Responsible Use of AI and Digital Tools
Lesson 1 of 3
1Responsible Use of AI and Digital Tools
Before using any AI tool in your learning, you need a clear framework for when these tools genuinely help you and when they harm your learning long-term.
What the picture shows: Understand the problem, try it yourself, then use the tool, check the result and say that you used it.
Map of this part
The concepts in order, click any one to jump to it:
Two ways into the same idea
AI in learning is like a calculator in maths class: very useful once you understand division, harmful if you use it instead of learning it. This lesson gives you four simple rules: say when you used it, understand before you copy, check everything it says, and never hand it your secrets.
Responsible use rests on transparency (disclose the help), understanding before copying (you can explain and change the code), verification (models hallucinate facts and sources), and data protection (never send API keys, personal data or private code to an outside service).
What You Need Before This Part
💻Basic familiarity with AI tools
A general understanding of what tools like ChatGPT, Claude, or Gemini are, you don't need deep expertise, just having tried typing a question into one and seeing how it responds.
🎓An educational context
Understanding that you're in a learning environment with academic integrity rules, meaning there's a clear difference between using a tool to help you understand and using it to bypass the learning process itself.
🌐Basic digital literacy
The ability to use web applications, type text, and distinguish between reliable and unreliable information online, skills you need to evaluate any AI tool's output.
The Core Idea: A Thinking Tool, Not a Replacement for Thinking
1Transparency
If you use an AI tool in work you're graded on, disclose it clearly per your institution's policy, transparency isn't an optional extra, it's part of academic integrity.
2Understanding before copying
If you can't explain a solution in your own words, you haven't actually learned it, even if the solution is 100% correct. A task's goal is building your understanding, not delivering a working file alone.
3Always verify
Large language models can "hallucinate" information that sounds entirely confident but is wrong, verify any important fact or piece of code with an independent source before relying on it.
4Privacy and data
Don't enter sensitive data (passwords, other people's personal information, confidential code) into general AI tools whose privacy policy you don't know precisely.
Why does hallucination happen specifically? At its core, the model generates the statistically most likely continuation of text based on patterns learned during training, its goal is a fluent, linguistically coherent continuation, not looking up a verified database and checking each fact against it. That's why a completely wrong answer can come out exactly as confident and polished as a completely correct one, stylistically the two are nearly indistinguishable; correctness and confidence are entirely separate things to the model.
Hallucination itself isn't a random glitch or a "bug" you can patch with a small fix, a 2025 OpenAI paper shows it has two compounding causes that work together. The first starts during training itself: the model learns from text with no attached "true/false" label at all; it learns style and general linguistic context, not the truth of every single fact inside it. If a particular piece of information (say, an obscure person's birth date, or a precise reference number) is extremely rare in the training data, there's no reliable statistical basis the model can build an accurate answer on, much like trying to guess someone's birthday from a photo alone; you might produce a plausible number, but there's no guarantee it's right. The second cause is more serious because it persists even after training: common model-evaluation methods only give credit for a correct answer, and give zero credit to a model that honestly says "I don't know", which pushes the model, through the whole training process, toward confident guessing rather than admitting uncertainty, because a guess might land correctly and earn credit, while admitting ignorance never earns any. Hallucination, then, is a predictable outcome of a training-and-evaluation equation that rewards apparent confidence over honest restraint, not random fuzziness inside the model.
And why do we insist so strictly on "understanding before copying," even though it looks slower? Because this isn't a purely moral judgment, it's a well-established finding in learning science called the testing effect, or retrieval practice: the act of trying to recall or build a solution yourself, even when you get it wrong on the first attempt, cements the knowledge in your long-term memory far more strongly than reading a ready-made correct answer. Research by Karpicke and colleagues in this area has repeatedly found that people who retrieve an answer themselves (even by attempting a short written answer) remember it for longer than people who simply reread it, even though their performance in the moment of retrieval can look weaker. Copying a large language model's solution directly robs you entirely of that critical moment, you get the correct answer instantly, but with no actual retrieval effort to anchor it in memory, which is exactly why a student who copies a "100% correct" solution often fails to solve a similar problem alone two weeks later.
As for privacy, the core of it is simpler than it sounds: text you type into a general AI tool is usually sent to a remote external server, and may be logged or used as future training data depending on the operating company's policy, a policy you usually don't know in detail for most free tools. Once text leaves your device, you no longer fully control it; so the practical rule is simple: anything you wouldn't want a complete stranger to see, don't type into a tool whose privacy policy you don't know precisely.
