Burned-in captions are one of the few format decisions that is close to unambiguously good. A large share of short-form is watched with sound off or partially off — in public, in bed, at a desk, in a room with other people — and for those viewers a video without captions is a video without content.
Since watch time is reported to be the heaviest ranking input, anything that keeps a sound-off viewer watching is not a courtesy. It is distribution.
Where it turns
The trouble is that captions are easy to escalate. One line becomes two, two become a block, the block gets a highlight colour, the highlight animates word-by-word, and at some point the frame is mostly text with a person behind it.
That endpoint is named directly in Instagram's own guidance: it aims to make “reels that are majority text” less visible, in the same sentence that covers watermarks, borders and low resolution. So the escalation has a cliff at the end of it, and nothing warns you as you approach.
Captions are additive until they become the frame. Then they are a demotion you cannot see in your own preview.
A working limit
Two lines at a time, occupying a band rather than a region. If a still frame from the middle of the video would read as a text post with a photograph behind it, you are over the line. If it reads as a video with words on it, you are not.
The animated word-by-word style deserves specific caution. It is popular, it does hold the eye, and it also tends to demand a larger type size and more screen area than static captions, which pushes toward the same cliff faster.
Placement, which is mostly about what you cannot see
Captions sitting too low collide with the post caption, the username row and the interface controls. Captions too high collide with the top furniture. Both produce the same outcome: words that are technically present and functionally missing, on the devices you did not check.
The band that is reliably clear is narrower than most editing previews imply, because the preview does not draw the interface. Checking a real post on a real phone once is worth more than any guideline, and almost nobody does it.
Accuracy is not optional
Auto-generated captions are good and not perfect, and they fail hardest on exactly the words that carry your meaning — names, jargon, numbers. A caption that mistranscribes the one term the video is about does more damage than no caption, because the sound-off viewer reads the wrong claim confidently.
Read them before export. It takes under a minute and it is the highest-value minute in the edit.
The honest summary
Caption everything, keep it to a band, read it before you post, and check it once on a real device. That is the whole of it. The gains are real and the failure modes are all mechanical — which puts captions in the same category as the five demotions Instagram names: not a creative problem, a checklist problem.