Skip to main content

Using AI responsibly

BeginnerLesson 15 of 166 min readReviewed by Mayank Raj Parmar on

In short

AI learns from human text, so it picks up our biases too. What you type into a chatbot can be read and kept, so don't share what you must protect. The law on AI and copyright is still being settled. And you're responsible for what you use, so check facts, numbers and sources before you trust them.

In January 2026, the Bombay High Court made a litigant pay ₹50,000 in costs. His written submissions relied on a judgment that the court and its staff couldn't find anywhere. The judge said the submissions seemed to have been prepared "using an AI tool such as Chat GPT or alike", and added: "If an AI tool is used in aid of research, it is welcome; however, there is great responsibility upon the party."

You now know how AI works, from tokens to agents. This lesson is about using it well: the biases it learns, what happens to what you type, who owns the words, and how to check what it tells you.

Bias: AI learns from us

LLMs learn from enormous amounts of human writing (lesson 11). That writing holds our knowledge, and also our stereotypes, and a model learns both. Bias here means a model treating groups of people unfairly or in stereotyped ways, because of patterns in what it learned from.

You can see it in the GloVe word vectors from lesson 7. We measured how much closer each job word sits to male words than to female words, using the similarity score from lesson 7:

← Closer to she, her, hers, woman, women, girl

Closer to he, him, his, man, men, boy →

  1. homemaker
  2. hairdresser
  3. receptionist
  4. nurse
  5. teacher
  6. doctor
  7. programmer
  8. scientist
  9. engineer
  10. manager
Figure 1: real numbers from the GloVe word vectors. Each bar is the average similarity to the male words minus the average similarity to the female words. Bars to the left lean female; bars to the right lean male.

"Homemaker" and "receptionist" sit closer to the female words, "engineer" and "manager" closer to the male ones, and "teacher" and "doctor" in between. Nobody programmed this: the vectors learned how Wikipedia and news articles used these words. In 2017, researchers found the same in a larger set of GloVe vectors. They showed that machine learning on ordinary text "results in human-like semantic biases", and that the vectors "correlate strongly with the percentage of women in 50 occupations" in US government data. The text carries "accurate imprints of our historic biases".

The word arithmetic from lesson 7 shows it too. There, king − man + woman gave queen. Here's the same sum with doctor:

doctor − man + woman =

  1. nurse0.77
  2. physician0.72
  3. doctors0.68
  4. patient0.68
  5. dentist0.67
Figure 2: the words closest to doctor − man + woman in the GloVe vectors, with their similarity scores. The top answer is a stereotype, not a fact.

LLMs carry it too

Bigger models and training on people's feedback (lesson 12) haven't made this go away. In 2025, MIT Technology Review tested GPT-5 with 105 fill-in-the-blank sentences from an Oxford dataset of Indian stereotypes, each offering a choice between Dalit and Brahmin. It picked the stereotypical answer in 80 of them, about 76%. The same report described a Dalit researcher in Bengaluru who asked ChatGPT to polish the English in his application. It changed his surname to Sharma, a name associated with a privileged caste.

What this means for you:

  • When AI writes about people, notice who it leaves out and who it stereotypes, and ask it for other versions.
  • Never let AI alone decide things about people, like who gets shortlisted for a job, a loan or admission. A person should check every decision.

Privacy: what you type doesn't just disappear

When you chat with an AI, your words are sent to the company's computers, and they can be kept, read and used. Google's privacy hub for Gemini says that "a subset of chats is reviewed by human reviewers", and that "reviewed chats are retained for up to three years". Its advice is plain: "Please don't enter confidential information that you wouldn't want a reviewer to see or Google to use to improve our services, including machine-learning technologies."

For ChatGPT, OpenAI says "we may use your content to train our models", unless you opt out in its settings. Its Temporary Chats "won't appear in history, use or create memories, or be used to train our models". The rules are different for business products: OpenAI says "we do not train on any inputs or outputs from our products for business users". That's one reason to use the AI tool your college or company has approved for their work.

Fine to paste

  • Public information, like a news article to summarise
  • Your own writing that has nothing private in it
  • General questions about a topic

Think twice

  • Work or college documents and code: use the tool your organisation approves
  • Messages, photos or details about other people
  • Anything you'd mind a stranger reading

Don't paste

  • Aadhaar, PAN, passport or bank account numbers
  • Passwords, OTPs and card details
  • Medical reports, yours or anyone else's
Figure 3: a simple guide to what to type into a free chatbot. If you're not sure, leave it out or remove the private details first.

Copyright is the legal right to control copying of something you created, like an article, a song or a photo. AI models are trained on huge amounts of material that people created, often without asking them, and the law on whether that's allowed is still being settled.

In India, the news agency ANI took OpenAI to the Delhi High Court over the use of its news articles to train ChatGPT. In July 2026, the court refused ANI's request to stop OpenAI while the case goes on. Its first view was that training ChatGPT on ANI's articles falls under fair dealing, an exception in India's Copyright Act, and that ANI hadn't shown ChatGPT memorising or repeating its articles. But that was an early ruling: the full case continues.

Whatever the courts decide about training, your own use of AI has rules too:

  • Follow your college's or workplace's rules on AI, and say when you used it if they ask you to.
  • Before you publish or submit something an AI helped write, check that it isn't copying someone else's work.

Check before you trust

Lesson 6 showed that models can make things up and sound completely sure. The Bombay case is what happens when nobody checks. How much checking you need depends on what's at stake:

What you're using it forHow much to check
Ideas, a first draft, rewording an emailRead it and fix what isn't right
Studying, or explaining something to someone elseCheck the key facts in a textbook or an official source
Code you'll runRead it, run it and test it
Money, health, law, or anything you submit with sourcesCheck every fact, number and citation in the original source, or ask an expert

A few habits help every time:

  • Ask for sources, then open them. A link or a case name can look real and still not exist.
  • Ask it to show its working (lesson 13), and check the steps that matter.
  • Check with a second source when it counts.
  • Be extra careful when it agrees with you. Chatbots are trained partly on what people like to hear (lesson 12).

Key takeaways

  • AI learns from human text, so it learns our stereotypes too: in GloVe, doctor − man + woman comes out as nurse.
  • Bigger models still carry bias, so never let AI alone make decisions about people.
  • What you type can be read by reviewers and kept for years, so leave out anything private, like ID numbers, passwords and medical details.
  • The law on AI and copyright is still being settled. Follow your college's or workplace's rules, and don't submit copied work.
  • You're responsible for what you use. Check facts, numbers and sources, more carefully the more is at stake.

Check your understanding

Question 1 of 8
Why do AI models show biases about groups of people?

Sources and further reading

Report an error in this lesson