AI Literacy and Responsible Use

AI Literacy and Responsible Use#

Before working with foundation models on Amazon Bedrock, it helps to be fluent in the everyday tools and risks of generative AI. Navigating AI responsibly means understanding what the tools are, how to manage your privacy, and the legal constraints around sensitive data. Learning how to configure your chat settings, protect personally identifiable information (PII), and use these platforms securely is essential for any modern workflow, and especially for educators and professionals who handle regulated data.

This primer is a practical foundation. It is deliberately platform-focused (ChatGPT, Gemini, and Claude) because those are the tools most readers touch daily, and it complements the deeper, architecture-level treatment in Module 1.

A note on accuracy

Consumer AI products change their interfaces and policies frequently. The settings paths and retention windows below were verified against vendor help centers and recent privacy reporting at the time of writing, but you should confirm the current steps in each product’s settings before relying on them. AI responses, including those in this book, may contain mistakes.

Part 1: AI and IT basics#

What is AI? Artificial Intelligence is a broad branch of computer science that builds systems capable of performing tasks that normally require human intelligence, such as learning, reasoning, language understanding, and problem-solving. It is an umbrella term: machine learning, deep learning, and generative AI all sit underneath it.

What is an LLM (large language model)? An LLM is a specialized type of AI trained on massive amounts of text. It learns context, grammar, and intent well enough to predict and generate human-like text, answer questions, and perform complex writing tasks. Mechanically, an LLM predicts the text most likely to come next given everything before it, an idea developed in detail in Chapter 2: Foundation Models and Large Language Models.

What does “multimodal” mean? A multimodal AI can process and generate more than one type of media. Instead of text alone, these models understand and combine text, images, audio, and sometimes video, either simultaneously or in sequence. Multimodality is covered hands-on in Chapter 5: Multimodal Prompting.

How these connect

AI is the broad field. An LLM is a kind of AI focused on language. A multimodal model is an AI (often built around an LLM) that also handles images, audio, or video. Keeping this nesting straight prevents most beginner confusion.

What is NLP? Natural Language Processing is the field of AI concerned with getting computers to understand and produce human language. It covers tasks such as translation, summarization, sentiment analysis, and question answering. LLMs are the current state of the art in NLP, but NLP is the broader, older discipline; an LLM is one (very powerful) way of doing NLP.

Single-modality AI: the basic input-to-output tasks#

Before “multimodal,” it helps to know the common single-modality tasks, named by what goes in and what comes out. Each is a specialized model (or an LLM applied to one job):

Task

In -> Out

What it does / examples

Text-to-text

text -> text

The core LLM task: translation, summarization, answering, rewriting.

Text-to-speech (TTS)

text -> audio

Reads text aloud in a synthetic voice (voiceovers, screen readers, voice assistants).

Speech-to-text (STT / ASR)

audio -> text

Transcribes spoken audio into text (dictation, captions, meeting notes). Also called automatic speech recognition.

Image-to-text

image -> text

Describes or reads an image: captioning, alt-text, and OCR (extracting printed text from a photo or scan).

A multimodal model simply combines several of these abilities in one system, so it can, for example, look at an image and discuss it in text, or take voice in and speak back out. Single-modality tasks are the building blocks; Chapter 5: Multimodal Prompting shows how they are combined.

Chatbots, agents, and autonomous systems#

These terms describe increasingly capable, increasingly independent AI systems. Knowing the ladder prevents a lot of hype-driven confusion:

  • Chatbots are conversational interfaces. A modern chatbot is an LLM wrapped in a chat interface that holds a back-and-forth conversation (ChatGPT, customer support bots). It responds; it does not act on its own.

  • Agentic AI / AI agents go a step further: the AI plans and takes actions to complete a multi-step task, using tools (search, code, APIs) and reacting to results, rather than just replying. Agents are covered in depth in Chapter 4: Agents.

  • Fully autonomous agents operate with little or no human intervention, pursuing goals, self-correcting, and running over long periods. Greater autonomy means greater capability but also greater risk, which is exactly why the responsible-AI practices in this book matter.

Where physical machines come in. The same ideas extend from software into the physical world, where AI controls hardware and the stakes rise because mistakes have real-world consequences:

  • Self-driving (driverless) cars use AI, computer vision, and sensor fusion to perceive the road and drive with reduced or no human input. They are a form of autonomous agent operating a vehicle.

