# Can an AI score predict which reels go viral?

> We scored 200+ real Instagram reels with an AI script checker and compared the scores to their views. The scores predicted nothing. Here's the data.

Canonical: https://www.algrohack.com/research/can-ai-predict-viral-reels
Last updated: 2026-11-02

By The algrohack team. Published 2026-11-02.

## The short answer

No, not from the script alone. We scored the spoken words of more than 200 real Instagram reels with an AI checker and compared the scores with the reels' real views.

Across four tests the scores didn't track performance: every overall correlation landed between -0.20 and 0.09, which at this size is noise. Given two reels by the same creator, the higher-scored one did better 46% of the time, worse than a coin flip.

What decided performance mostly wasn't in the words: the account that posted it, what was on screen, how it was delivered, the audio and the timing.

## Why we tested this

algrohack has a Virality score: it reads a script and rates it against the signals Instagram's own insights list under "What affects your views": skip rate, shares, likes, saves, reposts, comment rate and watch time. The 0-100 number is calculated from those ratings, not guessed by the model.

Before calling that number a virality score, we wanted to know whether it predicts anything. So we ran it on reels whose results we already knew.

## The data

The reels came from algrohack's own cache: reels creators brought into the app and reels our niche search found, each with a transcript of what is said and its view count. About 220 had both. Some tests needed more (follower counts, or several reels by one creator), so each test says how many reels it used.

The scorer was the live Virality score. To check whether it could be taught, we also built calibrated versions: a large language model studied the best and worst performers in a training set (70% of the reels, per language) and wrote a rubric, then scored reels it had never seen (the other 30%).

A correlation of 1 would mean the scores ranked the reels exactly by views, 0 means no relationship, and a negative number means higher scores went with fewer views. With samples this size, anything between about -0.3 and 0.3 can't be told apart from chance.

## Test 1: Score against views

Do higher-scored reels get more views? 20 reels, five from each views band.

| Views | Reels | Average score |
| --- | --- | --- |
| 1M and up | 5 | 41.8 |
| 100K to 1M | 5 | 52.4 |
| 10K to 100K | 5 | 53.6 |
| Under 10K | 5 | 56.6 |

The most-viewed reels scored lowest. The correlation between score and views was -0.20, and none of the seven signals tracked views on its own either.

A reel with 7.2 million views scored 23: a demo whose point was on screen, so its words alone read as flat. Two AI-tool demos with 2 million and 1.3 million views scored 30 and 34. Meanwhile a reel with 4,000 views scored 74, and one with 34,000 scored 83: tidy, well-built scripts, probably from smaller accounts.

## Test 2: Teach it with real reels

Does the score improve if the model first learns from real hits and flops? Trained on 70% of about 220 reels, tested on 58 it had never seen.

| Scorer | Correlation with views |
| --- | --- |
| The checker as it was | -0.14 |
| Calibrated on real reels | 0.09 |

Calibration removed a slight backwards lean but still ranked reels no better than chance. Run separately for each language in the set, results ranged from -0.28 to 0.12. The rubrics it learned were generic advice that shows up in hits and flops alike.

## Test 3: Remove the size of the account

Views mostly measure who posted, so we changed the target to views relative to the creator's followers, and compared reels by the same creator.

| Scorer | Views ÷ followers (62 reels) | Same creator: higher score won |
| --- | --- | --- |
| The checker as it was | -0.14 | 46% |
| Calibrated on real reels | 0.02 | 45% |

The same-creator check is the fairest test: 153 pairs of reels from 21 creators, so account size cancels out completely. Picking the higher-scored reel as the winner was right 46% of the time. A coin flip is right 50%.

## Test 4: Calibrate on the creator's own reels

Does it work if the model learns from the creator's own history? 9 creators, 65 of their reels, each reel hidden in turn and predicted from that creator's others.

| Scorer | Which of two reels by the same creator did better? |
| --- | --- |
| Calibrated on the creator's own reels | 39% right (71 of 181 pairs) |
| The checker as it was | 52% right (104 of 199 pairs) |
| A coin flip | 50% |

With five to nine reels per creator, the model learned coincidences ("their two hits were both about money"), not patterns. Below 50% here means unstable, not reliably wrong.

## Why words alone can't predict views

- **The account.** Across different creators, views mostly reflect who posted: followers, audience and how often they post. A strong script on a small account can get fewer views than an average one on a big account.
- **What's on screen.** Demos and screen recordings carry their value visually. A transcript of them reads flat, which is why the biggest demo in our set scored 23.
- **Delivery, audio and timing.** How it's said, the first frame, the sound, the trend and the day it goes out aren't in the words at all.
- **Too little data.** A few hundred reels, and under ten per creator, is too few to learn anything subtler than "have a strong hook", which hits and flops both had.

## What this means for creators

- Treat any "virality score" as a checklist for your script (is the hook clear, is there a reason to save or share), not a forecast of views.
- Judge a reel against your own usual numbers, never against someone else's views.
- Put as much work into the first frame and your delivery as into the words.
- After posting, learn from the reel's real data in Instagram's insights: skip rate, watch time, shares and saves.

## What we changed in algrohack

algrohack's Virality score still checks a script against the signals Instagram lists and suggests fixes for the weak ones. It never predicts views or the chance of going viral, and the panel says so: how a reel does also depends on how you speak, what you show on screen, the audio, the timing and the account.

## Limits of this study

- Small samples: 20 to about 220 reels per test, so modest effects could be missed.
- Not a random sample of Instagram: the reels are ones creators brought into algrohack or our niche search found.
- Views were read once, at a single moment, not at a fixed age.
- One model family and text only: the scorers read what is said, never the video.
- Run on 13 and 14 October 2026.

## Questions

### Can AI predict whether a reel will go viral?

Not from the script alone, in our tests. Scores from an AI checker didn't track the real views of more than 200 Instagram reels, and for two reels by the same creator the higher score did better only 46% of the time.

### Is an AI virality score useless, then?

As a forecast, yes in our data. As a checklist it still helps: it points out a weak hook or a script with nothing worth saving or sharing. Just don't read the number as a prediction of views.

### What does predict reel views?

In our data the account mattered more than the words: followers, audience and posting habits. Visuals, delivery, audio and timing aren't in a transcript at all. The fairest comparison is a reel against the same creator's usual numbers.

### Does algrohack's Virality score predict views?

No, and it doesn't claim to. It rates a script against the signals Instagram lists and suggests fixes. This study is why it says so.

## Related

- [About algrohack](https://www.algrohack.com/about)
- [AI script generator](https://www.algrohack.com/tools/ai-script-generator)
- [Script timer](https://www.algrohack.com/tools/script-timer)