How to Check if Text Was Written by AI

Step 1: Prepare the Text for Analysis

Before running any detection tool, make sure the text you are analyzing meets the minimum requirements for reliable detection. Most tools need at least 250 to 300 words of continuous prose to produce a meaningful result. Shorter passages, such as a single paragraph or a brief email, do not contain enough statistical signal for accurate classification.

Copy the full text into a plain text format, removing images, tables, and complex formatting. Most detection tools accept pasted text through a web form, and formatting artifacts can occasionally interfere with the analysis. If the document is a PDF or Word file, copy the text content rather than uploading the file directly, unless the tool specifically supports file uploads with reliable text extraction.

If the document is longer than the tool's character limit (typically 5,000 to 15,000 characters for free tiers), you have two options. You can paste sections individually and track the scores for each, or you can use a paid plan that accepts longer documents. When scanning sections individually, choose passages that represent different parts of the document rather than just the introduction, since AI generation patterns may vary across sections if the document has mixed authorship.

Step 2: Run the Text Through an AI Detector

Choose a detection tool appropriate for your use case. GPTZero offers the best balance of accuracy and transparency for general use. Originality.ai is the strongest choice for content publishers who also need plagiarism checking. Turnitin is the default for educators at subscribing institutions. For a quick, no-signup check, ZeroGPT provides instant results without requiring an account.

Paste the text into the tool and submit it for analysis. The scan typically completes within 10 to 30 seconds. Review the overall score first: most tools express results as a percentage indicating how much of the text appears AI-generated. Then examine the sentence-level or paragraph-level highlights, which show specifically where the detection signals are strongest. Pay particular attention to whether the AI signals cluster in specific sections or are distributed evenly throughout the document, as this can indicate whether the entire text was AI-generated or only certain portions were.

Record the overall score and note which sections were highlighted. A single detection result is informative but not sufficient on its own, since every tool has blind spots and biases that can produce misleading results.

Step 3: Cross-Check With a Second Tool

Run the same text through a different detection tool and compare the results. If both tools agree that the text is likely AI-generated (or likely human-written), the combined confidence is higher than either tool's result alone. If the tools disagree significantly, the text likely falls in an ambiguous zone that requires manual analysis to resolve.

Disagreement between tools is common, especially for text that has been lightly edited, paraphrased, or written by a non-native English speaker. Each detector uses different models and training data, so they react differently to the same input. A text that scores 90% AI on GPTZero but 40% on Originality.ai is genuinely ambiguous, and neither score should be treated as definitive.

For high-stakes decisions, consider using three tools rather than two. The consensus of three independent detectors provides stronger evidence than any individual result. However, even unanimous agreement among tools does not constitute proof, since all current detectors share similar underlying methodologies and can make the same systematic errors on certain types of text.

Step 4: Examine the Text Manually for AI Patterns

Automated tools analyze statistical patterns, but human readers can identify qualitative characteristics that algorithms miss. Read the text carefully and look for these common indicators of AI generation:

Uniform sentence length and structure. AI-generated text tends to produce sentences of similar length and complexity. Human writers naturally vary their sentence structure, mixing short punchy statements with longer, more complex constructions. If every sentence in a passage feels roughly the same weight and rhythm, that uniformity is a potential signal.

Generic vocabulary and hedging language. AI models favor common, broadly applicable words over specific, concrete terms. Phrases like "it is important to note," "there are several factors to consider," and "this can vary depending on" appear frequently in AI output because they are statistically safe choices. Human experts on a topic typically use more precise terminology and make more definitive statements.

Repetitive transition patterns. AI text often cycles through the same set of transitional phrases: "furthermore," "additionally," "moreover," "in addition." Human writers use a wider variety of transitions, and many experienced writers avoid explicit transitions entirely, relying on logical flow between ideas instead.

Absence of personal voice, anecdotes, or specific examples. AI models cannot draw on personal experience, so their output lacks the first-person anecdotes, specific real-world examples, and opinionated judgments that characterize human writing. A blog post about cooking that mentions no specific meals the author has cooked, a product review that describes no specific testing, or an essay that offers no original thesis beyond summarizing existing perspectives may have been generated rather than written.

Correct but shallow content. AI-generated text is often factually plausible but lacks depth. It covers the surface of a topic competently without offering insights that require genuine expertise or experience. If a 2,000-word article on a complex topic says nothing that would surprise someone already familiar with the subject, that flatness can indicate AI origin.

Step 5: Check Document Metadata and Context

If you have access to the document file (not just the text), check its metadata. Word documents and PDFs store creation dates, modification timestamps, author names, and editing time. A 3,000-word document with a total editing time of 2 minutes was not typed by a human. A document whose creation timestamp matches the assignment deadline to the minute, with no revision history, is worth investigating further.

Compare the text against other known writing samples from the same author. If a student who typically writes at a C+ level suddenly submits a polished A-grade essay with sophisticated vocabulary and flawless paragraph transitions, the quality shift itself is a meaningful signal, regardless of what any detection tool says. Similarly, if a freelance writer's latest submission has a dramatically different voice and style from their portfolio samples, that inconsistency warrants a conversation.

Check whether specific claims, statistics, or citations in the text are accurate. AI models sometimes fabricate citations, invent statistics, or attribute quotes to the wrong sources. Verifying a few key facts in the document can reveal whether the author actually researched the topic or relied on an AI model's plausible but potentially inaccurate output.

Step 6: Combine All Evidence and Make a Judgment

No single piece of evidence, whether from an automated tool or manual analysis, should be treated as conclusive. Instead, weigh all the signals together. Strong detection scores from multiple tools, combined with manual observations of uniform sentence structure and generic language, combined with metadata anomalies and quality inconsistencies, create a compelling case. A single elevated detection score with no supporting evidence does not.

If you are in a position to do so, the most reliable approach is to discuss the text with the person who submitted it. Ask them to explain their writing process, describe the research they conducted, summarize the key arguments in their own words, or discuss specific passages in detail. A person who genuinely wrote a document can discuss it fluently. A person who submitted AI-generated text often cannot explain their own reasoning, elaborate on specific points, or identify the sources behind their claims.