How to Check If Text Is AI-Generated

Start with the data point that reframes everything: according to independent benchmarking by AIDetectors.io (2026), when AI text is processed through a quality humanizer tool before submission, detection rates drop to 2–8% across all major platforms — including Originality.ai, which achieves 94% accuracy on unmodified AI text. That gap defines the challenge. Automated detection works well under one condition (unmodified AI output) and poorly under another (post-processing). Understanding this limitation shapes every practical decision about how to check for AI-generated text.

This guide covers what automated detectors actually measure, how to read and interpret their outputs, which tools perform best under which conditions, the manual signals that automation misses, and the workflow that produces reliable conclusions rather than false confidence in either direction.

What Automated Detectors Actually Measure

Every AI text detector — regardless of brand or price point — builds on the same two foundational signals. The marketing differs; the mechanism does not.

Perplexity: How Predictable Are the Word Choices?

Language models generate text by selecting, at each position, the statistically most probable next token given everything that preceded it. This produces text that is statistically expected — low-perplexity in formal terms. Every word in an AI-generated sentence is, by design, close to what any reader familiar with the training corpus would predict.

Human writing is different. People reach for less expected synonyms, shift registers mid-sentence, make choices influenced by personal experience that a language model has not modeled. A human expert might write “the data pointed somewhere uncomfortable” where a model would write “the data indicated significant findings.” The human sentence is higher-perplexity — less statistically predictable — and detectors flag low-perplexity text as a potential AI signal.

Per GPTZero's public methodology documentation, perplexity is one of their two primary signals. It is measured using a reference language model — typically one trained on a broad text corpus — to score how probable each word sequence is. The lower the average perplexity across a document, the higher the AI probability score.

Burstiness: Does the Sentence Structure Vary?

Burstiness measures variance in sentence length and structural complexity across a document. Human writing is highly bursty: writers naturally alternate between short, punchy sentences and long, complex ones. A paragraph written by a person often contains a three-word sentence followed by a forty-word sentence with three subordinate clauses. The variation is rhythmically natural and statistically distinctive.

AI-generated text is suspiciously uniform. A language model operating at a consistent temperature produces sentences of similar length, similar syntactic complexity, and similar information density throughout. Every paragraph feels equally weighted. Detectors operationalize this observation as a burstiness score — documents with high variance in sentence length read as human; documents with low variance read as AI.

Advanced classifiers — particularly Turnitin's AIW-2 model and Originality.ai's ensemble — also analyze discourse marker patterns (how transitions and connective language are used), argument structure coherence, and paragraph-level organization. These secondary signals extend detection beyond the sentence level to the document level.

AI Detector Accuracy Benchmarks: 2026

The numbers reveal two things. First, top-tier tools are genuinely effective against unmodified AI text. Second, the entire accuracy landscape collapses against humanized content. Originality.ai — the strongest performer in most benchmarks — catches only 7.8% of AI text that has been through a humanizer. GPTZero's 18% is better, but still means roughly four out of five humanized AI documents pass undetected.

This is not a flaw in the tools. It is a structural limitation: when AI output is edited to have higher perplexity and burstiness, it genuinely looks more human on the signals detectors measure. The solution is not a better detector — it is a more comprehensive detection approach that combines statistical tools with manual analysis and human verification.

Step-by-Step: How to Check Text for AI Generation

Step 1: Run a Multi-Tool Statistical Check

Never rely on a single detector. Different tools weight signals differently and use different training data. A text that scores 85% AI on GPTZero may score 40% on ZeroGPT because their classifiers are built on different corpora. Running the same text through two or three detectors and looking for concordance — multiple tools flagging the same passages — is more reliable than trusting any single result.

Specifically: run the text through EyeSift's sentence-level detector to identify which specific passages drive the overall score. Sentence-level granularity tells you where the AI pattern is concentrated — which is actionable information whether you are editing the text or investigating the authorship.

Step 2: Check the False Positive Context

Before treating a positive detection result as evidence of AI authorship, consider the demographic and stylistic context. The Stanford HAI study (Liang et al., 2023) is the definitive research here: non-native English speakers are misclassified as AI at a 61.22% rate across seven detectors. Native English students in the same study faced less-than-10% false positive rates.

Formal academic writing, technical documentation, and legal prose from experienced human writers also systematically triggers false positives — because expertise in formal writing produces low perplexity and low burstiness. If the text is from someone who writes formally by training or culture, a positive detection result requires additional investigation before any conclusion.

Step 3: Apply Manual Detection Cues

Manual detection targets content-level signals that statistical classifiers cannot measure. These cues survive humanizer processing because they reflect what the text says, not how it is structured.

Unverifiable or suspicious citations. AI models hallucinate. They generate plausible-sounding citations — journal names, author names, study titles — that do not exist. Look up every cited source. A paper that cannot be found in Google Scholar, PubMed, or JSTOR is a strong AI signal. This is the most reliable manual detection method because it identifies something a human author would not do: fabricate their own bibliography.

Discourse marker density. Words like “furthermore,” “moreover,” “it is worth noting,” “delve,” “leverage,” “paramount,” and “seamlessly” are dramatically overrepresented in AI-generated text relative to authentic human prose. Per Pangram Labs' 2025 analysis of detection failure modes, the co-occurrence pattern of these markers in a single document is a strong AI signal even after humanizer processing, because humanizer tools do not consistently remove all discourse markers.

