Coaching matters more than size (2022)
Source OpenAI study of ChatGPT's predecessor 11.
Finding People preferred answers from a small AI coached with people's ratings over an AI more than 100 times bigger without that coaching.
Apply it If I can't retrain the AI, I can copy the behaviour coaching produces by writing clear rules and showing examples.
A written list of values can coach an AI (2022)
Source Anthropic's "Constitutional AI" study 12.
Finding Giving the AI a short list of principles and having it check its own answers against them improved its behaviour, with far less human effort.
Apply it Write my rules as principles with reasons. The same rules can be used to automatically review replies.
AI can learn from "better vs worse" pairs (2023)
Source Stanford study on a simpler training method, called DPO 13.
Finding Showing the AI pairs of answers, one better and one worse, teaches it well, and is simpler than older methods.
Apply it If I ever retrain, collect pairs of better and worse answers rather than just scores. Some AI companies offer this as a service 52.
AI tells people what they want to hear (2023)
Source Anthropic study of five AI assistants 14.
Finding All five changed correct answers when users pushed back. People's own ratings rewarded agreeable answers, which is how the habit formed.
Apply it Tell the AI to stand by correct answers, and always test what happens when a user wrongly disagrees.
Friendlier AI makes more mistakes (2025)
Source Oxford study, later published in Nature 15.
Finding Five AIs trained to be warmer got facts wrong 10 to 30% more often and agreed more with false beliefs. Just asking for warmth in the instructions had a smaller but similar effect.
Apply it Any change to tone, including my instruction file, must be tested for accuracy before release.
AI learns to write too much (2023)
Source Study of how AI coaching rewards length 16.
Finding Much of the "improvement" from coaching came simply from answers getting longer.
Apply it Give the AI clear length rules for each kind of question, and use length settings where available 41.
AI skims the middle of long texts (2024)
Source "Lost in the middle" study 20.
Finding AI used information best when it was at the start or end of a long text, and worst when it was in the middle.
Apply it Keep instructions short and near the start. Sum up long conversations instead of sending them in full.
AI struggles over many messages (2025)
Source Microsoft and Salesforce study of 15 AIs 21.
Finding When a task was spread over several messages, AI did 39% worse on average. It guessed too early and stuck with wrong guesses.
Apply it When a request is missing key details, ask or state assumptions first. Before a final answer, sum up what the user has asked for.
AI can be taught whose instructions come first (2024)
Source OpenAI study on ranking instructions 17.
Finding AI trained to put the company's instructions above the user's, and both above text in documents, was harder to trick.
Apply it Keep my rules in the AI's main instructions. Never let documents or emails give it orders.
Examples teach quickly (2020)
Source OpenAI's GPT-3 study 18; Anthropic's writing guide for AI instructions 36.
Finding AI picks up a pattern from just a few examples, without retraining.
Apply it Include a few good and bad example replies in the instruction file, varied so it doesn't copy one template.
Looking things up makes AI more accurate (2020)
Source Study of AI that searches documents before answering (called RAG) 19.
Finding Giving the AI relevant documents to read before answering made its answers more accurate and specific.
Apply it Let the AI look up the organisation's documents for anything about the organisation. Tell it to say when the documents don't answer the question.
AI makes things up because tests reward guessing (2025)
Source OpenAI study 27; Anthropic study on AI knowing what it knows 28.
Finding AI often has some sense of when it's unsure, but tests that give no credit for "I don't know" teach it to guess confidently.
Apply it In my tests, score an honest "I'm not sure, here's what I can confirm" above a confident wrong answer.
Asking back only when needed works best (2022)
Source Study on when AI should ask clarifying questions 29.
Finding Asking a question back only when a request was genuinely unclear improved accuracy, without annoying users by asking every time.
Apply it Ask only when different readings would lead to different answers. Otherwise answer and say what you assumed.
Good memory means saving notes and picking the right ones (2023)
Source Stanford's "Generative Agents" study 23; "MemGPT" study 22.
Finding AI stayed consistent over time when the app saved notes and brought back the most relevant, recent and important ones.
Apply it Build memory as short saved notes, picked for each message, with a way for users to see and delete them.