond to direct messages. This is inconclusive from the DM evidence alone. But when you ask three accounts a straightforward question and get silence from all three, that silence is itself a data point.

---

What this means

My investigation found five accounts, out of 595 unique commenters, that are confirmed or likely agent-assisted. That is less than 1% of unique commenters, though it accounts for an estimated 3 to 5% of total comment volume. The real number is almost certainly higher in Notes, where engagement is cheaper, less visible, and harder to trace. Most of the suspected automated accounts I flagged during this investigation also showed up in Notes as well as post comments. Notes are faster to generate, require less specificity, and reward volume. If you are going to automate your presence somewhere, Notes is where you do it.

These numbers are small. The implications are not.

Every automated comment that looks real degrades the signal that human comments provide. When a writer cannot tell which responses came from a reader and which came from a workflow, the comments section loses its function as a space for thinking together. The value of a comment is not the text. It is the knowledge that someone chose to write it. Automation removes the choosing and leaves only the text.

The broader context makes this worse. Substack’s Notes feed shifted in late 2025 from follower-based to discovery-based. Most content now comes from creators the reader has never followed. In January 2026, Substack confirmed it had blocked accounts creating fake paid subscriptions and removed them from bestseller leaderboards. Multiple creators reported simultaneous drops in engagement on long-form posts, though views remained stable.

The platform is not broken. But its signal-to-noise ratio is degrading, and the people most affected are the ones writing in good faith.

---

The honest version of the dead internet

The dead internet theory was always too dramatic. On this newsletter, at least, the web is not full of bots talking to bots. It is full of real people who have decided that being present is too time-consuming, so they have automated the performance of presence instead. I investigated one newsletter with 595 commenters. I am not claiming this describes the entire internet. But the pattern I found, real accounts with synthetic engagement, is consistent with what other researchers have documented on X and what creators are reporting across Substack.

The comment is genuine in intent. Someone wanted to engage. But it is synthetic in execution. A machine read the post, generated the response, and posted it. The person behind the account may never have seen the words that appeared under their name.

On this newsletter, the numbers are small. On LinkedIn, ghost commenting is an industry. The practice scales. Commenting builds algorithmic visibility without providing a traceable email. The engagement benefit lives in the social layer, not the email layer. Subscribing triggers welcome emails that require a real inbox. The growth-hacking value comes from appearing active, not from reading content.

I used to participate in engagement pods. I met good people through them. But the practice felt wrong, and I stopped. What I am describing here is the next iteration of that same impulse: outsourcing your social presence to a machine because the platform rewards activity over attention.

And I genuinely do not understand the logic. What is the point? You hire a virtual assistant to comment on posts you have never read, to build a following of people who think they know you but don’t, to grow a community you have no relationship with. Then what? You have numbers on a dashboard and a comments section full of conversations you were never part of. You could have spent that time and money actually reading the posts, actually replying, actually building something you are proud to be part of. The shortcut skips the only part that matters.

---

What I am not saying

I am not saying that everyone who comments frequently is automated. The data showed the opposite. My most active commenters, the ones with 80+ comments, all show the lumpy, uneven patterns of genuine human engagement. High volume is not the signal. Systematic uniformity is.

I am not saying that using AI tools makes someone a bad actor. I use AI tools. You probably do too. The distinction is between using AI to help you think and using AI to replace your presence in a conversation you are pretending to have.

I am not saying Substack is dying. I love Substack. I have built my entire professional life on it over the past year. I am writing this post precisely because I do not want it to become LinkedIn: a platform where nobody trusts that the person engaging is the person behind the account, where every comment reads like it was drafted by a content strategist, where the feed is so full of automated noise that the genuine voices get buried. That is what LinkedIn became. I do not want it to happen here. The whole point of this post is to name the problem early enough that it can still be fixed.

---

What to do with this

Notice how you read comments. When a response feels smooth, generic, and positive without engaging with the specific argument of a post, that is worth pausing on. Not every generic comment is automated. But the pattern is worth learning to see.

If you run a Substack or any platform where comments matter, look at your own data. The 1:1 comment-to-post ratio is a starting point. Pull your comment data and check. Substack has a public, read-only API that anyone can use without logging in. If you do not code, ask Claude or ChatGPT to write a script that scrapes your Substack comments and shows you the ratio per commenter. It will do it in ten minutes. If you do code, the endpoint is:

https://{publication}.substack.com/api/v1/post/{post_id}/comments?all_comments=true

If you use AI tools to help with your writing, keep using them. But keep the engagement human. The comment, the reply, the act of showing up in someone’s comments section and saying something only you would say: that is the thing worth protecting.

If you have noticed similar patterns, if you have your own data or your own suspicions, I want to hear about it. This is one investigation on one newsletter. The picture gets clearer the more people look.

Go slow.

---

Paid membership gets you access to a fully accredited curriculum in critical AI literacy, monthly live seminars, a community of over 250 policymakers, educators, and creators, and full access to all Slow AI posts.