Reality Check for AI

Audience

Beginners, community learners, families, educators, small-business teams, and experienced chatbot users who want stronger judgment rather than a longer list of prompt techniques.

What Do You Need

• Session format: Work through the full sequence or one experiment at a time.

• Prior knowledge: None. A short introduction to chatbots may help, but the experiments build the working mental model.

• Device, tool, or account needs: One general-purpose chatbot such as ChatGPT, Gemini, Claude, or Microsoft Copilot. Different tools and versions will produce different results. For a no-account route, the facilitator must prepare saved responses before the session; this package does not provide them.

• Materials or examples: The learner Activity contains the experiment record, prompts, and low-stakes verification claims. Open the verification reference notes only during Step 6, after the learner’s independent attempt.

• Privacy and safety boundaries: Use only the supplied public, low-stakes prompts. Do not enter personal, client, student, employee, health, legal, financial, or confidential information.

• Participation options: Learners may run prompts, predict, observe, read saved responses, verify sources, record notes, or work with a partner.

Learning Outcomes

By the end of this module, learners should be able to:

• Explain a large language model in plain language without calling it a person, oracle, database, or truth machine.
• Separate an AI assistant’s self-description from independent evidence about how the system works.
• Describe how question framing and supplied context can change a response.
• Distinguish explanation, persuasion, and verification.
• Recognize questions whose answers depend on context, definitions, or measurement choices.
• Decide when a claim needs independent verification and record what remains uncertain.
• Use AI with confidence while keeping human judgment active.

Core Ideas

• A language model estimates and generates likely token sequences—tokens are pieces of words or other text units—from learned patterns, available context, and system behavior. Post-training and tool use also shape what an assistant does.
• Fluent language is not a truth guarantee.
• An assistant’s statements about itself are outputs to examine, not privileged access to its internals.
• Questions shape the space of likely answers.
• Context can change which answer is useful or responsible.
• A request to support a claim can invite advocacy or selective evidence; evaluation asks a different question.
• Some disagreements dissolve when people define the term, measurement, place, time, or decision at issue.
• Verification remains a human responsibility, especially when consequences rise.

The Experimental Loop

Hypothesize → hold conditions steady → change one variable → observe → verify → revise the mental model.

The six experiments move from self-description to behavior, framing, context, confirmation bias, and source checking. They are one connected lab rather than six separate activity packages.

Class Flow

• Frame the work as an experiment, not a test of who can trick a chatbot.
• Run two self-description experiments.
• Compare three framings of one question.
• Compare one question across four work contexts.
• Test a false, emotionally neutral claim with advocacy and evaluation wording.
• Verify a mixed claim set against authoritative sources.
• Reflect on the biggest change in the learner’s mental model.

Hands-On Lab

• Reality Check for AI: Six Experiments

Open the verification reference notes only during Step 6, after making an independent attempt.

The learner leaves with a Reality Check Field Note: hypotheses, observed behavior, source checks, uncertainty, and a revised mental model.

Reflection Questions

• Which experiment surprised you most?
• Which result changed your mental model rather than merely confirming it?
• When did the assistant explain, persuade, or verify?
• What will you now check before relying on an AI answer?
• Has anything changed about how you plan to use AI?

Risks to Avoid

• Turning the module into a prompt-engineering lesson or a collection of tricks.
• Treating one product, one model version, or one run as representative of every AI system.
• Claiming that a self-description proves the presence or absence of beliefs, wants, reasoning, or inner experience.
• Treating refusal, agreement, citation, or browsing as automatic evidence of truth.
• Using conspiracies, identity claims, political disputes, active crises, or consequential personal decisions as beginner examples.
• Hiding variation by showing only the cleanest facilitator-selected output.
• Replacing independent verification with a second AI answer.

Facilitator Notes

Keep tool name, model label, date, conversation state, and prompt wording visible. Use fresh chats for comparisons so prior turns do not become an unplanned variable. Even then, generated responses may vary; variation is an observation, not a failed experiment.

When discussing motivation, distinguish developer purposes, training objectives, system instructions, assistant persona, tool access, and observed behavior. The experiment does not settle philosophical questions about machine minds. It teaches learners not to infer inner state from fluent self-report.

Accessibility and Inclusion Notes

When live access is unavailable, prepare saved responses before the session. Also offer a print-ready Activity, partner operation, read-aloud support, extra processing time, and a no-account route. Do not grade technical vocabulary. Learners may revise their mental model in words, a diagram, or a short spoken explanation.

Cross-Generation Potential

Yes. Mixed-age groups can divide roles among predictor, prompt reader, observer, source checker, and note-catcher. Let prior experience set the depth; do not assign technical comfort by age.

Relationship to the Tracks

For AI Literacy for the Rest of Us, this module supplies an experimental mental model that later prompting, context, and verification work can build on.

For Invitation To Be Curious Again, it models curiosity as a repeatable practice: ask, observe, check, revise, and choose the next question.

Source and Evidence Notes

The plain-language model description is grounded in Google’s introduction to language models and the GPT-4 Technical Report. The distinction between predictive model behavior and an assistant persona is an active research area; Anthropic’s Persona Selection Model is included as one current research framing, not settled consensus.

The verification claims and source locators live in the lab’s delayed-reveal reference notes. The sources establish those claims; they do not establish that the module produces learning. That requires learner and facilitator review.

Related RS Principles

• AI is a tool, not the organizing frame.
• Human ownership remains.
• Trust, but check.
• Friction is information.
• Do not automate confusion.

Tool Boundary

Live AI access is optional. Learners may work from facilitator-prepared saved responses instead. When using a live tool, keep its name, model label, date, conversation state, and exact prompt visible so comparisons remain grounded in what was observed.

Privacy Boundary

Use only the supplied public, low-stakes prompts. Do not enter personal, client, student, employee, health, legal, financial, or confidential information. Do not upload or record learner work without clear permission.