Using an AI assistant well

A clear request

An AI assistant, the chat service a lawyer types into, runs on a model, the program that writes each answer by predicting, again and again, the word most likely to come next. The model learned to make those predictions in training, a process in which it was shown vast amounts of text and adjusted, a little at a time, so that its predictions came closer to that text. The assistant therefore works only from the request in front of it and from what training gave it: it has not read the file, met the client or seen the earlier drafts, so it fills every gap in a request with a guess. The guesses read as fluently as the rest of the answer, for the reasons set out in How a chatbot writes an answer, a Beginner module, so a gap in the request can pass into the finished work unnoticed. The request a user types into an assistant is called a prompt, and a prompt that supplies what the assistant would otherwise guess leaves it fewer gaps to fill.

The four largest vendors of these assistants, OpenAI for ChatGPT, Google for Gemini, Microsoft for Copilot and Anthropic for Claude, each publish guidance on writing prompts, and their advice on what a good prompt contains overlaps. OpenAI’s guide to prompting in ChatGPT names the goal, the context, the output and the boundaries. By boundaries OpenAI means the limits on the work: what ChatGPT is to leave unchanged, such as a passage or a figure the user does not want altered, and what it should avoid or check with the user before it acts. Google’s help for Gemini in Workspace, which covers Google’s email and office software, names a persona, meaning the role Gemini is asked to play, as well as the task, the context and the format. Microsoft’s help for Copilot names the goal, the context, the source, which is the material Copilot should use, and the expectations, which are how it should respond. Anthropic’s guide asks for clear, explicit instructions, and suggests showing the prompt to a colleague with little context on the task to see whether they could follow it.

Read together, the four vendors’ guidance describes a clear request as a task, the context it sits in and the form of the answer wanted, and each item in the vendors’ lists belongs to one of those parts. The task, which OpenAI and Microsoft call the goal, says what the assistant should produce, with a verb such as summarise, compare, list or draft, and OpenAI’s boundaries set limits on it. The context gives what a new colleague would need to know: who the work is for, what it will be used for and what has happened so far. Microsoft’s source belongs to the context too, since a new colleague would need to know which material to work from, and Google itself says that a persona gives Gemini context for how to behave and respond. The form, which OpenAI calls the output and Microsoft the expectations, describes the answer wanted: its length, its headings, whether it is a table or a letter, and how it should refer to the material it relies on.

An example of the answer wanted shows tone and structure faster than a description of them. Giving a model one or more worked examples inside the prompt is called few-shot prompting. Anthropic’s, Google’s and Microsoft’s guidance for developers, the programmers who build products on these models, each recommend few-shot prompting, although they differ on how many examples to give. The developer guidance holds for a lawyer’s request too, since a message typed into a chat is a prompt to the same models.

The vendors also differ on where a long document such as a judgment belongs in the prompt. Anthropic and Google advise putting such a document first and the instructions or questions about it at the end, while Microsoft’s guidance puts the task first, with the files and other sources the assistant should use after it. A label above each part, such as “Judgment” above pasted text, is therefore a safer habit than relying on order, since the label tells the assistant what each part is wherever it sits.

Large tasks in steps

A large task asked for in one request, such as reviewing a set of leases, drafting a note of the results and writing the covering email, gives the assistant many places to go wrong and leaves the reader one long answer to check. Each of the four vendors suggests splitting such work into smaller requests: Google’s Workspace help advises breaking related tasks into separate prompts, and Microsoft’s developer guidance says that language models often perform better if the task is broken down into smaller steps.

Anthropic and OpenAI apply the same idea of separate requests to a single piece of work. Anthropic’s guide describes a pattern it calls self-correction, in which the model writes a draft in one request, reviews it against criteria the user sets out in a second, and improves it in a third. OpenAI’s prompting guide suggests asking ChatGPT to set out its plan first whenever the way of doing the work could change the result. The guide also suggests starting with a single piece of output, such as one draft, which the user reviews before refining the instructions and asking for the rest.

For a lawyer, the value of working in steps lies in the checks between them. In the lease review, a list of the relevant clauses can be checked against the leases before a note is written on them, so an error caught at that stage stays out of every later step. A common prompting tip, asking a model to think step by step, makes the model write out its working within a single answer. Reasoning models, built to work through intermediate steps on their own before they answer, are described in Does AI reason?, a Beginner module, and OpenAI’s guidance for those models says that asking them to think step by step is unnecessary. Splitting the work into separate requests serves a different purpose, which is to let the reader see and check each stage.

