AI tools helped gather research and write the first draft of this article. I edited it and checked every policy, figure and quotation against the original paper, the platform's own page or a news report that quotes the platform, listed at the end. The cover illustration is AI-generated. Details at the end.
A creator programme starts with a sensible idea: reward authors whose work reaches people. Reach has to be measured, so the programme picks numbers: views, followers, likes, comments, visits from shared links. Then the numbers start to pay. From that day on, some participants will find it cheaper to produce the number than to produce the work.
This is not peculiar to creators; it happens to any measure that carries a reward. The useful question is how to notice it early and how to design numbers that are harder to fake.
This is the eighth part of Influence & Incentives, a series on how other people, rules and rewards shape what we decide, write and publish. The sixth part looked at what to count about an author beyond output. The seventh looks at how rewards change what gets published. This part stays with one narrow problem: how to tell when the numbers themselves are being farmed.
Why it always happens
In 1997 the anthropologist Marilyn Strathern, writing about audits in British universities, put it in one sentence: "When a measure becomes a target, it ceases to be a good measure" [1]. Her example was exam grades: the more a particular grade becomes an expectation, the poorer it becomes at telling individual performances apart. She noted that Hoskin, whose work she was drawing on, calls this "Goodhart's law", after the economist Charles Goodhart's observation about instruments for monetary control [1].
A creator metric follows the same path. Before it pays, a view count is a fairly honest side effect of people choosing to watch. After it pays, it is also something people can manufacture. The number keeps rising, but it describes less and less of what it was meant to describe.
Large platforms treat this as a permanent problem, not an occasional one. YouTube's fake engagement policy says the platform "doesn't allow anything that artificially increases the number of views, likes, comments, or other metrics", whether by automatic systems or by serving videos to unsuspecting viewers [2]. The policy holds creators responsible for methods used by "someone you've hired", not only by themselves [2]. In November 2018 Instagram announced it would begin "removing inauthentic likes, follows and comments from accounts that use third-party apps" to boost their popularity [3]. In May 2026 Twitch's chief executive, Dan Clancy, wrote that "effectively combatting viewbotting is challenging", as Tubefilter reported [4].
Three ways numbers get gamed
Most metric gaming falls into three families. They leave different traces, so it helps to keep them apart.
Buying. Paying for views, followers, likes or comments that come from automated accounts, scripts or farms of low-paid clicks. This is the crudest form and the easiest to detect, because the bought activity rarely does anything else: it arrives, registers, and leaves. Scripted visits also tend to stop just after the counting line.
Trading. Real people agreeing to inflate each other's numbers. Engagement groups, mutual-follow circles and chats where members click and comment on each other's links every day all work this way. Each individual action is genuine; the pattern is not. Trading is harder to catch because the accounts are human and sometimes the comments are, too.
Shaping. Producing work aimed at the rule rather than at readers, without anything fake in it. A giveaway post that asks everyone to reply. A title written to collect clicks the text does not repay. One article split into several short parts so each one counts separately. Shaping sits on the border between gaming and ordinary optimisation, which is why it causes the most arguments.
When a metric pays directly, services appear to produce it: the Steemit study discussed in part 07 also found a supply network behind bot curators [5].
Seven signs that a number is being farmed
None of the signs below proves anything alone. Each one says "look closer". Together, they often separate a lucky post from a manufactured one.
1. Proportions that do not fit
Real audiences behave in roughly consistent ratios: some followers see a post, some viewers like it, fewer reply. Bought activity usually inflates one number and leaves the others behind. A post with a large view count and almost no likes or replies, or an account with many followers whose posts reach only a sliver of them, is worth a second look. The exact "normal" ratio differs by platform, format and topic, so compare an author with their own history and with similar authors, not with a universal threshold.
2. A shape over time that has no cause
Organic growth is uneven, but its jumps usually have a reason: a post that was shared widely, a mention by a larger account, a topic in the news. A sudden step up in followers on a day when nothing was published, followed by a slow, steady decline as the platform removes fake accounts, is a pattern often associated with purchased followers. Ask what happened on the day of the jump.
