AI-assisted writing.Start1 Why AI2 The issues3 What people say4 Levels of quality5 Tools now6 What is new7 What has changed8 What works9 Day to day10 How it was made
Page 7 · What has changed since 2023

Seven pieces of 2023 advice are now false, and six are still true

Drop the advice about memory, short inputs and "delve". Keep the fact-check and the distrust of detector scores.

What to do differently

  1. Stop the hunt for "delve" and em-dashes. Read for the sentence patterns.
  2. Use the long input. Give the model the brief, the research and three examples of the client's best work in one go.
  3. Keep the fact-check. That advice has not expired.
  4. Plan for fewer clicks from search, and write for the reader who does arrive.

Advice that is now false

The old advice What is true in 2026 Source
"It forgets everything between chats." ChatGPT has had memory since 2024 and can refer to all past conversations since April 2025. Claude's memory is a list you can edit. OpenAI; Claude release notes
"It can only read a few pages." The 2023 GPT-4 took about 8,000 tokens, roughly 12 pages. Each of the ten top-ranked models in October 2026 lists 1 million tokens, roughly 1,500 pages. LMArena leaderboard, 2 October 2026
"Look for 'delve' and 'tapestry'." Those words are now listed as typical of 2023 to mid-2024 output. "Delve" began to drop in 2024. Wikipedia: Signs of AI writing; Geng and Trotta
"The em-dash gives it away." OpenAI made the instruction to avoid it work in November 2025, and Wikipedia's guide notes that the sign seems less common in current output. TechCrunch; Wikipedia
"Detectors are a coin toss." The best commercial detector had near-zero error on ordinary prompted text of medium to long length. A score is still not proof. See the next table. Jabarian and Imas, August 2025
"Google treats AI content as spam." Since February 2023 Google's test is quality, "however it is produced". Google Search Central
"Rank first and the clicks follow." When an AI summary appears, users click a normal result on 8% of visits, against 15% without one. Pew Research Center, July 2025, 68,879 searches

Advice that is still true

The advice The current evidence Source
Check each citation and source. Fabricated citations ran from 11.4% to 56.8% across ten models. Newer models were not reliably better. Naser, preprint, 2026
Do not treat a detector score as proof. Text from a fine-tuned model was flagged 3% of the time. OpenAI says detectors are not reliable enough. Chakrabarty and others, preprint; OpenAI
The sentence patterns give it away. "Not X, but Y", the rule of three and closing "-ing" phrases are listed as current signs. Wikipedia: Signs of AI writing
Out of the box it does not sound like your writer. In a test of literary style, expert readers strongly disfavoured imitation from a prompt. They preferred the model only after it was fine-tuned on the author's work. Chakrabarty and others, preprint
Pages made in bulk for rankings get penalised. Google's policy on scaled content abuse applies however the pages are made. Google, March 2024
More persuasive does not mean more accurate. Where training made models more persuasive, their factual accuracy fell. Hackenburg and others, 2025, 76,977 participants

Public comfort with AI text has not risen

People accept AI text more easily than AI images or AI people

In the same Reuters Institute report, 55% were comfortable with AI that edits spelling and grammar, 41% with AI that writes a headline, 26% with AI that makes a realistic image and 19% with an artificial presenter.

In the US, 75% accept AI for brainstorming and early drafts of ads, and 38% accept AI-created people or voices (Gallup, surveyed May 2026, 3,270 adults).

This supports Anthony's opening point on the start page: readers mind AI writing less than AI images and voices.