How to post about AI on LinkedIn without sounding generic

To post about AI on LinkedIn without sounding generic, publish something nobody else could publish: your own numbers, your own build, your own compiled work. Our cohort of 12,988 ranked posts contains 397 about artificial intelligence from 37 creators, and only 12 of them reached the cohort-wide top decile. By chance you would expect about 40. That gap is the most useful thing we can tell you about this topic. AI content is not underperforming because the subject is boring. It is underperforming because 40.6% of it is a link to someone else's article.

How saturated is AI content on LinkedIn?

Saturated enough that it shows up as a measurable performance penalty. Here is the clearest way to put it. Our top decile is defined cohort-wide: the top 10% of all 12,988 posts by engagement rate, which required 5.95 engagement points per 1,000 followers. If AI posts performed like everything else, roughly 40 of the 397 would clear that bar. Twelve did.

A second version of the same finding is harsher. When we rank AI posts against each other and take that topic's own strongest decile, 39 posts from 16 authors, the median engagement rate of that group is 4.22 per 1,000 followers. The best AI content in our data, at the median, would not qualify for the cohort-wide top decile. Almost every other topic's own top decile clears the bar comfortably.

The topic's medians tell the same story from the floor: 87 reactions and 7 comments for the median AI post, against 122 and 12 for the cohort. Comments are the scarcer signal in every cut of our data, and AI content earns fewer of them than anything else we measured. People are not arguing about these posts. They are scrolling past them.

Why does AI content underperform?

Three habits, all visible in the data, and all of them are the same habit wearing different clothes: the post contains nothing the author made.

It is mostly links

Shared articles account for 40.6% of all AI posts, the highest share of any topic in our corpus, and only 10.8% of AI posts are plain text, the lowest. That is the profile of a newsfeed rather than a body of writing. When the news cycle produces a model release, several thousand people share the same coverage within a day, and the feed treats the tenth copy the way you would.

It addresses nobody

Second person appears in 23.2% of AI hooks overall, one of the lowest rates in our data. For comparison, the productivity topic runs 49.8%. AI writing tends to describe a technology rather than talk to a person, which is a natural consequence of writing about a subject instead of writing about work.

It is buried in hashtags

63.2% of AI posts carry four or more hashtags, the highest stuffing rate of any topic we measured, and it rises to 71.4% in the topic's weakest half against 41.0% in its strongest decile. The most common tags in the category are broad and interchangeable. A tag stack is what people add when the post has nothing else to distinguish it. See the hashtag data for the cohort picture.

What separates the best AI posts from the rest?

The comparison below is within-topic: the AI topic's own top decile (39 posts, 16 authors) against its own bottom half (199 posts, 20 authors), out of 397 posts from 37 creators. It holds subject matter constant, so it answers the question you actually care about: among people posting about AI, what do the better ones do?

The first row is the loudest. Plain text posts are more than three times as common in the top decile as the bottom half, 20.5% against 6.0%. In most topics, format effects favour video. Here the standout is simply writing something. In a category where the default act is forwarding, an original paragraph is a differentiator.

The median top-decile AI post drew 312 reactions and 38 comments. The median bottom-half post drew 54 reactions and 3 comments. So the ceiling in this topic is normal. It is the floor that is unusually crowded.

That distinction matters for how you should read the saturation finding. It is not that AI content cannot work. Posts in this topic reach 300 reactions and 38 comments at the median of their own top decile, which is a perfectly good outcome. It is that the ordinary version of an AI post, the one most people publish, performs worse than the ordinary version of a post about anything else. The competition is not for attention on the subject. It is against several thousand people saying the same thing on the same afternoon.

What does a generic AI post look like?

You already know, because you have scrolled past forty of them this week. It is worth being precise about the failure modes, though, because most of them are recoverable with one edit. The examples below are invented illustrations, not posts from our dataset.

The pattern in the right-hand column is the same every time: a first-hand observation with a cost, a failure, or a number attached. Nothing in it requires you to have an opinion about artificial general intelligence.

What should you post about AI instead?

The strongest AI post in our exemplar set is a compilation the author built because it did not exist. That is the whole model for this topic.

