ViewDAO
Someone commented, "so it's just another AI detector?"
No. GPTZero answers "did AI write this?" Slop Check answers "does this read like AI to the people who'll actually read it, and what specifically do I fix?"
That's not a subtle difference. It shapes every decision after it.
GPTZero is a good project. We used GPTZero and Originality.ai ourselves when we were testing AI-generated content in our pipelines. They proved that statistical classifiers could reliably distinguish AI text from human text. For catching AI submissions in academic and editorial settings, it works. We have no complaints there.
As we scaled Awwra's content generation for crypto and AI creators on X, posts that passed GPTZero were still getting called out as AI slop in the replies. The detector said "human." The audience said "bot." We saw it happen dozens of times. Clean perplexity score, ratio'd in the comments.
We debated internally: keep using detection tools as our quality gate, or build our own scoring system from scratch.
Two things pushed us to build.
The creator workflow is different from the editor workflow
GPTZero was built for editors, teachers, and publishers. People who receive text and need to verify whether a human wrote it. They need a binary answer: yes or no.
Creators have a different problem. They already know they used AI. They used it on purpose. What they need to know is which specific patterns in their draft will make the audience stop trusting them. They need that before the post goes live, not after someone replies "nice ChatGPT essay bro."
A probability score ("87% likely AI-generated") gives a creator nothing to work with. Which sentence? Which word? Which structural pattern? GPTZero doesn't tell you. You get a number and you're on your own.
Short-form content breaks perplexity scoring
GPTZero uses perplexity and burstiness. It measures how predictable the text is at the token level. Low perplexity means the model picked the statistically obvious next word at every step. High burstiness means sentence complexity varies the way human writing does.
This works on essays, articles, long-form text. It falls apart on a 280-character tweet. There aren't enough tokens for the classifier to be confident. GPTZero's own documentation acknowledges reduced accuracy on short-form content.
Our creators post on X. Most of their content is short-form. We burned two weeks rerunning GPTZero on tweets before accepting this. We needed something that works at tweet length.
We built Slop Check to score what audiences actually react to, not what's statistically detectable at the token level.
Detection vs. craft scoring
Same input. Different question.
GPTZero asks "was this written by AI?" and returns a probability. Forensic answer. Years of production use, academic adoption, batch scanning, integration with plagiarism tools like Originality.ai. They've earned that ground.
Slop Check asks "does this read like AI to the audience that will see it, and what specifically triggers that reaction?" We think this is the right question for creators who use AI on purpose and need to ship content that sounds like them, and for autopilot systems that need to gate content quality without a human reviewing every draft. Whether a model generated the text matters less than whether the audience can tell.
What the difference looks like in practice
When GPTZero flags a post, you get a probability. 100% AI. Okay. Now what? You stare at the text and guess which parts sound robotic. Maybe you rewrite the whole thing. Maybe you change a few words and paste it back in. Trial and error.
When Slop Check flags a post, you see exactly which patterns fired. The em dash on line two. The word "leverage" in the middle paragraph. The three-bullet structure at the end. Fix those three things, and the score moves from a 7 to a 3. You know what to change and why.
Five dimensions instead of one number
We score every post 0-10 across five dimensions:
Original Take - does the post contain a position someone could push back on? A specific claim, a named example, a real number? "AI is changing the game" scores high (bad). "Claude Haiku runs the scorer in 1.2 seconds per post" scores low (good). Specificity is what separates a 3 from a 7.
No AI Tells - the pattern-matching layer. We check for vocabulary that spikes 3-5x in AI text compared to human posts on X. Words like "delve," "pivotal," "robust," "transformative," "tapestry." An em dash is an automatic +2. These aren't style preferences. They're the words that X audiences have already learned to flag. One "utilize" in an otherwise strong post and the replies shift from engagement to "bot account."
Voice Match - this one is harder to explain abstractly, so here's an example. We scored a crypto account's thread about Solana validator economics. Every sentence was grammatically clean. No banned words. But every paragraph opened with a claim, followed by a number, followed by an implication. Same cadence, every time. It could have come from any of 500 accounts running the same prompt. Voice Match: 8 (bad). Compare that to a dev who wrote "validator economics are cooked and here's why I'm still running one" and then rambled for six tweets with typos. Voice Match: 2 (good). The messiness was the proof.
