← AI Slop Detector
Updated April 2026

What the detector
is actually doing.

AI Slop Detector uses a weighted scoring pipeline rather than a one-rule shortcut. It combines transformer inference, phrase-density heuristics, sentence-level inspection, and explicit short-sample calibration to estimate whether a passage behaves more like AI-generated prose or ordinary human writing.

The short version

The transformer model is the primary classifier. Heuristics act as a smaller correction layer, sentence scoring makes the output inspectable, and low-word-count samples are intentionally compressed toward neutral because short text has high variance. The system is built to produce an evidence-weighted signal, not a proof of authorship.

Scoring Pipeline

Every input is normalized, split into chunks for long passages, passed through a RoBERTa-based detector, scanned for tracked phrase patterns, and then combined into a final scalar score on the interval [0, 1], where higher values indicate stronger AI-like behavior.

Primary Signal
82%

For longer passages, the transformer model dominates the blend because it captures broader language patterns better than surface-level rules.

Secondary Signal
18%

Phrase heuristics stay intentionally lighter. They are useful for explainability and nudging borderline samples, but they should not overpower the model.

m = (1 / n) Σ mi   where mi is the detector score for chunk i
h = min(((Σ cjwj) / (word_count / 100)) / 12, 1)
long-text score = 0.82m + 0.18h

In the heuristic term above, cj is the match count for tracked phrase j and wj is that phrase's configured weight. The density is normalized by words-per-100 so a long article is not penalized just for having more total tokens.

Signal Blend
The model is the main classifier. Heuristics remain a smaller but interpretable adjustment layer.
Short-Range Calibration

Short text is handled differently because one or two sentences produce unstable model variance. For samples under 120 words, the detector adds a sentence-average term. For samples under 50 words, it also applies an explicit trust factor that shrinks the score back toward a baseline.

s = average sentence score for sentences with at least 5 words
if w ≤ 60: b = 0.55m + 0.25s + 0.20h
if 60 < w ≤ 120: b = 0.62m + 0.23s + 0.15h
if w < 50: final = baseline + (b - baseline)t

The trust term t is bounded so very short text cannot swing too far in either direction. That is why a 20-word sample may return a low-reliability or unclear reading even when the raw model score is more extreme.

Weight Schedule by Length
The blend changes with sample length: shorter inputs get extra sentence weighting and stronger score compression.
Explainability

The app surfaces more than a raw number. It also ranks individual sentences and exposes weighted phrase matches so the output can be inspected, not just accepted at face value.

ML
Sentence signals are sorted by descending sentence score, so the highest-ranked lines are the ones that contributed most strongly to the AI-leaning side of the verdict.
P
Phrase hit breakdowns report the phrase label, its configured weight, match count, and aggregate contribution term c × w, which is then folded into the heuristic density.
R
Reliability messaging explicitly labels low-information cases instead of converting uncertain evidence into a fake high-confidence verdict.
Reliability by Word Count
Below 50 words, the detector explicitly treats the result as weak evidence. Reliability improves as sample size grows.
Short Text Handling

Very short text is the hardest case for AI detection. A 20-word sample simply does not contain enough syntactic, lexical, and distributional structure for a robust verdict, so the detector compresses those scores back toward neutral and may mark them as low reliability.

if w < 50:
t = clamp(0.08, 0.32, ((w - 50) / 30) × 0.24 + 0.08)
score = baseline + (b - baseline)t
20
20-word minimum keeps the app from pretending that tiny samples are trustworthy.
50+
50+ words recommended because the model and heuristics have enough material to become directionally useful.
90+
90+ words generally gives the detector its most stable and interpretable behavior.
Verdict Confidence Curve
Confidence grows with sample length, but short texts are intentionally flattened so they do not look more certain than they are.
Limits

No detector can prove authorship. Heavily edited AI writing can look human. Highly structured human writing can sometimes look model-like. Scraped text can also be noisy if the page extractor pulls boilerplate, disclaimers, or navigation text into the sample.

!
False positives are possible. Short, repetitive, or formulaic human writing can look suspicious to the model.
AI
False negatives are possible. Edited AI text can evade phrase rules and look less obviously synthetic.
← Back to detector