AI Detection Classifier
Definition
A machine-learning system trained to discriminate between human-written and LLM-generated text by measuring features such as perplexity, burstiness, and n-gram probabilities. Current classifiers are not reliable enough for evidential use.
- Task
- Discriminate human-written from LLM-generated text
- Features used
- Perplexity, burstiness, n-gram probabilities
- Current evidential status
- Not reliable enough for evidential use
- Failure modes
- False positives on human text, evasion by paraphrasing
Common questions
Why are these classifiers considered unreliable for evidence rather than just imperfect?+
They produce meaningful rates of both false positives, flagging genuine human writing as AI-generated, and false negatives, missing AI text that has been lightly edited or paraphrased, and courts require a much higher demonstrated reliability before a technique can support a factual finding.
Can a classifier's output be strengthened by combining it with other evidence?+
Yes, investigators typically treat a classifier flag as an investigative lead rather than proof, corroborating it with metadata, account activity logs, or stylometric comparison against known authentic writing samples from the same author.
Related terms
- Burstiness
- The variance in sentence-level complexity across a passage. Human writing tends to mix complex and simple sentences unevenly; LLM output tends to...
- Human-LLM Collaboration
- Text production in which a human contributes prompts, editing decisions, and intentional content while an LLM generates or transforms the prose. The...
- Large Language Model (LLM)
- A neural network trained on large text corpora to predict likely next tokens. At inference, it generates text by sampling from probability...
- Perplexity
- A measure of how surprising a sequence of words is to a language model. LLMs tend to generate low-perplexity text (predictable word...
- Stylometric Signature
- The set of measurable linguistic features : function-word frequencies, sentence length, punctuation patterns, vocabulary richness : that characterises a specific writer's output...