Over-resolution of arguments. AI writing almost never leaves a genuine loose end. Every paragraph begins with a topic sentence, makes a claim, provides evidence, and resolves the point before moving on. Human writing — especially expert human writing — is structurally messier. It revisits ideas, leaves qualifications unresolved, or acknowledges complexity it cannot fully address. Text that is suspiciously clean in its argumentation structure is worth flagging for further investigation.

Absence of specific personal or professional experience. Human writers, particularly in professional and academic contexts, tend to anchor claims in specific experiences: a particular project, a named colleague, a remembered case. AI models produce universalizing, impersonal prose: “organizations typically find that...” “researchers generally agree...” “most professionals report...” The absence of any specific, personal, unverifiable-but-plausible detail is a content-level AI signal.

Step 4: Use Targeted Follow-Up Questions

For educational and professional contexts, the most reliable verification method is direct: ask the author questions about their work. Not questions that can be answered by rereading the text, but questions that require genuine understanding:

• Can you explain the methodology of the study you cited in section two?
• What made you structure the argument this particular way — what alternatives did you consider?
• How did you find the data in the third paragraph, and what was the original source?
• What is your opinion on the counterargument — why did you address it the way you did?

Human authors who genuinely engaged with their work can answer these questions in their own voice, with personal reflection. AI-generated content submitted without meaningful human engagement cannot produce answers to questions requiring actual knowledge of process, source, and reasoning. This is why the “oral defense” model — asking students or applicants to discuss their submitted work — is increasingly considered the gold standard for AI authorship verification.

Model-Specific Detection Rates: ChatGPT vs. Claude vs. Gemini

Detection rates vary meaningfully by the underlying model that generated the text, which matters for educators and publishers targeting specific tools.

Per 2026 benchmarking, ChatGPT (GPT-4o and successors) generates the most detectable output — with detection rates of 92–96% on leading detectors. ChatGPT's training optimizes for coherent, expected prose at default temperature settings, which produces reliably low perplexity. Claude models generate text that is somewhat harder to detect (75–85% detection rate), partly because Anthropic's Constitutional AI training produces stylistically different output from RLHF-trained OpenAI models. Gemini shows the most variability, with detection rates ranging from 70–82% depending on model version and prompt structure.

These detection rate differences matter less than they might appear, because they all collapse against humanizer processing. The practical implication is that model-specific detection is more valuable as a baseline than as an ongoing strategy — once humanization is involved, model attribution becomes speculative.

When Detection Results Are and Are Not Actionable

The most important contextual judgment in AI detection is determining what a result can and cannot support. Being clear about this protects both the people investigating and the people being investigated.

Detection results can appropriately support: initiating a follow-up conversation, requesting additional documentation, applying heightened editorial scrutiny, requesting an oral explanation of submitted work, or flagging text for secondary human review.

Detection results cannot appropriately support: determining guilt or misconduct as a standalone finding, rejecting job applications without human review, failing students without additional evidence, or making legal claims about authorship. Turnitin's documentation explicitly states that AI detection scores should not be used as sole evidence in academic integrity proceedings. The same principle applies in all contexts.

This is not a limitation of the technology that will be solved by a better tool. It is an inherent property of probabilistic classification: a classifier that achieves 94% accuracy produces 6% false positives, and in a context with many legitimate human authors, the false positives matter. The false positive problem is well-documented and should be part of every detection workflow's design.

Practical Tool Selection by Context

For educators and academic institutions: Turnitin remains the standard for formal academic integrity processes because of its institutional integration, its well-documented model architecture, and its explicit guidance against using scores as standalone evidence. GPTZero is a strong supplementary tool for sentence-level granularity. Neither should be used without a human review process for contested results.

For publishers and content teams: Originality.ai delivers the best accuracy-to-cost ratio for bulk content review, with multi-document scanning and plagiarism integration. It is less suitable in ESL author contexts due to known false positive bias. The full publisher guide to AI detection tools covers workflow integration in more detail.

For HR and recruiting: AI detection of application materials should be treated as a flag for follow-up, not a disqualification criterion. The SHRM State of AI in HR 2026 report found AI adoption in HR functions at 43% — the irony being that HR departments increasingly use AI to process applications while also questioning AI use in applications. A candidate who used AI to polish a cover letter that accurately represents their experience is different from a candidate who fabricated qualifications. That distinction requires human judgment to make.

Frequently Asked Questions

The Practical Conclusion

Checking whether text is AI-generated is a workflow problem, not a tool problem. The tools are useful — genuinely so, for unmodified AI text — but they are neither sufficient alone nor reliable in all contexts. The workflow that produces defensible conclusions combines automated detection (for baseline statistical signals), citation verification (for the content-level signals automation misses), and human conversation (for the authorship verification that no algorithm can replicate).

Treat detection results as the start of an investigation, not the end. Use the sentence-level highlights to focus your manual analysis. And apply the false-positive context check — particularly for non-native speakers, formal writers, and technical documents — before drawing conclusions from any single score.