How to Tell if Text Was Written by AI: Methods, Limits and What to Do Next

You have a piece of text in front of you—an email, an essay, a job application, or a social media post—and something feels off. The writing is too smooth, too uniform, or just too perfect.

You want to know: was this written by AI? The honest answer is that no single method can give you a definitive verdict with 100% certainty.

But you can get very close by combining automated tools with your own human pattern recognition.

This article walks you through the three approaches to detection, the hidden limits of each, and what to do once you have a result.

Why Detection Is Harder Than It Sounds

AI language models like ChatGPT, Claude, and Gemini generate text by predicting the next most probable word based on patterns in their training data.

This means their output is statistically "average" in many ways—sentence lengths cluster around a mean, transition words appear at predictable intervals, and the overall tone stays neutral unless explicitly prompted.

Human writing, by contrast, is messy, rhythmic, and full of idiosyncrasies.

Detection tools try to measure how "surprising" a text is (perplexity) and how much its sentence structure varies (burstiness).

But because AI models are constantly improving and can be prompted to mimic human style, the line between human and machine writing is blurring fast.

The best approach is a layered one: start with a free automated checker, then verify with your own eyes, and finally consider the context in which the text was produced.

Three Approaches to Verification

Automated Detectors: What They Measure and Where They Fail

Free AI detectors like GPTZero, Originality.ai, and Copyleaks analyze text for statistical markers of machine generation. They typically return a percentage score and highlight suspicious sentences.

These tools are useful for a first pass, but they have well-documented blind spots:

• Short texts (under 50 words) produce unreliable results because the statistical sample is too small.
• ESL writing is often misclassified as AI because non-native sentence structures can appear "too regular" to the detector.
• Adversarial rewriting (using another AI to paraphrase AI text) can drop detection rates significantly.
• Model-specific gaps: no detector can keep up with every new version of every AI model.

Use a detector as a signal, not a verdict. A high AI score means “investigate further,” not “guilty.”

The 5 Red Flags You Can Spot Without a Tool

Before running any software, you can train your own eye to spot the most common tells of AI-generated text.

These manual inspection signals are especially valuable for short texts where detectors struggle.

1. Uniform sentence length. Humans naturally vary sentence length—some short, some long, some fragmented. AI tends to produce sentences of similar length, often 15–25 words, with very few very short or very long sentences.
2. Redundant transitions. Phrases like “in addition,” “furthermore,” “moreover,” and “consequently” appear at the start of nearly every paragraph. A human writer might use them once or twice and then switch to a subtle pivot like “But” or “Still.”
3. Vague specificity. AI loves to say “many experts agree” or “research shows” without naming the experts or the study. It gives the impression of evidence without committing to a real source.
4. Missing personal voice. Even in formal writing, humans leave fingerprints: a favorite metaphor, an occasional contraction, a slight opinion. AI text is generically polite and neutral unless explicitly prompted to be otherwise.
5. Unnatural paragraph structure. AI paragraphs often have a topic sentence, three supporting sentences, and a concluding sentence—every single time. Humans break this pattern, sometimes burying the main point in the middle or ending with a question.

Contextual Evidence: Metadata and Revision History

Sometimes the text itself is not enough. The circumstances around its creation can be more revealing than any pattern analysis. Look for:

• Document metadata (in Word or Google Docs): check the “last edited by” field, the creation date, and the total editing time. A document created in 30 seconds is suspicious.
• Revision history: a human writer typically makes many small edits, backtracks, and rephrases. An AI-generated document pasted in one go will show a single “paste” event.
• Typing pattern data: tools like GPTZero’s Origin Review can replay a video of the writing process, showing keystrokes and pauses. Real human writing has bursts and hesitations; AI writing appears as a smooth, constant flow.

How to Combine Methods for Different Scenarios

Not all texts require the same level of scrutiny. Use this framework to decide your approach:

When Detectors Lie: The Limits You Must Know

Even the best detectors—those claiming 99% accuracy—have failure modes that you need to understand before acting on a result.

• Short text unreliability: For a single sentence or two, detection accuracy drops below 50% for many tools. You cannot trust a “highly likely AI” flag on a tweet or subject line.
• Model-specific blind spots: A detector trained on GPT-3.5 may miss GPT-4o or Claude 3.5 Sonnet because the newer models produce more human-like perplexity scores.
• ESL bias: Non-native English writers often produce text with lower perplexity (less “surprise”) because they use simpler vocabulary and more straightforward sentence structures. Leading detectors like GPTZero have worked to reduce this bias, but many free tools still penalize ESL writers.
• Adversarial text: You can run AI-generated text through a second AI with a prompt like “rewrite this to sound more human.” This dramatically reduces detection rates. Some tools now claim to detect such rewrites, but the arms race is ongoing.

The Ethics of Detection: What to Do With a Verdict

An AI detection result is a starting point for a conversation, not a weapon. If you are a teacher, employer, or editor, follow these principles:

• Never punish based on a detector alone. False positives happen, and they can have serious consequences for students or job applicants. Always investigate further.
• Use detection as a teaching tool, not an accusation. Show the writer the flagged passages and ask them to explain their process. Many students who use AI to generate ideas, not final text, will have hybrid documents that confuse detectors.
• Document your evidence. If you confront someone, have the detector output, the manual inspection notes, and any contextual clues ready. A single screenshot from a free checker is not enough.
• Remember the limits of your own judgment. Even after reading this guide, you will occasionally be wrong. The goal is not to catch every AI-generated sentence—it’s to foster responsible use of AI while preserving trust in human writing.

Frequently Asked Questions

Can I detect AI in images or video as easily as text?

No. Image and video detection require different tools (e.g., metadata analysis, deepfake detectors). Text-based AI detectors are not applicable to visual media.

Does AI detection work on translated text?

Poorly. Most detectors are optimized for English prose. Translated text often has unnatural phrasing that can confuse both human and machine detection.

Can I tell if my own writing will be flagged as AI?

Yes. Run your text through a free detector before submitting it. If you get a high AI score, look for the manual red flags listed above—you may be writing in a style that mimics AI patterns.

Are there offline detection methods?

Few exist. Most detectors require cloud processing because they use large language models themselves for comparison.

Some basic statistical analysis (e.g., calculating perplexity locally) can be done with Python libraries, but it is not user-friendly.

How do detectors handle AI-generated code vs. text?

Code detection is less reliable. AI-written code often follows best practices perfectly, while human code has inconsistencies. However, detectors are not fine-tuned for code syntax.

What is the difference between perplexity and burstiness?

Perplexity measures how “surprised” a language model is by the text—lower perplexity means more predictable, which is typical of AI. Burstiness measures the variation in sentence length and structure.

Human writing generally has higher burstiness.

Can AI detection be fooled by editing?

Partially. Significant rewriting—changing sentence structure, adding personal anecdotes, varying word choice—can reduce detection rates. Minor edits like swapping synonyms have little effect.