What types of content can I scan with Winston AI?
Overview
Winston AI's detection engine is trained on a wide range of written content, but not every type of text produces equally reliable results. Understanding what works well — and what doesn't — helps you get the most accurate scans and set the right expectations when interpreting results.
Minimum Requirements
Before scanning, your text must meet a minimum threshold:
Minimum 500 characters to run a scan
The more text you provide, the more accurate the results
For best results, aim for 300 words or more
Short text snippets (a few sentences) can technically be scanned, but the prediction confidence will be lower. The overall Human Score is more reliable than the sentence-level AI Prediction Map, especially on shorter passages.
Content Type Reference
Content Type | Works Well? | Notes |
|---|---|---|
Essays and academic papers | ✅ Best results | Long-form prose is ideal. The more complete the document, the more accurate the score. |
Blog posts and articles | ✅ Best results | Full articles with natural prose produce highly reliable detection. |
Cover letters and personal statements | ✅ Good | Works well. Short length may reduce confidence slightly. |
Email body content | ✅ Good | Full email body works well. Subject lines alone are too short. |
Product descriptions | ✅ Good | Works well when descriptive prose is present. |
News and press releases | ✅ Good | Standard journalistic prose scans reliably. |
Social media posts | ⚠️ Limited | Usually too short for reliable detection. Treat results as indicative only. |
Bullet-point lists | ⚠️ Limited | Minimal prose means less signal for the model. Works better when mixed with full sentences. |
Translated content | ⚠️ Limited | Translation alters writing patterns significantly. Results may be less reliable. |
Heavily edited AI drafts | ⚠️ Limited | Significant human editing can lower the AI score even on AI-generated originals. |
Source code | ❌ Not suitable | Code does not follow natural language patterns. Do not use the text scanner for code. |
Mathematical formulas | ❌ Not suitable | Symbolic and numeric content lacks the prose characteristics the model relies on. |
Boilerplate legal text | ❌ Not suitable | Highly standardized language that looks similar whether written by a human or AI. Results are unreliable. |
Tables and structured data | ❌ Not suitable | Non-prose content does not provide meaningful detection signal. |
Transcripts (verbatim speech) | ⚠️ Limited | Spoken language patterns differ significantly from written prose. Results may skew toward human-written even for AI-generated scripts. |
Factors That Affect Accuracy
Text Length
The single biggest factor is length. Longer texts give the model more signal to work with. A 50-word snippet may return an uncertain result; the same content expanded to 500 words will produce a much more reliable score. If you receive an uncertain result on a short document, consider scanning a longer version if available.
Paraphrasing and Humanization Tools
AI content that has been run through a paraphrasing tool or "AI humanizer" will often score higher on the Human Score than unmodified AI output. Winston AI is specifically trained to detect humanized content, but heavily rewritten text can reduce confidence. A moderately elevated Human Score on paraphrased content does not necessarily mean the text is human-written.
Mixed Authorship
Documents where a human has edited AI-generated content — or vice versa — produce mixed signals. The overall Human Score will reflect the blend. The AI Prediction Map's sentence-level breakdown is the best tool for identifying which specific sections are driving the AI score in these cases.
Tips for Best Results
Submit the full document rather than excerpts when possible
Remove headers, footnotes, and bibliographies before scanning — these can introduce noise
If scanning a translated document, note that translation alone can significantly alter AI detection scores
Use the sentence-level AI Prediction Map to investigate specific sections rather than relying solely on the overall score for mixed-authorship documents
When in doubt, use the Explain button for an AI-generated narrative that contextualizes the result