Sources and uncertainty

A model asked something beyond what it knows tends to answer anyway, and the invented answers that result are what the vendors call hallucinations. Training may be the cause, as researchers at OpenAI and the Georgia Institute of Technology argue in a 2025 paper. On their account, a later part of training scores a model’s answers and adjusts the model towards those that score well, and developers also tune models to do well in the tests used to compare them. Both kinds of scoring give credit for a right answer and withhold it for an admission of doubt, so a model learns that a confident guess pays better than admitting it does not know. A request can still push against the habit of guessing, since a model also follows the instructions it is given, and an instruction that uncertainty is welcome makes admitting doubt part of the task. Anthropic’s guide to reducing hallucinations suggests explicitly permitting the model to admit it does not know, and OpenAI’s prompting guide suggests telling ChatGPT to use only the sources supplied and to flag missing information.

Keeping what a source says apart from the assistant’s own reasoning lets the reader tell which statements come from the material and which the assistant supplied. For long documents, Anthropic suggests asking first for word-for-word quotations, each numbered, and then for an analysis based only on those quotations that refers to each one by its number. Anthropic’s worked example of the technique is a review of a privacy policy against the GDPR. OpenAI’s guide gives a sample request, for turning quarterly reports into a brief, that asks ChatGPT to keep the facts the reports state distinct from its own analysis and to tie each number to the file it came from. Google’s developer guidance suggests confining the answer to the material supplied and saying when the answer is not in that material, and Microsoft suggests pointing Copilot at a specific file and asking it to show the exact section relied on.

Quotations give the reader something to test, since each one can be found in the material supplied or shown to be absent from it. Grounding, the vendors’ word for basing an answer on material supplied with the request or found by a search, reduces invented answers without removing them. Anthropic says that the techniques in its guide significantly reduce hallucinations but “don’t eliminate them entirely”. Microsoft’s description of Copilot for organisations says that, in summarising content from various sources, it may include information that is not present in those sources. Google’s help for Gemini says that Gemini can hallucinate and present inaccurate information as factual.

The Divisional Court commented on the practice of telling a model to admit uncertainty in a judgment on lawyers who had cited cases that do not exist. In R (Ayinde) v London Borough of Haringey it said that it did not consider the risks materially reduced by “asking the tool not to provide an answer if it is not sure”, and that the safeguard is “to check any output by reference to an authoritative source”.

Checking what it cites

Asking a model whether its own citations are real gives little protection, since the confirmation is a second answer from the same model, written in the same way as the first. A model that invented a case can therefore describe it confidently when asked whether it exists, because its reply, like the invented citation, is built from the words most likely to come next. Anthropic’s help centre warns that Claude can display quotes that look authoritative while not being grounded in fact, a phrase it uses in the everyday sense of resting on fact, and that users should not rely on Claude as a singular source of truth.

Assistants that search the web or a set of documents show where their answer came from. Claude’s web search puts citations and source links in its answers, and ChatGPT shows search results and citations in the chat. Gemini shows a Sources button at the foot of an answer or links within it, and Copilot’s Sources button shows the exact query it sent to Bing, Microsoft’s search engine, as well as the sources it used.

A citation shows where a statement may have come from, and the vendors themselves say it falls short of proving that the source supports the statement. Google’s help on sources in Workspace says that Gemini might cite a document that was not directly used for a specific piece of information, can sometimes make up a source, and may base its answer on only some of the content where there are too many sources. Microsoft’s transparency note for Copilot says that even answers drawn from high-authority web data might misrepresent that content, where high-authority appears to mean pages ranked highly in a web search, since the note describes Copilot as centring its answers on such pages.

The check that counts is a reader opening each source, finding the passage cited and reading whether the passage supports what the answer claims. For a case or a statute, that means an authoritative source, which the Divisional Court in Ayinde listed as the government’s legislation database, the National Archives’ database of judgments, the official Law Reports and the databases of reputable legal publishers. The facts of Ayinde and the court’s reasoning on checking sources are worked through in Does AI reason?, a Beginner module.

Saved instructions and memory

Most assistants now keep something from one conversation to the next, either as a package of instructions the user saves or as details from past conversations that the assistant keeps. Each saved package holds standing instructions, and often files, that the user applies whenever similar work comes up. Anthropic and OpenAI let users group chats, files and standing instructions into a project, a workspace with its own chats and files. Anthropic says that a project’s instructions tailor Claude’s responses within it, and OpenAI’s help on projects says that a project’s instructions apply across all the chats it holds. Anthropic’s skills are folders of instructions, scripts and resources that Claude loads for particular tasks, the scripts being optional pieces of code that Claude can run. Google’s Gems hold a name, instructions and optional files, and Microsoft’s agents in Copilot hold instructions and knowledge sources, such as uploaded files, documents in SharePoint or OneDrive, Outlook email and public websites.

