AI DETECTOR · WRITING · PHISHING
How to detect AI-generated text: a guide to stylistic clues
Generative AI helps people write and can also help fraudsters produce polished, targeted phishing messages. Error-free writing is not a guarantee of safety. Conversely, using a writing assistant does not make content malicious.
Learn how to use an AI text detector to review stylistic clues, then independently check the facts and the request.
1. Examine structure: perplexity and sentence variation
These concepts are often confused. Only the measurements shown in the report have actually been performed.
Perplexity: a concept, not a measurement in this tool
Perplexity describes how predictable a word sequence is to a particular language model. It varies with the model, language and genre. Low perplexity does not prove AI authorship: formal or predictable human writing may appear similar. IA Detector does not calculate model-based perplexity.
Sentence variation: a measured signal
The tool measures the spread of sentence lengths relative to their average, a form of variation sometimes described as burstiness. It also reviews paragraph-length regularity. Uniform style can equally come from standardised human writing, editing or a template.
2. Review writing patterns and verify references
Generic phrasing and repetition
IA Detector flags certain formulaic expressions, the frequency of transition words, repeated three-word groups and similar sentence openings. It also measures lexical diversity and paragraph structure. These clues describe writing style without identifying software, author or intent.
Verify references and facts separately
Generated writing may cite nonexistent sources, but humans can also make mistakes or fabricate references. Check laws, numbers, links and names against reliable sources. The stylistic tool does not automatically fact-check claims or assess link safety.
Help your teams assess suspicious correspondence
With SecurCheck Business, HR, communications and risk teams can use stylistic analysis and a suspicious-text assessment to pause before acting on an urgent request. Sentence variation, repetition and vocabulary provide clues, never a verdict on origin or fraud.
Explore our business offer for your teams3. Understand the limits before drawing conclusions
Allow for false positives
Short messages, legal writing, technical reports and heavily standardised correspondence can resemble patterns flagged by the detector. The report expresses lower confidence with limited samples. Do not attribute authorship or make disciplinary decisions on the score alone.
Account for editing and rewriting
Proofreading and writing assistants can make human writing more uniform; AI writing may be extensively rewritten. No stylistic signal reliably traces these changes. If appropriate, compare writing from similar contexts under your applicable rules.
Check the message and confirm the request
Assess a suspicious email by its sender, domain, links, attachments and requested action. SPF, DKIM and DMARC offer authentication signals from the original message but do not prove the sender’s honesty. For payments, access or sensitive data, contact the person through a known independent channel.
Question polished writing methodically
Polished writing guarantees neither the sender’s identity nor their honesty. A writing-pattern assessment can create space to reflect if you also examine the request, its evidence and context. Always independently confirm a message that asks for access or payment.