We Ran 100 Social Media Posts Through Devoke. Here's What We Found.
We pulled 100 posts from public feeds across Twitter/X and Reddit — a mix of news accounts, political commentary, sports, technology, and general content — and ran each through Devoke's rewriter. We categorised the results and looked for patterns.
The posts are not attributed here. We're reporting patterns, not calling out individuals. The findings are what's interesting.
How they broke down
| Category | Count | % of sample |
|---|---|---|
| Unchanged — already calm and factual | 21 | 21% |
| Lightly rewritten — mild loaded language removed | 34 | 34% |
| Significantly rewritten — substantial manipulation removed | 31 | 31% |
| Flagged with [unverified claim] | 14 | 14% |
The headline finding: roughly 79% of posts in our sample had some detectable manipulation layer. One in five was already calm enough that Devoke left it alone.
What "unchanged" looked like
The posts Devoke left alone tended to fall into a few clear categories:
- Personal updates ("Just got back from two weeks in Japan. That was exactly what I needed.")
- Straightforward factual announcements ("The match kicks off at 3pm BST. Here's the lineup.")
- Clearly flagged opinion ("I think the new policy is misguided, and here's why — [thread]")
- Humour that was obviously humour ("My cat has once again rejected the expensive food and is eating cardboard.")
- Grief and personal loss
The model is reasonably good at recognising genuine human expression. It almost never rewrote personal posts — which matters, because the goal is to filter manufactured outrage, not to sanitise how people talk to each other.
What "lightly rewritten" looked like
The largest category. These were posts where the framing was slightly loaded — not egregiously manipulative, but carrying a charge that wasn't warranted by the facts.
Common patterns:
- Unnecessary intensifiers: "absolutely devastating" → "significant"; "utterly failed" → "failed"
- Unearned certainty: "will destroy" → "may affect"; "proves" → "suggests"
- Vague but alarming qualifiers: "sources say something alarming is coming" → the claim was so vague it became "reports suggest an upcoming development" — and revealed how little was actually being said
The light rewrites were often the most revealing. When the loaded language was stripped and you could see what remained, some posts contained almost no information at all. The emotional charge had been the content.
What "significantly rewritten" looked like
The strongly sensationalist posts. These were the ones where the rewrite was most dramatic.
Representative pattern — original:
"The MAINSTREAM MEDIA is TERRIFIED of this story getting out — big tech is doing everything they can to suppress it. SHARE THIS EVERYWHERE before it disappears!!"
Rewritten:
"This story has received limited mainstream coverage. The author believes it is being suppressed by major tech platforms."
The rewrite doesn't validate or invalidate the claim — it just removes the engineered urgency and presents what's actually being said. The reader can evaluate the claim itself, rather than the emotional packaging around it.
Another common pattern in the significant category: news headlines that dramatically overstated what the underlying story contained. The headline would say "SHOCKING new development EXPOSES government failure" — and the article (when you read it) described a moderately critical audit finding that had been covered elsewhere without drama.
Where the model struggled
Roughly 8% of the rewrites in our sample were ones we'd flag as imperfect. Three types of error:
Softening genuine urgency. Some posts expressed real alarm about real situations — a developing emergency, a time-sensitive public health message — using language that was urgent because urgency was warranted. The model occasionally softened these when it shouldn't have. This is the most significant error type, and it's why the toggle to the original is non-negotiable.
Missing dry irony. Deadpan posts that appeared sensationalist on the surface but were clearly ironic in context occasionally got rewritten when they should have been left alone. The model reads text, not tone of voice.
Factual precision loss. In a small number of cases, the rewrite generalised a specific claim into something vaguer — technically more neutral but also less informative. "The species faces an 80% population decline" becoming "the species faces significant population decline" loses the specific number that might actually matter to the reader.
The most surprising finding
We expected the most sensationalist content to come from explicitly political accounts. It did — but the second-largest source was sports commentary, which we hadn't anticipated.
Sports commentary on social media uses almost exactly the same linguistic patterns as political outrage content: superlatives, ALL CAPS, certainty about outcomes, tribal framing ("WE were robbed," "THEY cheated"). The emotional register is identical. The stakes are lower, but the manipulation techniques are the same.
Devoke rewrites sports posts the same way it rewrites political ones. "That referee DESTROYED us with that call — absolutely DISGRACEFUL" becomes "The referee's decision was controversial and cost us the match." Whether you want that depends on how much of sports' emotional intensity you want preserved. (For what it's worth, at Deep intensity, the model is more aggressive; at Light, it would have left most sports commentary alone as relatively harmless emotional expression.)
What this suggests about the feed
If our sample is representative — and we have no strong reason to think it isn't — then roughly 80% of social media content contains some detectable manipulation layer. Not all of it is egregious. Most of it is mild: a word here, a framing choice there. But it accumulates.
The average social media user sees hundreds of posts per day. If 80% of them carry even a small emotional charge that isn't warranted by the content, the aggregate effect on how you feel after a scrolling session is not nothing. It's not the single outrageous post that does the damage — it's the sustained ambient charge of content that's always slightly more alarming than it needs to be.
That's what Devoke is trying to address at scale.