Before you type:
what to do when you start an AI conversation in a clinical trial setting

01A tool that remembers more than you think

Most people using AI in clinical trials treat it the way they treat a search engine. You open a tab, type a question, read the answer, close the tab. Clean. Contained.

It isn't.

The AI you use for work is not a search engine. It is a system that may carry context from previous conversations into every new one. Silently, automatically, and without labelling it. When you start a new session, the model may already "know" things about you, your organisation, your prior sessions, and the documents you have shared before. You did not write any of that into today's prompt. You cannot see it, and the tool does not tell you it is there.

For most tasks, this is a convenience. For regulated clinical work, it is a data governance problem.

02What is actually happening

When you type a question into an AI tool not built for regulated work (AI tool from here on), what reaches the model is not just your words. It is your words plus whatever the tool has accumulated from your previous sessions (stored as memory, conversation history, or project context), injected invisibly into the model's working window before your first character appears.

The model does not separate this background information from what you just wrote. It processes everything together. Your question about a protocol amendment is answered through the lens of every prior conversation that happened to be in scope. The model cannot tell you which part of its response came from your document today and which part came from a session three weeks ago.

In regulated environments this matters for a specific reason: Traceability. If an output cannot be traced to a defined, reviewed input, it cannot be defended in an audit. A response shaped by undisclosed prior context has no traceable input. It is not auditable, and neither is any document that was produced using it.

If an output cannot be traced to a defined, reviewed input, it cannot be defended in an audit.

03What you can do yourself

These steps will not solve the underlying problem, but they will reduce your exposure when working in an AI tool.

1. 1

   Check and clear memory before any regulated task

   Most AI tools store memories of your previous conversations. Before starting work on anything regulated, go to the tool's settings and check what has been stored. Delete anything that does not belong in scope for today's work.

   In Claude Settings → Capabilities → Memory.

   In ChatGPT Settings → Personalization → Manage memories.

   Make it a habit to review and delete as needed before each work session, not a one-time setup.

2. 2

   Use incognito or private mode where available

   Some tools offer an incognito or private conversation mode that disables memory and history for that session. This is the closest these tools come to a clean context window. Use it for any work involving clinical data, draft documents, or regulatory content.

   In Claude Start a new conversation and select the incognito option, or check your account settings for a persistent option.

   In ChatGPT New chat → model/menu area (top right corner) → Temporary Chat.

3. 3

   Use separate projects for separate studies

   If the tool you use supports projects or workspaces, keep each study in its own isolated project. This is not full isolation, as shared memories can still cross boundaries, but it limits casual context bleed between studies. Never carry a conversation from one study into a session where you are working on another.

4. 4

   Declare your context at the start of every session

   Do not assume the model knows what you are working on today. Open every regulated query with an explicit declaration:

   "I am working on [task]. The only document relevant to this question is [document]. Please answer only from the content I provide in this session."

   The more specific you are about what you need right now, the less room the model has to fill in gaps from its memory. It does not eliminate the problem, but it shifts the balance toward what you intended.

5. 5

   Stop if you see specific details you did not type or share in this session

   If the model mentions a study name, protocol number, sponsor, or any detail specific to your work that you did not type, ask where it came from. "You mentioned X. What is that based on?"

   ⚠

   If the model cannot point to something you provided in this session, it is drawing on context from a previous conversation. That is a red flag!

04What these steps do not solve

These are behavioural controls. They help, but they have limits.

They work when the person applying them knows the risk exists, remembers to apply them every time, and does so correctly under the time pressures of real work. In a regulated environment, that is not a sufficient control. In GCP carefulness is not a replacement for a validated process. The same is true here.

Specifically, the steps above do not solve:

These gaps are not fixed by prompting more carefully. They require architectural controls: stateless API calls, contractual data agreements, schema-validated outputs with source references, and enforced scope at the infrastructure level. AI tools do not provide these by design.

05Pre-session checklist

06The question to keep asking

The discipline that protects you in regulated AI work is not technical. It is the habit of asking one question before accepting any AI output:

If you cannot answer that question, if the source is the model's general knowledge, an undisclosed prior session, or a document you cannot point to, the output is not ready to use in regulated work.

That is not a limitation of AI. It is the standard of evidence that clinical research has always required.