How to Check If Text Is AI Generated

Humans identify AI-generated text at roughly the same rate as a coin flip. A meta-analysis of six detection studies found that participants correctly distinguished AI writing from human writing just 53% of the time, where random guessing would yield 50%. Even when researchers trained participants on common AI patterns, accuracy barely moved past 59%.

This matters more now than it did a year ago. Open-source AI agents like Nous Research Hermes Agent can draft emails, write reports, and generate documentation autonomously. Commercial models from OpenAI, Anthropic, and Google have gotten better at mimicking natural human prose. The volume of AI-generated text online has grown faster than our ability to spot it.

The gap between human intuition and actual detection creates real problems. Educators need to evaluate student submissions. Editors need to verify that freelance writers deliver original work. Teams running AI agents need to know which outputs require human review before publishing. In every case, "it sounds like AI" is not a reliable signal on its own.

Detection tools help, but they bring their own limitations. False positive rates across commercial detectors range from 2% to 28%, depending on the tool and the type of text being scanned. That spread is wide enough to flag innocent human writing as AI-generated, or to miss AI text entirely.

This guide covers both sides: the manual signals you can train yourself to spot, and the detection tools that perform better than guessing. Neither approach works perfectly alone, but combining them gets you closer to a reliable answer.

Before reaching for a detection tool, you can catch a surprising amount of AI-generated text by reading carefully. These patterns show up consistently across GPT-4, Claude, Gemini, and open-source models like Hermes 3.

Uniform sentence length

This is the most reliable manual signal. Human writing naturally alternates between short punchy sentences and longer explanations. AI text tends to produce sentences of similar length throughout, creating a metronomic rhythm. Open any suspicious text in a word processor and eyeball the line lengths. If they all look roughly the same, that is a red flag.

Hedging language everywhere

AI models are trained to be helpful and non-committal, which produces a distinctive pattern: "It's worth noting that," "One might consider," "It's important to remember." Humans hedge too, but AI does it in nearly every paragraph, often in the same structural positions.

Vocabulary that nobody actually uses

Words like "delve," "underscore," "noteworthy," "pivotal," and "foster" appear far more often in AI output than in natural conversation or typical business writing. These words are not wrong, but their frequency is a signal. If a 500-word email uses "noteworthy" twice and "underscore" once, you are probably reading generated text.

Repetitive paragraph structure

AI models often follow a pattern: topic sentence, two or three supporting points, summary sentence. Every paragraph. Humans are messier. We start paragraphs in the middle of a thought, write one-sentence paragraphs for emphasis, and sometimes skip the summary entirely.

Perfect grammar with zero personality

AI text rarely contains typos, sentence fragments, or colloquialisms. Human writing usually has at least a few rough edges, whether that is a run-on sentence, a deliberately incomplete thought, or slang. If a piece reads like it was copy-edited three times but has no distinctive voice, that is worth investigating.

Generic openings and conclusions

Phrases like "In the rapidly evolving landscape of..." or "As we navigate the complexities of..." are AI signatures. So are conclusions that restate the introduction almost word-for-word. Humans usually open with something specific to their experience or perspective, not a sweeping generalization.

None of these signals is proof on its own. A careful writer might produce uniformly structured paragraphs, and someone who reads a lot of AI output might start using "delve" without realizing it. The value is in stacking signals: when three or four of these patterns appear together, the probability of AI generation is much higher.

AI detection tools analyze text using statistical models that measure how predictable the writing is. The two core metrics are perplexity and burstiness.

Perplexity measures how surprised a language model would be by the next word in a sequence. Human writing scores higher on perplexity because we make unexpected word choices, use idioms, commit errors, and write creatively. AI-generated text scores lower because language models choose statistically probable words at each step. A detection approach based on perplexity alone achieves an F1 score of about 0.81 in controlled tests.

Burstiness measures how much sentence complexity varies across a document. The formula is straightforward: B = (standard deviation of sentence lengths / mean sentence length) x 100. Human writing has high burstiness because we naturally mix short and long sentences. AI output tends toward low burstiness, with sentences clustering around similar lengths. When researchers combined perplexity and burstiness with other stylistic features, the detection F1 score jumped to 0.94.

