AI Learning Hub › Learn AI › Self-Paced Courses › Introduction to AI › Getting Started › Files, Documents, Data, and Media

Level 1 of 4 · Getting Started · Lesson 4 of 6

Files, Documents, Data, and Media

The kinds of files, data, and media AI systems can work with, and the operations they can perform on them.

About 15 minutes

Skip to lesson

By the end of this lesson you should be able to

  • Describe common inputs, operations, and outputs: what you give a tool, what it does, and what comes back
  • Tell apart pairs of features that sound alike, such as describing an image and generating one
  • Read a product’s file claim and identify what it leaves unsaid

What can I give it, and what can I get back?

AI tools can work with more than typed questions: documents, slides, spreadsheets, screenshots, recordings. It helps to think of every use as three parts: an input (what you give the tool), an operation (what you ask it to do), and an output (what you get back). The operation is where tools differ most. Reading a file, pulling specific facts out of it, summarizing it, comparing it with another, calculating from it, converting it, and creating something new from it are different jobs, and a tool can be good at one and weak at another.

Input, operation, output: six examples

Low-risk examples using fictional material, or public material you are permitted to use.

A public PDF
Input: a published report. Operation: summarize the sections on one topic. Output: a short summary with page references to check.
A Word document or slide deck
Input: your own draft handout. Operation: suggest a clearer structure. Output: suggestions in the chat, or a revised file if the tool can create one.
A small spreadsheet
Input: a fictional or public table of registration counts for each event, with no names or other personal information (holding a public event does not make its registration records public). Operation: total by month. Output: a table of monthly totals, with the method shown so you can check the range and recompute a total.
A screenshot or diagram
Input: a public chart. Operation: describe what it shows. Output: a description you can compare with the image.
An authorized audio recording
Input: a recording you have permission to use. Operation: transcribe. Output: a transcript, where names and numbers need checking.
A video or its transcript
Input: a public lecture transcript. Operation: list the main points. Output: an outline based on the words, not on what appears on screen.

Features that sound alike but are not

  • Describing an image reads a picture you supply; generating an image makes a new one.
  • Transcribing a recording turns speech into text; a voice conversation is talking with the assistant.
  • Summarizing a transcript works from the words; inspecting video frames means looking at what is shown. A tool may do one and not the other.
  • Showing a table in the chat is not the same as creating an editable spreadsheet you can download and reuse.
  • Displaying a calculation is not proof that one was run. A figure on screen does not show whether a calculator, spreadsheet formula, or code was executed. And when a tool does run a calculation, a wrong range, formula, unit, or interpretation can still produce a wrong answer. Check the inputs, method, and result; for a small calculation, recompute it with a calculator or spreadsheet.
  • Writing slide content is not the same as producing a presentation file.

The same goes for file support. Being able to upload a file, reliably extracting what is in it, editing it in its native format, keeping its tracked changes or formulas, and exporting an editable file are five separate capabilities. A supported file extension tells you only the first.

Two documented examples

OpenAI documents two different ChatGPT capabilities here: file uploads, after which ChatGPT “can then summarize, extract information, or answer questions based on the contents,” and data analysis, in which ChatGPT “can run code in a secure environment to analyze and visualize data.” (OpenAI: ChatGPT capabilities overview (opens in a new tab); provider documentation.) Running code lets you inspect how a figure was produced, but it does not settle whether the method was right: OpenAI’s data-analysis page says to “review the generated code, outputs, and assumptions before relying on the result.” (OpenAI: Data analysis with ChatGPT (opens in a new tab)) Anthropic describes Claude’s artifacts as things it makes that open beside the conversation (a document, a deck, a small interactive tool) that you can edit, return to, and share. It says everything you make is saved to the Artifacts tab, that artifacts start private to you, and that template-based documents can be exported to formats such as Word or PDF. (Anthropic: What are artifacts? (opens in a new tab)) Creating content, retaining it in a service, exporting a copy, and sharing or publishing it are different events, and services handle them differently. Running code and sharing are also capabilities an institution can configure. Check where a result is stored, who can access it, and what your own account offers.

Try it: what does the card leave out?

A fictional task and fictional product card.

The task. A staff member has a public PDF of a 30-page annual report containing a table of program attendance by year. They want (a) the attendance figures moved into an editable spreadsheet, (b) the percentage change for each year calculated, and (c) a one-paragraph summary of the report’s section on outreach.

Product card: Quill Assist. “PDF upload supported. Spreadsheet analysis. Clear summaries.”

  1. Which capabilities does the task need?
  2. For each part of the task, what does the card establish, and what does it leave open?
Answers and discussion
  1. Reading the PDF, including its table; extracting the table accurately; creating an editable spreadsheet file; calculating percentage changes with a method you can check; and summarizing one section.
  2. (a) “PDF upload supported” says you can give it the file. It does not say whether it extracts tables reliably or can export an editable spreadsheet. (b) “Spreadsheet analysis” might mean analyzing a spreadsheet you upload, not converting a PDF table into one, and it does not say how calculations are done or whether you can see the method. (c) “Clear summaries” describes a style, not accuracy or whether the summary stays within the outreach section.

None of this means Quill Assist cannot do the job. It means the card does not tell you. A short pilot on the public file can show whether the tool is worth using: compare a few extracted figures with the PDF and recompute one or two percentages yourself. A pilot tests capability on a sample; it does not verify the whole table. Before relying on the spreadsheet, check every figure you will use against the PDF, and confirm the formula and range behind each percentage.

Before you upload

Practice with made-up files, or public files you are permitted to use. Before using real material, check that the service and your account are approved for that information, that you have the right to use the content, and, for recordings, that the people recorded were told or consented as required and that sharing the recording is allowed. Beginner Lesson 5 covers how to think about information and approved services. How to extract data carefully and check it is taught in Intermediate Lesson 5.

Name the operation, not just the file type. “It takes PDFs” is only the beginning. Decide what you need done (extract, summarize, calculate, create) and check that part of the output against the original.

Resources