Problem: when generated posts underperform, the Instagram "AI info" label gets blamed, and people look for ways to strip AI indicators from files instead of checking what their audience responds to. Result: a tiny script that splits your own posts into live and fully generated, and compares average views, reactions and production time per group. Stack: Python 3.10+ standard library (tested on 3.15.2). Boundaries: you export the numbers yourself from Instagram insights; the script does not call any API, and a handful of posts is a signal, not a statistically solid result
Background: how the label works
Under Meta's rules the label appears when Meta detects industry-standard AI indicators that generators write into the file, or when the creator discloses AI use. Since September 2024, content only edited with AI tools has the label moved into the post menu; fully generated content keeps it visible. Removing indicators from files works against a platform rule and does not change how people react to the content itself
In my own work, live Reels, photos and posts get many times more views and reactions than fully generated ones. That is the comparison worth running on your account before deciding anything
Input, processing, output
Input: a list of posts, each with a generated flag (true only if the whole visual was made by AI), views, reactions and the minutes it took to produce
Processing: split by the flag; compute averages per group
Output: one line per group with post count and averages
The example
The numbers below are invented to show the output shape; they are not my account's statistics
from statistics import mean
def compare(posts: list[dict]) -> dict:
"""Split posts into live and generated, return averages per group."""
groups = {"live": [], "generated": []}
for p in posts:
groups["generated" if p["generated"] else "live"].append(p)
out = {}
for name, items in groups.items():
if items:
out[name] = {
"posts": len(items),
"avg_views": round(mean(p["views"] for p in items)),
"avg_reactions": round(mean(p["reactions"] for p in items), 1),
"avg_minutes": round(mean(p["minutes"] for p in items)),
}
return out
if __name__ == "__main__":
posts = [
{"generated": False, "views": 2100, "reactions": 96, "minutes": 25},
{"generated": False, "views": 1650, "reactions": 71, "minutes": 30},
{"generated": True, "views": 420, "reactions": 9, "minutes": 10},
{"generated": True, "views": 380, "reactions": 6, "minutes": 8},
{"generated": False, "views": 1900, "reactions": 88, "minutes": 20},
]
for name, s in compare(posts).items():
print(name, s)
Output, exactly as printed when I ran it:
live {'posts': 3, 'avg_views': 1883, 'avg_reactions': 85, 'avg_minutes': 25}
generated {'posts': 2, 'avg_views': 400, 'avg_reactions': 7.5, 'avg_minutes': 9}
Read it as a ratio, not as absolute numbers. In this invented sample generated posts are faster to make but bring a fraction of the response. If your real export looks like that, removing the label would not fix anything
Design notes
Define "generated" strictly. A photo you shot and retouched with AI is live for this comparison; only content with no real person or footage counts as generated. Otherwise the groups blur and the comparison says nothing
Include production time. Generated posts usually win on minutes. The question is whether the time saved is worth the response lost, and you can only answer that with both numbers side by side
Compare like with like. Reels and static posts behave differently; run the script per format if you have enough posts
Use a window, not a single month. Twenty or thirty recent posts per format is a reasonable start
Record the flag at publishing time. Deciding later which posts were generated is unreliable. Add the flag to whatever sheet or content plan you already keep, so the comparison is ready whenever you want to run it
What I do instead of hiding the label
I run a content factory that scales my work rather than replacing it. It watches 13 sources around the clock, checks facts before drafting and writes separate versions per platform, which I approve one by one. A post that took about an hour by hand now reaches a ready preview in about 5 minutes; 84 posts have been published that way, none without my approval
For live video, AI handles the tedious part: I send one word in Telegram and a raw clip gets subtitles and is ready in about 2 minutes. The face and voice are mine, so there is no label to worry about
Errors and limits
Small samples swing wildly; look at trends over several weeks
Reach also depends on timing, topic and hashtags. Keep those roughly comparable between groups or note the differences
The script treats every post equally. A single viral post can dominate an average; consider a median if your numbers vary a lot
Reproduce it
- Save the code as
compare.py - Run
python3 compare.pyand compare with the output above - Replace the list with your own last twenty or thirty posts from Instagram insights
- Run again per format: Reels separately from photos
The non-technical version is on my site: why a content factory should scale, not replace
This article was created with the help of AI.



Top comments (0)