  • Drones and robots apply the same perception-and-control AI to flying or moving machines, for delivery, inspection, mapping, or manufacturing.

The autonomy ladder

A useful way to hold these together: a chatbot talks, an agent acts within software, an autonomous agent acts on its own over time, and a robot, drone, or self-driving car is an autonomous agent that acts in the physical world. Each rung adds capability and, with it, the need for stronger safeguards (Part 3 and the responsible-AI module).

The six levels of self-driving (SAE)#

Self-driving capability is not all-or-nothing. The Society of Automotive Engineers (SAE) defines six levels (0 to 5) in its J3016 standard, and they are the industry’s common vocabulary. They also make a useful analogy for AI autonomy in general, how much the system does versus how much a human must supervise.

Level

Name

Who does the driving

0

No automation

The human does all driving; the car may only warn (e.g. blind-spot alerts).

1

Driver assistance

AI handles either steering or speed (not both); the human stays fully engaged.

2

Partial automation

AI controls both steering and speed (e.g. adaptive cruise plus lane centering), but the human must constantly monitor and be ready to take over instantly. As of 2026, most mainstream “self-driving” cars are here.

3

Conditional automation

Under specific conditions (e.g. a clear highway) the car drives itself and the human can disengage, but must take over when the system requests it.

4

High automation

The car drives itself fully within a defined Operational Design Domain (ODD) and needs no human takeover there, though it may not work in extreme conditions.

5

Full automation

The car handles all driving in all conditions with no human input; no steering wheel or pedals needed, all occupants are passengers.

How AI powers the shift. Recent progress moved self-driving away from rigid, rule-based programming toward learned behavior, the same trend this book describes for language:

  • Neural networks and computer vision let vehicles perceive and spatially reason about the world, mimicking human sight (the deep-learning ideas from AI and Tools Reference).

  • Training, not coding. Instead of writing millions of lines of rules for every scenario, automakers train models on vast real-world driving data, exactly the foundation-model shift from Chapter 2: Foundation Models and Large Language Models.

  • Next-generation planning and reasoning. End-to-end architectures and multi-stage perception-and-planning systems let a car interpret its environment and plan through unforeseen situations in real time, rather than only matching pre-programmed cases. (These planning architectures are related in spirit to, but distinct from, the LLM chain-of-thought prompting in Chapter 4: Advanced Prompting Techniques.)

Verification note

The six SAE levels (J3016) are an established industry standard. That today’s mainstream driver-assistance systems are generally Level 2, and that the field is moving toward higher levels with AI-based perception and reasoning, reflects the current state of the industry; specific vehicles, brands, and regulatory approvals change, so confirm any specific claim against current sources before citing it.

Part 2: Chat management and prompts#

Deleting chats#

Removing a conversation from your history is the simplest privacy hygiene step. The exact controls shift over time, but the current patterns are:

  • ChatGPT: in the sidebar, open the menu next to a conversation (the three dots) and choose Delete.

  • Gemini: in the sidebar, open the menu next to a conversation and choose Delete. You can also manage Gemini Apps Activity in your Google Account to auto-delete interactions on a schedule.

  • Claude: open your Chats history, then delete a conversation from its menu (hover to reveal the selection control, or open the chat and use its menu).

Deletion is not always instant

Deleting a chat removes it from your visible history, but providers typically retain backend copies for a short period for safety and legal reasons, for example, Claude states deleted conversations are removed from its backend within about 30 days. Deletion reduces exposure; it is not a guarantee that data vanishes immediately.

Saving and reusing prompts#

None of the major consumer tools has a perfect prompt library, so people improvise:

  • ChatGPT: bookmark a chat’s URL, or use Custom Instructions to persist standing guidance across chats.

  • Gemini: export a useful exchange (for example, to Google Docs) or bookmark the conversation.

  • Claude: there is no dedicated prompt-saving button, so bookmark the chat URL or, better, keep your best prompts in a separate document you control.

A simple, durable habit is to maintain your own prompt file (a plain text or Markdown document) with your most effective, reusable prompts. It is portable across tools and never breaks when a vendor changes its UI.