Watermark detection is a third approach, but it only works when the AI provider embeds signals during generation. Google's SynthID watermarks Gemini outputs by biasing token probabilities in a pattern that is statistically detectable but invisible to readers. As of June 2026, OpenAI has not deployed public watermarking in ChatGPT, and Anthropic has not released watermarking for Claude. Open-source models like Hermes 3 have no watermarking at all. So watermark-based detection catches some Gemini text but misses most AI output in the wild.

Where tools fall short. Detection tools work best on raw, unedited AI output where the statistical signatures are strongest. Once a human edits the text, adds personal anecdotes, or restructures paragraphs, detection accuracy drops 20% to 60%. Paraphrasing tools reduce it further. No detector is reliable enough to serve as a sole decision-maker for high-stakes judgments like academic misconduct charges or freelancer termination.

Audit every file your AI agents create

Fast.io workspaces log every agent action with full attribution. 50GB free storage, no credit card, and an MCP endpoint that connects to Hermes Agent and other AI tools.

Sign Up Free

This process combines manual review with tool-based analysis. No single step is conclusive, but following the full sequence gives you the highest confidence.

1. Read the text aloud and listen for rhythm. AI text often has a smooth, even cadence without natural variation. If every sentence feels the same length and weight, mark that as your first signal.

2. Scan for hedging patterns. Search the text for phrases like "it's worth noting," "one might argue," and "it's important to consider." More than two hedging phrases per 500 words is unusual for human writing.

3. Check sentence length variation. Copy the text into a word processor or text editor that shows word counts per sentence. If most sentences fall within a narrow range, say 15 to 25 words each, that low burstiness suggests AI generation. Human writing typically shows a wider spread, with some sentences under 10 words and others past 30.

4. Look for AI vocabulary markers. Search for words that AI models overuse: "delve," "underscore," "noteworthy," "pivotal," "landscape," "foster." One or two is fine. A cluster of them in a short text is suspicious.

5. Run the text through a detection tool. Paste the text into GPTZero, which provides sentence-level highlighting showing exactly which sentences triggered detection. Independent benchmarks place its accuracy between 87% and 88% on standard AI text. For a second opinion without creating an account, use QuillBot's AI Detector, which categorizes text into four levels from fully human to fully AI-generated.

6. Cross-check with a second detector. If the first tool flags the text, run it through at least one more. Grammarly's AI Detector and Scribbr's detector are both free options. Agreement between two independent tools increases your confidence. Disagreement tells you the text sits in a gray zone where you need more context before making a call.

7. Evaluate the results with context. Consider who wrote the text, why, and whether the stakes justify the false positive risk. GPTZero's published false positive rate is 0.24% on controlled benchmarks, but real-world university testing found 15% of human essays incorrectly flagged. Short texts under 500 words and non-native English writing produce the most false positives across all tools. If you are evaluating student work or freelance submissions, a positive detection result should start a conversation, not end one.

The goal is not binary classification. It is building a weight of evidence that helps you make a judgment call. When manual signals and tool results both point the same direction, you can be reasonably confident. When they disagree, more investigation is needed.

When you use AI agents to generate content, the question shifts from "Is this AI?" to "Is this good enough to publish under our name?" Detection becomes part of a quality assurance workflow rather than a forensic exercise.

AI agents like Nous Research Hermes Agent, Claude, and GPT-based tools can draft blog posts, client emails, documentation, and reports. The output is always AI-generated, but the real question is whether it reads like AI output to your audience. Running your agent's drafts through the detection process above tells you which sections need human editing before publication. Sections with low burstiness or heavy hedging language are the ones to rewrite first.

A practical agent workflow looks like this: the agent generates a draft, a human reviewer runs the manual signal check and a tool scan, and flagged sections get rewritten. For teams producing content at scale, this process needs shared infrastructure. You need a workspace where agents deposit drafts, humans review them, and the final versions are tracked with clear attribution.

Fast.io handles this kind of handoff. AI agents connect through the MCP server or API to upload drafts into shared workspaces. Reviewers access the same workspace to read, annotate, and approve content. Every action is logged in an audit trail, so you always know which version was generated by the agent and which was edited by a human. When the reviewed content is ready, ownership transfers from the agent to the team.

This pattern applies beyond content creation. Any team using AI agents for document drafting, report generation, or client communication benefits from separating "what the agent wrote" from "what we published." The audit trail matters for accountability, and shared workspaces prevent drafts from sitting in someone's local files where nobody else can review them.

The free agent tier includes 50GB of storage, 5,000 AI credits per month, and 5 workspaces with no credit card required.