3. Activity bunched just past the counting line
Every programme rule has a threshold: a visit counts after a certain number of seconds, a post qualifies after a certain number of distinct visitors. Real behaviour spreads out around such lines. Scripted behaviour stops right after them. If an unusually large share of an article's visits end a moment after the minimum needed to count, the visits were probably built for the rule. This is why programmes should keep the distribution of visit lengths, not only the count of visits that passed.
4. The same small crowd every time
Trading leaves a social fingerprint. If the same few dozen accounts are always the first to like, reply or click, across every post by an author, the audience may be a circle rather than a readership. A practical check is overlap: what share of the people who engaged with one post also engaged with the author's previous posts? Some overlap is healthy loyalty. Nearly total overlap, with few newcomers, suggests a closed group.
5. Replies that say nothing
Comments from engagement groups and bot farms tend to be short, generic and interchangeable: a word of praise, an emoji, a phrase that would fit any post. Real replies refer to something specific in the text, ask a question or disagree. If most replies under a post could be pasted under any other post, and the author never answers them, the comment count is not measuring conversation.
6. An average far above the typical post
One giveaway, contest or controversy can produce a single post with many times the usual replies or shares. An average over the author's posts then looks impressive, while the typical post is ordinary. Comparing the average with the median, the middle value, shows this at once. A programme that uses medians instead of averages removes much of the incentive to stage one inflated post.
7. Nothing happens downstream
Real readers do things after reading: open a second article, subscribe, return the next week, link to the piece, send a correction. Farmed visits stop at the counter. If a post's external traffic produces no further reading, no returns and no later references, the traffic may have been delivered rather than earned. It is probably the strongest sign, because downstream behaviour is the most expensive part to fake.
Designing numbers that are harder to farm
Detection after the fact is expensive. It is cheaper to choose numbers that resist gaming from the start. A few design choices recur:
- Count active attention, not presence. When Medium moved writer pay mainly to reading time in 2019, it said it would count active reading time, so if a reader leaves the window open and steps away, "that shouldn't count", according to TechCrunch [6]. The sixth part used the same rule to ask what to count; here it matters because each such condition raises the cost of faking a visit.
- Use medians and windows. A median over the last 30 or 90 days rewards consistent work and blunts single staged posts.
- Limit how often the same reader counts for the same author. If one person, device or account can confirm an author's work only once in a set period, a closed circle of fifty friends cannot carry an author for long.
- Pair every reward number with a downstream one. Visits next to further reading, followers next to reach, comments next to author replies. Gaming one number is cheap; gaming two that must agree is not.
- Delay rewards. A holding period before rewards can be withdrawn gives time for removed fake accounts and reversed activity to show up before the money is gone.
- Penalise the inflation, not only the person. Twitch's May 2026 measure is an example: for channels identified as persistently viewbotting, Clancy wrote that Twitch would cap concurrent viewership for a fixed period, with the cap "based upon historical data regarding that creator's non-viewbotted traffic" [4]. The penalty takes away exactly what the bots were bought to produce.
A sign is not a verdict
The same signs that catch cheaters can wrong honest authors. A post can go viral in one community and collect a closed-looking crowd. A competitor can buy bots for someone else's account to get it penalised. A short, warm reply can come from a real reader with nothing more to say. YouTube's policy holds creators responsible for methods used by someone they hired [2]; from the outside, a programme usually cannot see who paid for the activity at all: the author, a promoter or a rival.
Three habits keep detection fair:
- Treat signs as reasons to review, not as automatic penalties. Hold a reward and ask a person to look; do not ban on a ratio.
- Give authors a way to explain and appeal. Many sudden spikes have innocent causes that only the author knows.
- Describe evidence, not character. "Most visits on this post ended right after the counting threshold" is a fact that can be checked. Calling someone a fraud in public is an accusation that can rarely be proven, and it carries legal risk.
A worked example (fictional)
The programme, the authors and all figures below are invented to show the signs in use.
A small publication pays authors when an article brings in a set number of visits from links they share. At the end of a month, two articles have each crossed the line with about the same count.
