AI Writing
AI writing is the use of language models to suggest, draft, rewrite, summarize or translate text, ranging from the next-word suggestion in a phone keyboard to a large language model producing a complete report, article, email or chapter from a short instruction. It sits on a continuum rather than being a single technology: at one end the software completes what a person is already typing, at the other the person supplies a brief and edits the machine's draft, and the practical, ethical and detection questions differ along that continuum.
The history of AI writing is a story of the unit of assistance growing from the word to the document. Predictive text and spelling correction on phones were statistical n-gram models trained on typing patterns; Google's Smart Compose, launched in Gmail in 2018, completed phrases inside emails using a neural language model. OpenAI's GPT-2 in 2019 showed that a transformer trained on web text could continue a paragraph coherently, enough that its full release was staged over months, and GPT-3 in 2020 made few-shot prompting a practical interface for writing tasks. The public release of ChatGPT in November 2022 added instruction following and conversation, which turned drafting into a dialogue: ask, revise, ask again. Tools for marketing copy, code documentation, customer replies, translation and academic editing followed, and word processors and email clients built the capability into their own interfaces.
Hallucination is the defining factual risk, because a language model optimizes for plausible continuation rather than verified truth. A fluent paragraph can contain invented quotes, nonexistent sources or confidently wrong numbers, and the polish of the prose gives no signal of which parts are accurate. The best-known case is Mata v. Avianca in 2023, in which lawyers in New York filed a brief citing court decisions that ChatGPT had fabricated, and the judge sanctioned them. In early 2023, the technology site CNET also corrected a batch of finance explainers produced with an AI tool after errors were found. Retrieval-augmented generation, which grounds the draft in supplied documents, and citations that a reader can click through reduce but do not remove the problem, so verification remains part of the job whenever accuracy matters.
Detection is much less reliable than people assume, and that matters most in schools and workplaces. Detectors typically estimate how predictable a text is under a language model, or use a classifier trained on human and machine samples, and both approaches break down when text is edited, paraphrased or produced by a newer model. OpenAI released its own AI text classifier in January 2023 and withdrew it in July 2023 citing its low accuracy. A 2023 study by Weixin Liang and colleagues at Stanford found that several commercial detectors disproportionately flagged essays by non-native English writers as machine-generated, since simpler, more predictable phrasing resembles model output. Watermarking, proposed for example by Kirchenbauer and colleagues in 2023, biases the choice of words in a hidden, statistically detectable pattern, but it only works when the provider applies it and can be weakened by rewriting. The practical consequence is that a detector score is weak evidence about a particular person and should not by itself decide misconduct cases.
Style is the quieter risk: unreviewed machine prose tends toward a recognizable, averaged register. Models favor common phrasing, balanced structures, safe hedges and generic summaries, so heavy reliance flattens a writer's voice and makes documents from different authors resemble each other. Prompting with samples of one's own writing helps, and editing remains the step that gives text particularity, such as specific examples, a point of view and the details only someone with direct knowledge would supply. Authors who write to think, such as students and researchers, can also lose the learning that drafting provides if the model does the composing.
The legitimate uses are broad and often unglamorous. Common applications include first drafts of routine correspondence, summaries of long documents, rephrasing for audience or reading level, translation, brainstorming of structure and headings, accessibility support for people with dyslexia or limited fluency in a language, and boilerplate documentation. Disclosure norms vary: many journals and publishers require that AI use be stated and that AI not be listed as an author, while some employers set their own rules. The misconception to avoid is that AI writing equals AI thinking; the model supplies fluent wording, while the claims, the judgment and the responsibility for what is published remain with the named author.
Read AI writing as drafting acceleration with a verification burden: the tool moves from autocomplete to full drafts, detectors are too unreliable to police it, and the quality of the result depends on how carefully a person checks facts and restores a voice of their own.