AI Text Detection Explained: 'AI Tells', Machine Learning and the Limits of These Tools

The trigger: The dispute centres on the 2026 Commonwealth Short Story Prize, one of the world's most global literary awards (open to writers across Commonwealth countries, with regional winners published by the magazine Granta). After the Caribbean winner's story appeared, a University of Pennsylvania professor ran it through the AI detector Pangram, which flagged it as overwhelmingly AI-generated, and other readers pointed to tell-tale stylistic patterns.

The dispute: The author denied any AI use, Granta said it was investigating, and the Commonwealth Foundation stood by its judges, noting that entrants must declare their work is their own and not AI-assisted. The episode became a flashpoint in a larger debate over so-called "AI slop" creeping into literature, journalism and research — and over whether detection tools can settle such questions at all.

Machine learning (ML): ML is the branch of AI in which a system learns patterns from data and statistics rather than from explicit rules. Large volumes of data are fed to a model so that it can recognise regularities and make predictions or classifications.

How detectors are built: To detect AI text, developers feed a model many examples of both human-written and AI-written content and train it to tell them apart. The model learns statistical patterns — in word choice, phrasing and punctuation — that distinguish the two, and then classifies new text as likely human or likely AI.

A key distinction: AI detectors are not the same as plagiarism checkers. A plagiarism checker compares text against a database of existing sources; an AI detector instead examines the internal statistical signature of the text itself.

Surface signals: Commonly cited "tells" include excessive use of em dashes, certain favoured words (such as "delve" or "imperative"), text organised in bullet points under headings, and conclusions that neatly wrap up without introducing new ideas (human conclusions often do). Another is "negative parallelism" — the formulaic "Not X, but Y" construction (for example, "not just hearing devices, but sound-cancelling devices").

Statistical signals: Beyond surface cues, classic detectors measure perplexity and burstiness. Perplexity is how "surprising" or unpredictable the text is — AI tends to produce low-perplexity (highly predictable) text. Burstiness is the variation in sentence length and complexity — human writing is "burstier" (mixing short and long sentences), while AI output is often more uniform. Some advanced approaches also use watermarking, where a hidden statistical signal is embedded at the moment of generation.

Pre-training vs post-training: A large language model (LLM) is first pre-trained on vast text to predict the next word, then post-trained (often via human feedback and curated examples) to make it safe, useful and able to follow instructions.

The likely source: Researchers say the answer is not settled, but a leading hypothesis is that the post-training data — answer examples written by hired annotators and data vendors for frontier labs — carry particular stylistic habits. Because models learn and replicate the writing style of these examples, the resulting "house style" (em dashes, tidy structure, parallelism) shows up across many AI systems.

The two error types: A false positive is when a detector flags human writing as AI; a false negative is when AI text passes as human. In high-stakes settings (academic integrity, prizes, jobs), a false positive can wrongly damage a real person, so most detectors are tuned to keep false positives low, even at the cost of more false negatives.

The reliability picture: A tool like Pangram advertises a false-positive rate of roughly 0.01% (1 in 10,000) and performs well in some independent tests. But, as experts note, no ML system is perfect — much as email spam filters, however good, occasionally misclassify messages. Reliability also varies widely across tools; notably, OpenAI withdrew its own AI-text classifier because of low accuracy.

Too little text: Detectors are more likely to err when there are few words, because short passages offer too few signals to judge confidently.

"Low-entropy" and structured text: Text that is precise and has essentially one correct form — a factual list (say, the states of India in alphabetical order) — is hard to classify, since a human and a machine would write it almost identically. Computer code is similarly tricky, as there are only limited ways to write a given instruction.

Editing and evasion: When a person lightly polishes their own writing with an LLM, detectors may still flag it as fully AI-generated, misrepresenting genuinely human work. Conversely, paraphrasing or heavy editing can help AI text evade detection, and watermarks can be stripped by editing or translation. Most detectors are also trained mainly on English, limiting reliability elsewhere.

High-stakes fairness: Because the costs of a wrong call are high, detector outputs should be treated as one signal, not proof, combined with human judgement and due process — wrongly branding a student or writer a cheat can cause real harm. The shift from "did they use AI?" to "how much assistance was used?" reflects this nuance.

Norms and governance: Publishing and academic experts argue for openness and disclosure standards — agreeing on where AI touches the writing process and tracking it transparently — rather than denial or a patchwork of hidden practices. This connects to wider debates on AI governance, authenticity and content provenance, including India's growing focus on responsible AI in education and the IndiaAI Mission, and to global efforts on content-provenance/watermarking standards.

The central lesson is that AI detectors should be used as decision-support, not verdicts — their outputs must be paired with human review, context and a fair process before any action is taken against a writer or student. Institutions (universities, journals, literary prizes and employers) need clear, published policies that distinguish prohibited AI use from permitted assistance and require disclosure of the degree of AI involvement, moving past the simplistic human-versus-AI binary. Continued research investment — into more robust detection, reliable watermarking and content-provenance systems, and multilingual tools — is essential, alongside safeguards against false accusations. Finally, building broad AI literacy among teachers, editors and the public will matter more than any single tool in preserving trust in writing.