Article A's visits arrived over ten days, with lengths spread from a few seconds to many minutes. A fifth of the visitors opened a second article. Two readers replied, one with a correction. The visitors had little overlap with the author's previous articles.
Article B's visits arrived in two short bursts. Most of them ended a few seconds after the minimum that counts. Almost no visitor opened another page. Nearly all the accounts that commented had also commented on the author's last four articles, each with a few words of praise.
On the paid number, A and B are equal. On the signs, B shows three at once: bunching at the threshold, the same crowd, and nothing downstream. A fair programme would not accuse B's author. It would hold B's reward, ask the author how the article was shared, and look at the traffic with a person rather than a formula.
Five questions to ask before a metric starts paying
- What would it cost to fake this number? If the answer is "less than the reward", it will be faked.
- Which second number must agree with it? Name the downstream signal that a fake would not produce.
- What does the distribution look like, not just the total? Keep visit lengths, timing and repeat visitors.
- How often can the same person count? Set a limit per reader and author.
- What happens when a sign appears? Write down the review and appeal steps before the first case.
What this article cannot tell you
The policies described here are what YouTube, Instagram and Twitch said at particular dates; they change often, and enforcement details are mostly not public. The Steemit study measured one platform in one period, and its bot curators are the authors' suspicion, not confirmed identities. The seven signs are patterns drawn from the research notes behind this article and from platform practice; they are not validated thresholds, and the right ratios differ by platform, format and audience. Twitch's statements are cited through Tubefilter's report, which quotes the chief executive's post. Nothing here assesses any real author or any programme on Dapp.Expert. How incentives shift what authors choose to publish is the subject of the previous part, and whether audience size reflects authority on a topic is the next.
How this article was made
| Task | What the AI did | What a person did | How it was checked |
|---|---|---|---|
| Research | Collected platform policies, studies and detection patterns into research notes | Chose the question and which sources to rely on | Policies, figures and quotations checked against the paper, platform pages and news reports quoting the platforms [1]–[6] |
| Draft | Wrote the first draft from the notes | Edited the text, the structure and the worked example | A separate review compared figures and claims with the sources |
| Images | Generated the cover photograph-style illustration; built the diagrams from a code template | Chose the illustration and the content of the diagrams | Labelled as AI-generated in the disclosure line and alt text |
| Not done by AI | — | The decision to publish and responsibility for the text | — |
Conflicts of interest: none. The author has no commercial relationship with the platforms, researchers or publications named; the programme and authors in the example are fictional.
Responsible for this article: Dmitriy Lastovka, author. Corrections: contact form.
Sources
Labels: official (the platform's own policy or announcement), paper (research paper, read in the original or via its official record), report (a news report that quotes the platform).
- Marilyn Strathern, "'Improving ratings': audit in the British University system", European Review 5(3), 305–321, 1997; quotation on page 308. https://doi.org/10.1002/(SICI)1234-981X(199707)5:3%3C305::AID-EURO184%3E3.0.CO;2-4 Paper.
- YouTube Help, "Fake engagement policy", read 15 November 2026. https://support.google.com/youtube/answer/3399767 Official.
- Instagram, "Reducing Inauthentic Activity on Instagram", 19 November 2018. https://about.instagram.com/blog/announcements/reducing-inauthentic-activity-on-instagram Official.
- Sam Gutelle, "Twitch is punishing streamers who use viewbots by capping their concurrent viewership", Tubefilter, 8 May 2026. https://www.tubefilter.com/2026/05/08/twitch-new-viewbot-policy-dan-clancy-announcement/ Report.
- Chao Li, Balaji Palanisamy, "Incentivized Blockchain-based Social Media Platforms: A Case Study of Steemit", arXiv:1904.07310, April 2019. https://arxiv.org/abs/1904.07310 Paper.
- Anthony Ha, "Medium says it will compensate writers based on reading time, not claps", TechCrunch, 22 October 2019. https://techcrunch.com/2019/10/22/medium-reading-time/ Report.