Using the Tool Instead of Letting It Use You
| Healthy use | Use that harms your learning |
|---|---|
| Ask for an explanation of a concept you don't understand yet, with different examples | Ask for the complete solution to a task you're required to solve yourself |
| Use it to review your code after writing it and find bugs in it | Copy its code directly without understanding a single line of it |
| Ask for ideas to design a solution, then write the code yourself | Ask it to write the complete solution, then "edit" it only cosmetically |
Why does this difference actually matter for your skill, not just your grade? Take a simple example unrelated to any of your real projects: imagine you asked a language model to solve a small problem, printing a list's items in reverse order, and it gave you code using recursion, a technique you hadn't studied before. If you copy that code as-is and submit it, you're handing in something that works, but your mind never went through a single step of actually building that logic: you never identified the base case that stops the recursion, never traced how calls stack on top of each other, and never made a mistake and then caught and fixed it yourself, and it's exactly those three steps that actually cement an understanding of recursion in your mind, not just seeing the finished code. But if you instead asked the model for just a general direction ("walk me through this: how could I solve this without a for loop?") and wrote the code yourself step by step, you'd be going through the same real retrieval-practice process described in Part One above, and you'd come out of the exercise with deeper, lasting understanding, even if it took longer and you made mistakes along the way. Making a mistake during the attempt isn't failure; it's part of the very mechanism that builds long-term memory.
Figure 1, what actually happens in your memory depending on how you handle the model's answer: copying gives an instant solution but weak memory; real retrieval practice is slower but builds lasting understanding.
Worked Example: Two Scenarios Using the Same Tool
Imagine a student named Saad learning to write functions in Python who encounters a bug in code he wrote himself, the function returns None instead of the expected value:
def calculate_average(numbers):
total = 0
for n in numbers:
total += n
average = total / len(numbers)
# forgot to write return average
| Approach | What Saad does | Long-term result |
|---|---|---|
| Direct copy | Pastes the full code into ChatGPT and asks "fix the code", gets the corrected version with return, copies it and submits | Didn't understand why the problem existed in the first place, next time he forgets return, he'll fall into the same error and won't recognize it on his own |
| Educational use | Asks the model: "my function returns None, what are common causes for this in Python?", gets a list of causes, traces them one by one in his code until he discovers the missing return | Built a mental pattern for diagnosing this error, next time he'll recognize the symptoms of a missing return immediately without help |
Myth vs. Reality
| Myth | Reality |
|---|---|
| AI models always give correct answers because they're "intelligent" | The model predicts the statistically most likely text, it can be entirely confident and entirely wrong at the same time (hallucination) |
| Using any AI tool in studying = cheating | It depends entirely on how you use it: asking for a concept explanation or reviewing code you wrote yourself is completely different from asking for a ready-made solution to a task assigned to you |
| The model "understands" what you say and processes it the way a human does | The model processes statistical patterns in text, it has no consciousness or human-like understanding, but captures linguistic correlations deep enough to appear like comprehension |
Real-World Applications
🏫Academic integrity policies
Universities and educational institutions worldwide are updating their policies to define when AI tool use is permitted and when it constitutes a violation, transparency through disclosure has become a fundamental requirement.
🔎Professional code review
Professional developers use AI tools to review their code after writing it, not to write it from scratch, but to catch weak patterns or security vulnerabilities they might miss.
📰Fact-checking in journalism
Journalists use AI tools for initial drafts, then manually verify every fact, because fabricating even one piece of information in a published article could destroy the entire institution's credibility.
🧑⚕️Responsible AI in medicine
Medical diagnostic systems that rely on AI always require a human doctor's review, because a single misdiagnosis could be a matter of life or death.
Deep Dive: Trust, Bias and Personal Boundaries
This section extends the core idea: why smart tools pull us toward over-trust, and how to build habits that protect our learning and accountability. Read it after the previous section and think about your behavior, not "other people’s".
1. Fluency is not evidence of correctness
A language model produces tidy, coherent text because that is what it was trained to do: predict the likely next word. There is no "check the facts" step inside it. So a wrong answer can arrive in the same confident tone as a right one, and the model usually cannot say "I don’t know" unless it was built or asked to. There is a second side that concerns you as the reader: a smooth explanation gives you the feeling of having understood, while real understanding shows only when you try to re-explain or apply it yourself. This is the Illusion of Competence.