Some of these packages are made to be shared, and sharing one shares what is inside it. Google’s help on sharing a Gem says that its instructions and uploaded files can be viewed by every user with access to it. Microsoft says that a file added to an agent shapes the answers every user of the agent can see, so the file’s contents can surface in answers given to any of those users.

The products that hold these packages also change quickly, so a firm that has put client material into one needs to know where that material goes when the product changes. Google is replacing Gems with a Gemini feature it also calls skills, reusable instructions that a user can call up in any chat, and Gems in personal accounts become skills from 17 November 2026. OpenAI’s own shareable packages, called custom GPTs, are going the same way, since OpenAI has scheduled their retirement for 11 December 2026 and is moving custom GPTs to plugins, which are reusable packages of instructions and connections to other services.

Memory, the other form, means details from past conversations that the assistant keeps and uses in later ones. Claude’s memory is on by default on the Free, Pro and Max plans and off on Team and Enterprise plans until each member switches it on, and it lists what it remembers as topics a user can read, edit or delete. OpenAI’s guidance on ChatGPT describes saved memories, the details a user asks ChatGPT to remember, and references to past chats, which let ChatGPT draw on recent conversations. Each of those settings can be switched off, and a page lets the user review and delete what is stored. Gemini’s memory of past chats works only while the account keeps its activity, meaning while the Keep Activity setting saves the user’s Gemini chats to their Google account, and Google says that removing something it remembered means deleting the chats that hold it.

A client’s name or circumstances mentioned in one conversation can therefore reappear in another, and in some products deleting the conversation leaves the memory in place. Anthropic says so of Claude and Microsoft says the same of Copilot for organisations; for its consumer app, Microsoft adds that switching memory off leaves saved memories in place until they are deleted. Temporary chats, which Anthropic calls incognito chats, keep a conversation out of memory and history while it is still stored for a time. Anthropic keeps incognito chats for 30 days by default and Google keeps temporary chats for 72 hours. Microsoft says that an organisation’s retention policy may keep temporary chats where its administrators can still access them.

Client information

Whether a vendor may use what a user types to train its models depends on the plan. Anthropic’s consumer terms for its Free, Pro and Max plans let each user choose whether to allow training, and Anthropic keeps the data of users who allow training for five years and of those who do not for 30 days. Google’s privacy hub for Gemini says that Google uses activity to improve its services, including training generative AI models, that human reviewers read some chats, and that users should not enter confidential information they would not want a reviewer to see. For every user aged 18 or over, the Keep Activity setting is switched on by default, and Google deletes saved activity automatically once it is 18 months old unless the user changes that period. Microsoft’s privacy questions and answers for its older consumer Copilot app say that it uses data from Copilot for AI training except for certain users and those who opt out. Microsoft’s questions and answers cover only the version of the app offered before an updated one arrived on 18 August 2026, and Microsoft sets out the newer app’s terms on separate pages.

Business plans, the versions of each assistant offered to firms and other organisations, are generally sold on different terms. OpenAI says that by default it does not use business customer content from ChatGPT Enterprise and ChatGPT Business to train its models, and Anthropic says the same of its commercial products. Google says that Workspace does not use customer data for training models without the customer’s permission or instruction. Microsoft says that prompts and responses in Copilot for organisations are not used to train foundation models, the large general-purpose models its products are built on, and that it acts as a data processor under its data protection terms.

Retention periods, the length of time a vendor keeps conversations, differ between business products as well. OpenAI says that chats in its business products are kept until the user deletes them, unless an Enterprise owner sets a retention period. Google says that Workspace administrators set how long Gemini conversations are kept, up to 36 months. A firm’s own agreement with a vendor may set other terms, and a feature such as memory or a shared project can keep information longer or show it to more people than the firm expected.

These differences give a lawyer a practical rule: before client information goes into a tool, the lawyer needs to know which plan and terms apply, whether the firm has agreed them, and how training, memory and retention are set. Client information stays out of a tool whose terms allow training on what is typed, or retention beyond what the firm has agreed. Work that needs an assistant can use invented or anonymised facts in place of the client’s own.