Field Guide · level 1 · Start here · How it works, optional

Who taught it all this?

Where AI's knowledge comes from, and why you hand it the document.

5 min · try it 10 min · as of Oct 2026

Bring this to your team · Team workshopsShow your team what AI learned, what it never saw, and what to hand it. Get in touch

Right topic, wrong policy

You asked about your policy. It answered about policies in general.

How it got this way

In four steps.

Read, guessing next words
Trained to help
Reading stopped at a date
It doesn't know your files

It read, guessing next words

It read huge amounts of text, guessing each next word. Misses nudged it slightly.

Then trained to help

Reading alone didn't make it helpful. Example answers and people's rankings did.

Its reading stopped at a date

That's its knowledge cutoff. Anything newer needs a search, or you.

Your report isn't on its shelves

It read public text and other data. It knows reports, not yours.

Learned the world. Hand it yours.

Paste or attach it. With the real text in front of it, the likely answer is usually true.

My pick for your own files

Gemini logo Gemini NotebookAlso good: Claude logo Claude Projects

Built for answers drawn only from the files you add. It was called NotebookLM until July.

All my picks

Try it · 10 minutes

Without it, then with it

A made-up policy, safe to paste.

New chat: ask only the question.3 min
New chat: paste all of it.3 min
Compare the two answers.4 min

The made-up policy

Our team's travel policy (made up):
1. First class only on trains over 2 hours.
2. Hotel stays over 3 nights need a manager's sign-off.
3. Book through the travel desk.

Question: Can I book first class on a 90-minute train? Answer only from the policy above, and quote the line you used.

Good looks likeOnly the second answer quotes rule 1.

In your assistant · as of Oct 2026 · Claude, ChatGPT, Gemini

Hand it over, and know its date

Menus move often. For work documents, use the tool your company approves.

CheckYour company allows it
CheckThe newest version
CheckIts knowledge cutoff

Check your workplace's rules before you paste a confidential document into any AI tool.

Claude

Attach it, or keep it in a project

Attach

+thenAdd files or photos

Up to 20 files a chat. Or paste the text.

Keep

Projectsthen+ New Projectthen+

Add files to the project's knowledge, for documents you reuse. Free: up to five projects.

Date

Knowledge cutoff: June 2026

For Opus 5.5, Sonnet 5.5 and Haiku 5.5. Anthropic's help centre lists other models' dates.

ChatGPT

Attach it, or add it to a project

Attach

Attach the file, or paste the text

Pastes over 10,000 characters become an attachment. Free: 3 uploads a day.

Keep

New project

Then add its files: 5 on Free, 25 on Go and Plus, 40 on Pro and up.

Date

Knowledge cutoff: 20 Apr 2026

For GPT-6 Sol, behind GPT-6 in Chat on Plus and up, on OpenAI's developer page for that model. Free and Go use a Luna model with its own date.

Gemini

Attach it, or use a notebook

Attach

Add filesthenUpload

Or Add from Drive. Up to 10 files a prompt, subject to availability.

Keep

Gemini NotebookthenCreate new notebook

Upload your sources, then ask about them.

Date

Knowledge cutoff: March 2026

For Gemini 4 Argon and Gemini 3.6 Flash, in Google DeepMind's model cards. Some areas are older.

Quick check · three questions

Did it stick?

1A teammate asks AI about a rule your company changed last month. No file, no search. What is the answer built on?
The answer: A Your rule wasn't in its reading, and its reading stopped at a date. With nothing handed over, it writes what's likely.
2Why does pasting in the real contract usually give a better answer about it?
The answer: B The pasted text is its working memory for that chat, not new training. With the real text in front of it, the likely answer is usually true.
3Its answers come out polite, organised and on topic. Where does that mostly come from?
The answer: C Reading alone didn't make it a good assistant. Example answers and people's rankings did. OpenAI says its models keep no copies of their training data.

0 of 3 answered

Pasting isn't trainingSide note

What you paste sits in that chat's working memory, Anthropic's term, apart from what the model was trained on. For documents you reuse, keep them in a project.

Search helps with news, not your filesSide note

Web search can bring in public pages newer than the cutoff. Your private files still need you to attach them, or a connector working with your own access.

