A lot of YC startups get initial traction thanks to their Launch HN posts (unfortunately Lettercamp isn’t a YC startup, so we don’t quite qualify - but Show HN has still been immensely helpful). The Launch HN guidelines don’t mention or prohibit the use of AI in the posts, but they do say to “Write in your own voice”.
I wanted to research if there were any interesting trends in Launch HN posts so that I could make a better Show HN post for my own product. I looked at over 100 Launch HN posts over this past year and used Pangram to see if they were written by AI or not, and then tried to see if there was any correlation between the use of AI and the amount of engagement the posts got.
Here’s the summary of data in case you don’t want to read the whole article:
- I looked at 108 Launch HN posts over the last year
- 22 (20.4%) were marked as AI-generated
- 44 (40.7%) were marked as mixed
- 42 (38.9%) were marked as human written
- Human written posts had a median of 83 upvotes, mixed posts had a median of 80 upvotes, and AI-generated posts had a median of 56 upvotes
So interestingly, it does seem that writing your Launch HN post by hand instead of using AI results in better engagement, although the difference between fully-human written and mixed is small. In the rest of this post, I’ll share more of the details behind the methodology, as well as link to some of the post examples.
Methodology and Data
I first scraped the last years worth of Launch HN posts from Hacker News, then isolated the text of the post. I then passed the text itself to Pangram to understand whether it was written by AI or not.
I used Pangram’s classification to determine whether the post was AI-generated or not. Pangram returns a score and classification based on what it determines the text to be. If the inputted text scores more than 80% AI-generated, then it’s marked as AI. If it’s 90% or more human written, then it’s marked as Human. Everything else and in between is marked as mixed.
| Pangram label | Posts | Share |
|---|---|---|
| AI | 22 | 20.4% |
| Mixed | 44 | 40.7% |
| Human | 42 | 38.9% |
| Total | 108 | 100% |
Some examples of Launch HN posts marked as AI: Chamber, Constellation Space, General Instinct, Livedocs, and TeamOut.
Examples of Launch HN posts marked as human-written: AgentMail, BrowserBook, Momentic, Onyx, and Parsewise.
If you’d like to audit the full dataset, you can do so through this public spreadsheet here.

The results for the Chamber Launch HN post as specified by Pangram
Do human-written posts get upvoted more?
In short, yes. The median number of upvotes for human-written posts was 83, while the median number of upvotes for AI-generated posts was 56. I think it’s clear that if you want more engagement, you should write your Launch HN and/or Show HN posts by hand instead of with an LLM.

Median upvotes for each of the categories of posts
Do posts marked as AI really sound like AI?
I think they do. Here’s an excerpt from the Chamber launch:
One thing that surprised us, even coming from Amazon where we’d seen large GPU fleets: most teams we talk to can’t even tell you how many GPUs are in use right now. The monitoring just doesn’t exist. They’re flying blind on their most expensive hardware.
The short sentences and rhythm of the writing does sound like it was generated by an LLM.
Interesting outliers
Here are some interesting outliers from the data:
- Among human-written posts, Freestyle had the most points (322), and Sentrial had the fewest (31).
- The mixed group had the highest-scoring post overall, Tweeks (351 points), a browser extension for customizing websites. Lowest in this category is Sonarly (30 points).
- RunAnywhere was the most upvoted post in the AI-written category with 240 points (for some reason, the original post text doesn’t show up on this post anymore, but it is available still in our dataset). Chamber, an AI agent for managing GPU infrastructure, had the lowest score overall at 26 points.
So if we take just the outliers, it is interesting that even the most upvoted AI-generated Launch HN post scores lower than the high scoring posts of the other categories.
Takeaways
I think it’s clear that on some subconscious level, audiences might be able to detect when a post isn’t fully written by a human, and that may affect their willingness to upvote a post. I would suspect that the Hacker News audience is very in tune with the style of an LLM’s writing and that may affect it even more.
In the age of AI writing, it still feels that human-written content is more effective at reaching an audience. I hope to keep writing like that, directly to readers, in Lettercamp launches and blog posts in the future as well. Thank you for reading!