Platform Fit - three bullet points work on LinkedIn. On X, they read as ChatGPT. We saw one account lose 30% of their reply engagement the week they switched to a bullet-point format for every post. Went back to prose, engagement recovered. Platform fit isn't about aesthetics. It's about what the audience on that specific platform has been trained to distrust.
No Bait - fake urgency ("This changes everything"), symmetric takes that refuse to pick a side, hollow closers ("Watch this space"). Content optimized for impressions instead of credibility.
GPTZero gives you one score. We give you five scores and the specific phrases that triggered each one.
Inline scoring vs. external checkpoint
GPTZero lives outside your workflow. You write something, copy it, paste it into GPTZero, read the result, go back to your draft, make changes, paste it again. It's an audit tool.
Slop Check runs inside Awwra's content generation pipeline. Every draft gets scored before it hits your queue. The autopilot uses the score as a gate: anything above 6 gets held as a draft instead of posting. That threshold is configurable per creator.
1-3 publishes automatically. The post reads human. 4-6 gets held. Passable but detectable. You review it, or the generator retries with the specific flags as constraints. 7+ gets queued for full rewrite.
The scoring model is Claude Haiku. Fast enough to run on every single post. The check adds about 1-2 seconds. You don't notice it. Your audience would notice if it wasn't there.
Academic and editorial detection
If you're a publisher vetting 2,000-word articles for AI generation, perplexity-based detection is the right approach. Long-form text gives the classifier enough tokens to be confident. Originality.ai adds plagiarism scanning on top. For catching AI submissions at scale in academic and editorial contexts, these tools are built for the job, and they've been doing it longer than we've existed.
Slop Check is not trying to be a forensic detector. It doesn't tell you whether a human or machine wrote the text. It tells you whether the text reads like a machine wrote it to the specific audience that will see it.
The GEO layer
Slop Check has a sibling we built called the GEO (Generative Engine Optimization) scorer. Different concern, different score.
Slop scoring asks "does this sound human?" GEO scoring asks "will ChatGPT, Perplexity, or Google AI Overviews cite this when someone searches?"
Those AI search surfaces are real discovery channels now. Content that gets cited by them compounds traffic. Every AI answer that references your post sends readers back to you.
The GEO scorer looks at whether the opening sentence directly answers an implied question (AI search engines extract the first self-contained answer they find), whether the passage length sits in the 130-170 word extraction window, how many specific facts are included, and whether the content is self-describing. A post can have a perfect slop score and a terrible GEO score. We tracked this across 1,200 beta posts.
GPTZero and Originality.ai don't do anything like this. They were built to detect AI. We're trying to optimize for how AI systems discover and rank content.
Pricing
GPTZero: Free tier with limited scans, then paid plans starting around $10/month. Originality.ai: Pay-per-scan model. Awwra Slop Check: Included in Awwra. Free for ViewFT users.
Where each tool wins
GPTZero wins when you need forensic detection on long-form submissions at scale. Slop Check wins when you already use AI, your content is short-form, and you need to know what to fix before your audience fixes it for you in the replies.
In our beta, 43% of posts that passed GPTZero with a "likely human" classification still scored above 6 on Slop Check. The detector said clean. The audience wouldn't have.
This article was scored using Slop Check
First creation: Slop Check score 6.6. Three flags.
The closer was engagement bait, the five scoring dimensions were stacked as abstract definitions with no real examples between them, and the ending zoomed back out into a mission statement after spending 1,500 words being specific.
Second creation: Slop Check score 7.5. We cut the engagement bait, broke the definition stack by adding a real Solana validator thread example under Voice Match and a 30% engagement drop anecdote under Platform Fit, and replaced the mission statement with the 43% beta stat. The middle sections still had a rhythm problem, every paragraph was the same weight, same cadence.
Third creation: Slop Check score 8.1. We broke the rhythm. The tool flagged that every section opened with the same structure. Context sentence, explanation, payoff. So we changed one section opener from a full paragraph to just two short sentences, and let the surrounding sections carry the detail.
It failed twice before it passed. Every fix came from specific flags the scorer surfaced, not gut feel. That's the difference between a probability and a diagnostic.
Try it at https://awwra.app/#slop-check
Check out the GitHub repo here: https://github.com/Viewfin-Labs/Slop-Check