Five kinds of AI post that carry information rather than commentary:

• Something you compiled. A list, a benchmark, a comparison you ran yourself. Expensive to make and impossible to copy, which is why it travels.
• A deployment with numbers. What you shipped, what it cost, what it broke, what the actual time saving was after the second month. Almost nobody publishes the second month.
• A failure. The pilot that did not work and the reason. This is the single most under-supplied kind of AI content on the platform, and the easiest to write honestly.
• A domain-specific take.What this technology means for structural engineers, or veterinary practices, or bid writers. General AI commentary competes with everyone. Your industry's version competes with almost nobody.
• An argument with a position. Not a prediction, which nobody can check, but a claim someone could disagree with today.

That second example is instructive on its own. The account has around 3,000 followers, and the post produced more comments than reactions, which is rare anywhere in our data. It did that with one concrete fact and one inference. No tools list, no prediction about 2030.

Who is actually posting about AI, and does it work?

The tag data on this topic is revealing in a way we did not expect. The most common hashtags across our 397 AI posts are digitalhealth (194 posts), technology (192), healthcare (185), medicine (164), and future (162), ahead of ai itself (143) and artificialintelligence (126). In other words, the AI conversation in this corpus is disproportionately medical, written by people whose actual subject is healthcare and whose AI content is a wing of that.

Two things follow. First, a methodological one: a large slice of this topic comes from a small number of prolific domain writers, which is exactly why we publish the author counts alongside every figure. Second, and more useful: the writers who sustain an audience on this subject are not AI commentators. They are specialists whose domain happens to intersect with it. The strongest post in our exemplar set is a list of FDA-approved algorithms, which is a healthcare post that happens to be about AI rather than an AI post that happens to mention healthcare.

That is the most portable lesson in this article. Nobody needs another generalist explaining transformers. There is an open lane in every industry for the person who can say what this technology does to their specific work, in the vocabulary of that work, with examples their peers recognise.

How do you post about AI if you are not technical?

Better than the technical people do, usually, because the shortage is not in explanation. It is in first-hand reports from the places where the technology meets an actual job. You do not need to understand how a model works to write the most useful post on your feed this week.

• Report the buying decision. What you evaluated, what you rejected, what you paid, what the procurement process was like. Almost nobody writes this and everybody reading is about to face it.
• Report the workflow change. Which step in your process disappeared, which one got worse, what your team argued about. Concrete and completely uncopyable.
• Report the customer reaction. If your industry has started using these tools with clients, what the clients actually said is information that almost never gets published.
• Report the policy. What your organisation decided about usage, disclosure, or client data, and how the decision was made. Every company is writing this document right now, in private, badly.

What to avoid: explaining the technology to people who can search for the explanation, and predicting adoption curves. Both are freely available and neither is yours.

A useful test before publishing anything in this topic: could a competent stranger have written this post after twenty minutes of reading? If yes, they probably already did, several times, this week. The posts that survive the saturation are the ones where the answer is obviously no, and the reason is usually that the author spent time or money finding something out.

Should you post about AI news the day it happens?

Only if you have run it. The reflexive same-day post about a launch is the highest-volume, lowest-return act in this category, and the shape of the data supports that: 42.2% of the weakest AI posts are shared articles, and the weakest half of the topic drew a median of 3 comments.

The version that works is deliberately late. Wait a week, use the thing on a real task from your own week, and publish what happened, including the parts where it failed. The audience for launch-day commentary is enormous and has thirty options. The audience for “I tried this on our actual workload and here is where it broke” is smaller and has almost none.

There is a cost to being late, and it is worth naming: you give up the small reach spike that comes from posting into a live conversation. Our data cannot price that trade-off directly. What it can say is that the shared-link pattern which dominates same-day commentary sits heavily in the bottom half of this topic, and the original-work pattern sits at the top.

Do AI posts need to be optimistic or pessimistic?

Neither, and the choice is a trap. The optimistic version and the doom version are equally generic, because both are positions rather than contributions. Our data cannot score sentiment, so this is judgment: what the strong posts in this topic have in common is that they are about a specific thing that happened, not about where the technology is heading.

The exception is when the position is unusual for the person holding it. An engineer saying the technology is oversold, or a sceptic describing the first thing that changed their mind, carries information because it is costly to say. A consultant saying AI will transform your business carries none, because it is what a consultant would say either way.

Should you say that AI helped write your post?

This comes up constantly and it deserves a straight answer. There is no LinkedIn rule requiring you to label an ordinary text post as AI-assisted, and opinion among readers is genuinely divided rather than settled. What we can say from our own data is narrower and more useful: nothing in the winning profile of any topic is about how a post was drafted. It is about whether the post contains a specific, checkable, first-hand thing.

The line most people converge on in practice is between assistance and fabrication. Using a tool to structure, tighten, or draft your thinking is editing. Publishing an experience you did not have, a client story that did not happen, or a number you did not measure is something else, and it is the only version that reliably damages people when it comes out.

We have a longer treatment of detection, disclosure, and what actually gives generated text away in the AI content detection guide, and a practical version of the writing problem in how to make AI writing sound human. If you are choosing tooling, the post generator comparison covers what the category can and cannot do.

What did the AI posts that worked have in common?

Rereading the twelve AI posts that reached our cohort-wide top decile, plus the wider set of 39 that top their own topic, the shared properties are unglamorous.

1. The author did something. Built, tested, compiled, deployed, or paid for it. The post reports rather than reacts.
2. It is grounded in a domain. Healthcare, advertising, logistics. Not AI in general.
3. It is written natively. Text or video, not a link with a caption. Plain text runs 20.5% at the top against 6.0% at the bottom.
4. It has a person in it.First person runs 17.9% at the top, and second person 33.3% against 24.1%. Both pronouns matter more here than the topic's reputation for technical detachment would suggest.
5. It is not tagged to death. Four-plus hashtag stacks fall from 71.4% at the bottom to 41.0% at the top.

More hooks from this topic with the numbers attached are on the AI examples page.

How should you judge whether an AI post worked?

Adjust your expectations before you publish, because the baseline in this category is low. The median AI post in our cohort drew 87 reactions and 7 comments; the median post overall drew 122 and 12. If you hold AI content to the numbers your other posts get, you will conclude it does not work and stop, which is the wrong conclusion from the right data.

Two things to watch instead. The first is comments from people who do the work you are describing. One reply from a practitioner in your field is worth fifty reactions from strangers, because it is evidence the post reached the audience it was written for. The second is what happens off-platform: saves, forwards, and the meeting that gets booked because somebody read it. AI content skews toward decision-makers researching a purchase, which is a small, quiet, valuable audience that shows up in your calendar rather than your notifications.

The failure signal worth acting on is comments that could have been written without reading the post. If the replies are all agreement and emoji, the post did not say anything specific enough to answer, and the fix is more detail, not more frequency.

How do you write an AI post that is not generic?

1. Start from something you did this month. Not from a headline. If nothing comes to mind, you do not have a post yet, and that is a legitimate outcome.
2. Put the specific detail in the first line. The model, the task, the number, the industry. The visible hook runs about 206 characters, and generic openers spend all of them saying nothing. See the hooks study.
3. Include the part that failed. It is the most credible sentence available to you and almost nobody uses it.
4. Cut every prediction. Anything about 2030 can be deleted without loss.
5. Write it as text or film it. If your instinct is to paste a link, write your argument first and put the link at the end. The link analysis covers the trade-off.
6. Cap the hashtags at two. Ideally a domain tag rather than the generic ones.
7. Address someone. Name the job title this actually matters to. Broad audience, no reader; narrow audience, real replies.

What this analysis cannot tell you

These are correlations, not causes. The AI topic is 397 posts from 37 creators, and its own top decile is 39 posts from 16 authors, so individual habits carry weight in the percentages. The cohort skews toward established creators with at least 1,000 followers across 65 accounts. Engagement counts are lifetime-cumulative at scrape time and post ages vary, which matters here because this topic moves faster than most: a post about a model release ages in weeks. Every text finding describes the visible hook rather than the full body. And the saturation figure is a statement about this cohort, not a law of the platform. Full method, filters, and caveats are in the 34,000-post engagement study, with percentile benchmarks in the benchmarks guide.

Writing about a different subject?

Each subject behaves differently, and we compared every topic against its own top decile rather than against the whole cohort: DEI, burnout and mental health, productivity, remote work.