Reusing a prompt: turn it into a saved assistant#

Bookmarking a chat saves the conversation; the more powerful move is to save the prompt itself as a reusable assistant so you can run it on demand without pasting anything. Each major tool has its own mechanism:

Tool

Feature

How to reuse a prompt with it

ChatGPT

Custom GPTs (and Custom Instructions)

Build a Custom GPT (Explore GPTs -> Create), paste your refined prompt as its instructions, optionally attach reference files, and save. Launch it anytime from the sidebar, no re-pasting. Use Custom Instructions for standing preferences that apply to every chat.

Gemini

Gems

Create a Gem (a saved custom assistant), put your prompt and persona in its instructions, and reuse it from the Gems list. Good for a fixed role you call repeatedly.

Claude

Projects and Skills

Use a Project to bundle standing instructions plus reference files the model reads on every chat in it. Use Skills (reusable instruction folders, each a SKILL.md with steps and resources) to package a repeatable task the assistant can invoke on demand.

The pattern is the same everywhere: take a prompt you have refined until it works, then save it as a Custom GPT (ChatGPT), a Gem (Gemini), or a Project / Skill (Claude), so the quality is reproducible and one update improves every future use. For a deeper, tool-agnostic version of this workflow, see A Practical AI Workflow: Making AI Do the Heavy Lifting; the workspace features are summarized again under “Workspaces for reusable context” below.

Organizing your chats#

As your history grows, a flat list becomes unusable. The major tools offer the same basic housekeeping, even if the buttons differ:

  • Rename a conversation to something descriptive instead of the auto-generated title.

  • Pin or star the conversations you return to so they stay at the top.

  • Archive chats you want out of the active list but do not want to delete.

  • Group related work into a workspace (see “Workspaces for reusable context” below).

A few minutes of naming and archiving each week keeps your history searchable and makes it far easier to find, and to safely delete, the right conversations.

Finding past conversations#

All three major assistants now let you search your chat history by keyword, so you can recover a prompt or answer without scrolling. If your tool’s search is weak, this is another argument for the personal prompt file above: anything you keep in your own document is searchable with tools you control. When you cannot find a past chat, check whether it was an archived or temporary chat (the latter is never saved, see below).

Temporary and private chats#

Most assistants offer a temporary or incognito mode (for example, ChatGPT’s Temporary Chat) for a conversation that is not saved to your history and is not used to build memory or, where you have opted out, to train models. Use it for one-off, sensitive, or experimental prompts you do not want retained. Two caveats: temporary chats still pass through the provider’s systems and may be kept briefly for safety, and because they are not saved, you cannot return to them later, so copy anything you want to keep before you close the window.

Memory and personalization#

Newer assistants can remember information across conversations rather than treating each chat as a blank slate:

  • ChatGPT has a Memory feature that saves facts it infers about you (your name, preferences, recurring tools) and reuses them. You can view, edit, or delete individual memories, or turn the feature off, in settings.

  • Claude offers memory and document-shaped Projects that carry context across chats within a project.

  • Gemini offers Gems (saved, persona-shaped custom assistants with standing instructions) and personalization tied to your Google account.

Memory is convenient but is a privacy surface: review what your assistant has stored periodically, and remember that anything in memory may influence future answers and may persist until you remove it. For sensitive work, use a temporary chat (which bypasses memory) or turn memory off.

Workspaces for reusable context#

Beyond single chats, the major tools provide workspaces that bundle standing instructions and reference files so every conversation in them starts with the same context:

Tool

Workspace feature

ChatGPT

Projects (group chats and files) and Custom GPTs (reusable assistants with their own instructions and knowledge).

Claude

Projects (add files the model reads on every chat in the project).

Gemini

Gems (custom assistants with persistent instructions).

These are the consumer-facing version of the prompt-template and master-prompt ideas in A Practical AI Workflow: Making AI Do the Heavy Lifting: set the context once, reuse it everywhere. Keep the same data-protection rules in mind, do not load regulated or sensitive files into a consumer workspace (see Part 3).

Exporting your data#

You can usually export your data (your conversations and account information) from the assistant’s settings, often labeled “Export data” or “Download your data.” This is useful for keeping your own backup of valuable chats, for moving prompts into your personal prompt file, and for exercising data-access rights. Treat the exported archive like any sensitive document: it contains everything you ever typed, so store it securely and delete it when no longer needed.