Still check what mattersSide note

The real text makes a true answer likely, not certain. Anthropic says grounding techniques reduce made-up answers but “don't eliminate them entirely”. Check the lines your work stands on.

Sources, checked 19 Oct 2026Source

Product details as of Oct 2026, each checked on the vendor's own page on 19 Oct 2026. Recheck by 19 Nov 2026. The examples in the pictures are made up for this lesson.

Anthropic, Glossary (pretrained to predict the next word; pretrained models not inherently good at answering questions; fine-tuned to be a helpful assistant; RLHF ranks example texts; the context window is a working memory, apart from the training data)
Anthropic, How up-to-date is Claude's training data? (knowledge cutoff June 2026 for Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1, with older models' dates; models may not be aware of events after their cutoff)
Anthropic, Models overview (reliable knowledge cutoff and training data cutoff Jun 2026 for the four current models)
Anthropic, Transparency Hub (training data: publicly available information, public and private datasets, user data, synthetic data, and for Sonnet 5.5 data users explicitly permitted for training; pretrained, then post-training to be an effective assistant)
Anthropic, Tracing the thoughts of a large language model (27 Mar 2025; models like Claude are not programmed directly by humans but trained on large amounts of data)
Anthropic, Reduce hallucinations (word-for-word quotes from the document; allow "I don't know"; the techniques "don't eliminate them entirely")
Anthropic, Upload files to Claude (+, then Add files or photos; up to 20 files a chat)
Anthropic, How can I create and manage projects? (Projects, + New Project, + to add files to project knowledge)
Anthropic, What are projects (Free users can create up to five projects)
Anthropic, Use the connectors directory (Claude inherits each person's permissions from the connected service)
Anthropic, Claude release notes (Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 launched Sep to Oct 2026; no change to uploads or projects since)
OpenAI, How ChatGPT and our foundation models are developed (predict the next most likely word; parameters adjusted slightly; pre-training and post-training; no stored copies of training data; training data from the public internet, third-party partners, users, trainers and researchers, and synthetic data)
OpenAI, Does ChatGPT tell the truth? ("Knowledge cutoff: The models are trained on data up to a certain point"; without search, responses are based on what the model learned in training)
OpenAI, GPT-6 Sol model page, OpenAI's developer site (Apr 20, 2026 knowledge cutoff)
OpenAI, GPT-6 Luna model page, OpenAI's developer site (May 18, 2026 knowledge cutoff)
OpenAI, GPT-5.6 Luna model page, OpenAI's developer site (Feb 16, 2026 knowledge cutoff)
OpenAI, GPT-6 and other models in ChatGPT (with GPT-6 selected, GPT-6 Sol powers Instant to Extra High; Free and Go receive GPT-5.6 Luna)
OpenAI, ChatGPT release notes (4 Aug 2026: pastes over 10k characters become an attachment, on all plans; 7 Oct 2026: Plus, Pro, Business and Enterprise use GPT-6 Sol, Free and Go use GPT-6 Luna)
OpenAI, Uploading files and audio to ChatGPT (Free users are limited to 3 file uploads per day)
OpenAI, Projects in ChatGPT (New project in the sidebar; files per project: 5 Free, 25 Go and Plus, 40 Edu, Pro, Business and Enterprise)
Google, Learn about generative AI (machine learning models, generative AI included, learn through training, by observation and pattern matching; LLMs pick probable next words)
Google DeepMind, Gemini 4 Argon model card (13 Oct 2026; knowledge cutoff March 2026, less up to date in some domains; training data includes publicly available web documents, licensed data, user data under Google's terms and user controls, and synthetic data; post-training with instruction tuning, reinforcement learning and human-preference data)
Google DeepMind, Gemini 3.6 Flash model card (21 Jul 2026; knowledge cutoff March 2026; some domains limited to January 2025)
Google, Gemini Apps release notes (13 Oct 2026: Gemini 4 Argon; no change to uploads or notebooks since)
Google, Upload & analyze files in Gemini Apps (Add files, then Upload or Add from Drive; up to 10 files a prompt, subject to availability)
Google, Create a notebook in Gemini Notebook (Create new notebook; Upload a source; ask questions about your sources)

Free skillCheck a draft against its sourcesBring this to your teamTeam workshops · Get in touchBack to level 1, Start here

Paste in the real document.

Next lesson · 15 minHow do you brief AI so the first answer is close?