2. Automation bias and how it creeps in
Automation bias is the tendency to favour suggestions from automated systems even when they conflict with what you can see for yourself. It grows under time pressure, tiredness, and when the tool is usually right, because repeated success lowers your vigilance. The cure is not avoiding tools but a small habit: before accepting an answer, ask "what independent evidence can I reach within two minutes?"
Figure 2, The dependency loop: each time effort is removed too early, understanding weakens and reliance grows.
3. Bias and neutrality
Models learn from text written by people, so they inherit people’s patterns: what is written often seems "normal", and what is written rarely may be sidelined or distorted. That does not mean the model "intends" anything; it means its outputs are not automatically neutral. In practice, when you ask about people, cultures, or decisions that affect others, request several perspectives and compare them against a trusted source. Well-known responsible-AI frameworks, such as the OECD AI Principles, stress accountability, transparency, and human responsibility.
4. When a tool helps and when it steals the attempt
| Situation | Use that builds learning | Use that replaces learning |
|---|---|---|
| A fuzzy concept | You explain it in your words, then ask for corrections | You ask "explain it" and just read |
| A bug in your program | You ask for a hint on where to look | You paste the program and copy the fixed version |
| Review before submission | You ask for critique of your solution | You ask for a full solution and compare |
| Summarizing a reading | You write your summary first, then compare | You read only the ready summary |
5. Your personal boundaries and disclosure
Personal boundaries are short rules you set before you get tired and lenient. Examples: "I try the problem for twenty minutes before any help", "I never copy what I cannot explain", "I never send private data about myself or others". Disclosure (transparency) is not punishment but trust: when you state which tool you used, for what, and what you changed yourself, others can honestly evaluate your effort and understanding.
"I used an AI assistant to clarify two-dimensional arrays and asked for a hint only; I wrote the solution myself and checked the result by trying it on three cases."
Figure 3, A decision tree before reaching for an AI tool.
Try it yourself
An assistant confidently gave a rare reference number, and it was wrong. What explains it best?
Hallucination is an expected result: the likeliest text is not necessarily the true one, and common evaluations reward a confident guess more than “I don't know”. That is why we verify with an independent source.
Which behaviour applies “understand before you copy”?
Retrieving it yourself, even with mistakes, fixes knowledge better than reading a finished answer (the testing effect). Disclosure alone does not replace understanding.
Which of these should you never type into a general AI tool whose privacy policy you do not know?
General tools may keep what you type. Sensitive data, passwords and secret code stay out.
Quick Review Flashcards
Check Your Understanding
If a language model gave you code that "works" but you don't understand why, should you submit it?
Not recommended, submitting code you don't understand means you haven't gained the skill the task required, and this can show up later in an interview or a bigger project that relies on exactly that understanding.
What is the fundamental difference between asking the model "what are common causes for this error?" versus asking it to "fix the code"?
The first question keeps you in the role of thinker and analyst, you receive information that helps you diagnose the issue yourself, going through real retrieval practice. The second eliminates this process entirely: you get the solution directly with no mental effort that cements understanding.
If you published information from an AI model in a study report without verifying it, and it later turned out to be fabricated, who bears responsibility?
You bear full responsibility, the tool has no legal or academic liability. Your responsibility as a user includes verifying every important piece of information before relying on or publishing it.
Spot the two mistakes: a student wrote in his report, "The model answered confidently and in an organized way, so the information is correct and I didn't check it. And since I mentioned at the end that I used the tool, I'm not responsible for any error in it."
Mistake one: treating confidence and polish as proof of correctness, to the model, correctness and confidence are separate things (hallucination). Mistake two: believing that disclosing the tool transfers responsibility to it; disclosure is required but never replaces verification, and responsibility stays with the author.
A tool gave you a convincing answer you cannot verify within two minutes; what do you do?
Treat it as a hypothesis, do not build an important decision on it, and look for an independent source or ask the tool for sources and open them.
Resources for this part
Sorted by kind: start with the official docs when you want the precise reference, or with an article or video when you want the explanation first.
Resources for this lesson
Grouped by type. Start with the official docs for precision, or an article or video for the explanation.
Official docs3
Done with this lesson?
Sign in to save your progress.