Sharing chats safely#

Most tools can create a shareable link to a conversation. Two rules keep this safe: first, a shared link usually makes that chat viewable by anyone who has the link, so never share a conversation that contains personal, confidential, or regulated information; second, sharing typically captures a snapshot, later messages may or may not appear, so re-check what the link actually exposes before you send it. When in doubt, copy the specific text you want to share rather than the whole conversation.

Part 3: Sensitive data, PII, FERPA, and HIPAA#

This is the part that matters most professionally. Misusing a consumer AI tool with regulated data can create real legal exposure.

Should you put sensitive data or PII into a consumer AI tool?#

As a default, no. Do not input PII, Social Security numbers, confidential financial records, unreleased business intellectual property, passwords, or similar, into standard consumer AI models. Public, consumer-tier models often use your input to further train their systems unless you have opted out or are on a contract that forbids it.

Opt out of training where you can

On consumer accounts you can usually turn off model-training on your data. As of this writing the paths are roughly:

  • ChatGPT: Settings -> Data Controls -> turn off “Improve the model for everyone.”

  • Claude: Settings/Privacy -> turn off “Help improve Claude.”

  • Gemini: turn off Gemini Apps Activity (“Keep Activity”).

Even after opting out, providers may retain logs briefly (on the order of days) to monitor abuse. Note that policies change: confirm your current settings rather than assuming a default. For genuinely sensitive work, a contractual tier (business/team/enterprise) that prohibits training on your content is safer than a single toggle.

FERPA: protecting student data (education)#

The Family Educational Rights and Privacy Act (FERPA) restricts disclosure of protected student information to unvetted third parties. To use AI tools without violating it:

  • Anonymize first. Strip names, student IDs, and other direct identifiers before any text touches an AI tool.

  • Use an institutional “walled garden.” Prefer tools your institution has vetted under an agreement that keeps data private and excludes it from model training (for example, an enterprise/education subscription), rather than a personal consumer account.

HIPAA: protecting health data (healthcare)#

The Health Insurance Portability and Accountability Act (HIPAA) protects Protected Health Information (PHI). To stay compliant:

  • Never input PHI such as patient names, medical record numbers, or specific conditions tied to an individual into a general consumer tool.

  • Verify compliance. Only use AI platforms that will sign a Business Associate Agreement (BAA) and that run on secure, encrypted infrastructure with auditable access logs.

The common thread

FERPA and HIPAA differ in the data they protect (student records vs. health records), but the safe-use recipe is the same: de-identify the data and use a vetted, contractually bound tool. A signed agreement (an institutional walled garden for FERPA, a BAA for HIPAA) is what legally separates “private, not used for training” from “consumer default.”

Part 4: AI for educators#

A frequent question: can educators upload documents to help grade assignments? Yes, with a critical condition, de-identify the documents first.

  • What you can do: upload an anonymized rubric, a syllabus, or student work with all names and identifiers removed.

  • Best practice: check whether your institution has an enterprise or education subscription (for example, a university-approved enterprise LLM or Copilot tenant). These accounts contractually guarantee uploaded documents stay private and are not used to train external models, which a personal account does not.

A practical grading workflow

  1. Remove names, student IDs, and identifying details from the document.

  2. Use an institution-approved, agreement-backed tool, not a personal consumer login.

  3. Keep the AI’s role to drafting feedback against an anonymized rubric; the educator makes the final judgment.

  4. Never paste a student’s identity back into the tool to “personalize” feedback; reattach names only in your own private records.

Key takeaways#

  • AI is the broad field; an LLM is language-focused AI; a multimodal model also handles images, audio, or video.

  • Know how to delete chats and save prompts, and remember deletion is not always immediate.

  • By default, keep PII and regulated data out of consumer AI tools, and opt out of training where the setting exists.

  • FERPA and HIPAA share one safe pattern: de-identify the data and use a vetted, contractually bound tool (walled garden or BAA).

  • Educators can use AI for grading support only on anonymized material, ideally through an institution-approved account.

Sources verified for this chapter

Deletion and opt-out steps were checked against the Claude Help Center and recent 2026 privacy guides; FERPA and HIPAA guidance reflects standard compliance practice (de-identification, walled-garden agreements, and Business Associate Agreements). Because vendor settings change often, treat specific menu paths as starting points and confirm them in the product. Selected references: