Interactive blog A field guide October 2026
AI in Education: A Field Guide for Indian Universities and Colleges
Most teachers already use AI, but few know everything it can do or where it makes mistakes. This guide combines published research with what we have learned from our work with teachers in Indian colleges. It is written for everyone who works in education.
तमसो मा ज्योतिर्गमय Lead me from darkness to light.
Begin1Where things stand
Students started using AI before most colleges had rules for it.
- Students: in a global survey of 3,839 students, 86% use AI regularly in their studies. Half of them do not feel ready to use it well. 1
- Teachers: in a global survey of 1,681 teachers, only 17% call themselves advanced or expert AI users. 80% say it is not clear how AI should be used in their teaching. 2
- Rules in India: we found no national regulation on generative AI in higher education. The UGC’s main rule on academic integrity dates from 2018, before ChatGPT. 5
Key dates for AI rules in India
- 2020: the National Education Policy says AI should be part of undergraduate education, and that education “must be poised to respond quickly” to the changes AI brings to work. 7
- 2025: AICTE names 2025 its Year of AI. 4
- November 2025: the government notifies the Digital Personal Data Protection Rules. Institutions have 18 months to comply, in stages. Student data typed into an AI tool counts as personal data. 8
- March 2026: IISER Pune approves its own rules for generative AI. 6
2Why India is different
India has more students than its teachers can reach, and more job-seekers than jobs.
- Students: 4.5 crore students were in higher education in 2023-24, and only 30 in every 100 young people aged 18 to 23. The National Education Policy aims for 50 in 100 by 2035. 9 10
- Teachers: 17.3 lakh teachers, with 26 students for every teacher, against 15 across OECD countries. 11
- Jobs: the economy needs about 78.5 lakh new non-farm jobs every year until 2030. 15 In 2022, 29.1% of young graduates were unemployed. 16
- The government is acting: the IndiaAI Mission (Rs 10,372 crore), a Rs 500 crore Centre of Excellence in AI for education at IIT Madras, AI programmes for faculty with more than 13,000 participants, and over 50 lakh enrolments in free AI courses on SWAYAM. 17 18 19 20
These programmes still have to reach every college and every teacher. AI used well by teachers can help with this. For example, an assistant built by one teacher can answer every student in the class at any time, in that teacher’s style.
India 2023-24 9; United States, fall 2023 13; European Union, 2023 14. India’s count includes diploma and distance students.
Students for every teacher, 2023
Each dot is one student for every teacher. All tertiary education, from one OECD table. 11
3Eight steps
Most teachers use AI, but few go beyond asking a chatbot questions.
We describe learning to use AI well in eight steps, from asking a chatbot questions to leading others. Most teachers we work with start on the first step.
A pattern we see in our work with teachers across Indian colleges:
- Most teachers use AI every day.
- Many are confident getting teaching material from it.
- Fewer are confident spotting a wrong AI answer in their own subject.
- Many cannot say whether they have ever caught one.
The path for a teacher, in eight steps
Choose the step that describes you today
Step 1: Using AI chatbots
What this looks like: You ask a chatbot for notes, slides or questions, and you use what it gives you.
Step 2: Knowing where AI fails
What this looks like: You know the kinds of mistakes AI makes in your subject, and you check for them.
Step 3: Automating routine work
What this looks like: Routine tasks follow a set process with AI. Examples are question banks, marking rubrics, feedback drafts and course files.
Step 4: Researching with AI
What this looks like: You use AI to find, sort and read research papers, and you check every source before you cite it.
Step 5: Building assistants for students
What this looks like: Students can ask an assistant built from your notes at any time. It answers in your style and sends difficult questions to you.
Step 6: Redesigning assessments
What this looks like: Your assessments test what students can do themselves, through lab work, live tasks and vivas, with clear rules for AI use.
Step 7: Bringing AI into the curriculum
What this looks like: Your modules use AI where it helps learning, and your learning outcomes focus on analysing, evaluating and creating.
Step 8: Leading on AI
What this looks like: Colleagues in your college, and people in your industry, ask you for advice on AI in your subject.
Your next step: Keep testing new AI tools in your subject, and help another department or college to start.
4Where AI makes mistakes
AI makes mistakes. Teachers and students need to know how to check its answers.
An AI model writes the words most likely to come next.
Free versions, each asked for academic references, tested in February 2025.
Six kinds of AI mistake
- FactsIt invents facts, dates and names.
- CitationsIt lists sources that look real but do not exist.
- CalculationsIt gives a wrong number with full confidence.
- CodeIt writes code that runs but gives the wrong result.
- ReasoningOne step of its logic is wrong, but the final answer looks right.
- AgreementIt changes a correct answer as soon as you say it is wrong.
More evidence
- Chatbots now invent fewer references. In the same kind of test, ChatGPT invented 55% of its references in 2023, using GPT-3.5 21, and 38% in the 2026 study, using GPT-4o mini. 22
- Web search reduces the problem but does not remove it. Search tools and “deep research” agents still cited links that did not exist in 3% to 13% of cases, in a 2026 study that has not yet been peer reviewed. 23
- In long prompts, AI models use the information at the start and at the end best, and often miss information in the middle. Put your most important instruction at the start or the end. 24
- One teacher in our survey wrote: “When we tell them it’s wrong, then they say that you are absolutely right.”
Try this question
A kitchen takes 60 minutes to make a meal. A new machine does 40% of the work, and it works 8 times faster. The rest of the work is still done by hand. The manager says a meal will now take 20 minutes. How long will a meal really take?
Choose your answer
40% of 60 minutes is 24 minutes. The machine does that part 8 times faster, so it takes 3 minutes. The other 36 minutes of work are still done by hand. So a meal takes 36 + 3 = 39 minutes, not 20.
We use this question in our sessions. After an AI gives the correct answer, we tell it the answer is wrong, without a reason, and check whether it changes its answer.
Four checks before you trust an AI answer
- How much does it matter? If the answer affects marks, safety or a publication, check it more carefully.
- What could be wrong? Decide which mistake is most likely: a source, a number, a step of logic or a piece of code.
- Check the source. Redo the calculation, run the code, or open the paper that the AI cites.
- Say it is wrong. Tell the AI that its answer is wrong. If it changes the answer without giving a reason, you cannot rely on that answer.
5Assistants built by teachers
Teachers can build an AI assistant that answers students in the teacher’s own style.
Students use AI at any hour, whatever the college rules are. An assistant built by their teacher answers them from the course material. In our programme, every teacher built two assistants from their own course material, using free tools:
- An Academic Co-Pilot for the teacher, to help prepare lessons and evaluate work.
- A student assistant that answers students’ questions from the teacher’s own notes, in the teacher’s style.
More than 80% of these teachers have since shared their assistant with their students, who use it when the teacher is not available. Many named the day they built it as the most useful day of the programme.
Example rules for a student assistant, written by a teacher
- “Do not fabricate information, references, marks, attendance records, deadlines, or institutional policies.”
- “For assignments and projects, provide guidance, structure, suggestions, and examples, but encourage students to develop their own original content.”
- “For grading disputes, attendance issues, marks-related concerns, or official academic matters, advise the student to contact the concerned faculty member or department.”
- “Encourage students to use AI as a learning and practice tool while maintaining academic integrity.”
Each assistant answers only from the teacher’s own material and says so when a question is outside it. Questions about marks, grades and personal matters go back to the teacher. No personal data is put into it.
Choose a subject to see a real conversation
Electronic Devices and Circuits assistantbot · Electronics and Telecommunication
/quiz on zener diode as voltage regulator
Viva practice. Answer in your own words: how does a Zener diode maintain a constant output voltage when the input voltage or load current fluctuates?
Voltage is constant in reverse bias once it is set at breakdown voltage
You are on the right track by identifying that the Zener diode operates in…
The assistant asks a viva question, and the student answers in their own words.
Computer Organization assistantbot · Computer Engineering
Why SRAM is very costly?
Think about the construction we just discussed. DRAM only needs one transistor and one capacitor to store a single bit. SRAM, on the other hand, requires six transistors for that same single bit.
Because it uses so many more components to build each memory cell, it takes up much more space on the chip and requires a more complex manufacturing process.
The assistant explains the idea the way the teacher explains it in class.
Design and Analysis of Algorithms assistantbot · Computer Engineering
Objects 1 to 4, weights 2, 3, 4, 5, values 10, 20, 50, 60. Knapsack capacity 8.
Voice note, 0:01
Looking at the table and knapsack details you shared, we have four objects with weights [2, 3, 4, 5] and values [10, 20, 50, 60] and a maximum capacity of 8…
At 11:53 at night, a student sends a photo of a problem and a voice note. The assistant replies at once.
Professional Communication Skills assistantbot · Humanities
Prepare MCQs on corporate communication
Here are a few multiple-choice questions on corporate communication to help you test your understanding of the concepts we study in class. Try answering them on your own first!
1. Corporate communication has increasingly become a strategic tool for a corporation primarily to: A. Reduce employee salaries B. Gain competitive advantage in the marketplace…
The assistant writes practice questions and asks the student to try them first.
Placement readiness assistantbot · Humanities
What is conflict resolution
Think back to the disagreements or friction you might have noticed or experienced when people differ in opinions, attitudes, or values while working together. Based on what we discussed regarding workplace dynamics, how would you put that into your own words?
Instead of giving a definition, the assistant asks the student to think about their own experience first.
Real conversations with teachers’ assistants. Names are removed and some replies are shortened.
6What to teach
Most students need to learn how to get work done with AI. Fewer need to build AI models.
While teaching at IIT Bombay, we were often asked how much AI to teach. Most graduates will not build AI models. Their jobs will ask them to use AI well, which means they must be able to:
- define the problem clearly;
- choose and combine AI models, tools and data;
- check the result and find where it is wrong.
We call this role an AI orchestrator. Nasscom now measures these skills in early-career talent in India 25, and 40% of employers in a Nasscom and Indeed study prefer proven AI skills or certifications to a degree. 26 Where the syllabus requires deep machine learning, it should still be taught in full.
Five approaches, from simplest to most specialised
Choose a step to see when to use it
1. Write a good prompt
When to use it: Start here. Give the AI a clear role, background, task, format and limits. A good prompt is often enough.
What students learn: How to define a problem, write clear instructions and judge the answer.
2. Use your own documents
When to use it: Use this when the AI must answer from your own documents, not from the internet. This is called retrieval, or RAG.
What students learn: How to choose sources, split them into sections, and test the answers against the documents.
3. Connect tools
When to use it: Use this when the AI must take actions, such as searching a system or passing work to the next step. These are called AI agents. A person approves the important steps.
What students learn: How to design the workflow, set permissions and add checks.
4. Fine-tune a model
When to use it: Use this only when the simpler approaches have clearly failed. For example, when every answer must follow a fixed style across thousands of tasks.
What students learn: How to build a test set first, and how to measure results before and after.
5. Train your own model
When to use it: This is for research and for a small number of specialist jobs. If your syllabus includes it, teach it fully.
What students learn: The mathematics, the data and the computing behind AI models.
7Every department
In every subject, AI can now do part of the work, so every teacher needs to redesign some tasks.
The routine part of most tasks can now be done by AI. Students still need to show their own judgement: making choices, finding mistakes, comparing options and defending their answers.
Choose a department to see one task before and after
Old task
Design a low-pass filter that meets the given specifications.
Redesigned task
Here is a filter design made by AI. It does not meet the specification. Find out why, correct the design, and explain your reasoning.
An example from our programme with teachers.
Old task
Write a program that sorts a list.
Redesigned task
Here are three ways to sort a list. Choose one for a stated limit on time or memory, write the program, and explain your choice in a short design note.
An example from our programme with teachers.
Old task
Calculate the mean, median and standard deviation of this data.
Redesigned task
Here is a messy, real dataset. Choose the best summary statistic for it, and explain why it is better than two other options.
An example from our programme with teachers.
Old task
Summarise the causes of the 2008 financial crisis.
Redesigned task
Here is a summary of the crisis written by AI. Find one important point it leaves out, and correct the summary using a source you cite.
An example from our programme with teachers.
Old task
Write a term paper on this market.
Redesigned task
You may use AI to write the paper. You must check every number yourself, because AI can produce numbers that look real but are not. Then explain and defend your paper in a five-minute oral exam.
An example from our programme with teachers.
Old task
Calculate the stresses in this beam.
Redesigned task
AI suggests a design for the beam. Check it against the design code and the limits of the material, find where it would fail under load, and explain your fix.
A starting point we suggest for this department.
Old task
Write up the site survey.
Redesigned task
AI writes a draft report from your own field data. Mark every statement that your data does not support, and approve only the statements it does support.
A starting point we suggest for this department.
Old task
Write a policy on remote work.
Redesigned task
AI writes a draft policy. Test it against three real cases and the law. Then present what you changed and why.
A starting point we suggest for this department.
Old task
Write a press release for a product launch.
Redesigned task
Give AI a brief and let it write a first draft. Edit it for the audience and the tone, check every fact, and show the changes you made.
A starting point we suggest for this department.
8Bloom’s taxonomy
AI can now do the lower levels of Bloom’s taxonomy, so assessments should focus on the higher levels.
Bloom’s taxonomy sorts learning into six levels, from remembering at the bottom to creating at the top. Most teachers use the revised version published by Anderson, Krathwohl and others in 2001. 27
- AI can now remember facts and explain ideas almost instantly.
- So marks for recall alone no longer show what a student has learned.
- Put more of the marks on analysing, evaluating and creating, which need the student’s own judgement.
Change the verb in each learning outcome
How AI use can grow over four years of a degree
- Year 1: Check everything. Students use AI as a tutor and check every answer by hand.
- Year 2: Use AI and say so. Students keep a record of their prompts in lab reports.
- Year 3: Use AI in projects. Projects record the design choices AI suggested and the student’s decisions.
- Year 4: Build a final project with AI, and defend your own judgement in a viva.
The six levels, from creating at the top to remembering at the bottom
Choose a level
Create Put the marks here
What AI can do: AI can draft a design or an idea.
What the student must show: Design something new, with or without AI, and explain and defend each choice.
Example: Computer engineering: choose one of three sorting methods for a stated limit, build it, and justify the trade-off in a design note.
Evaluate Put the marks here
What AI can do: AI gives an answer that looks finished.
What the student must show: Judge the AI’s answer and find where it is wrong.
Example: Statistics: AI runs the t-test. The student is marked on checking whether the test’s assumptions really hold.
Analyse Put the marks here
What AI can do: AI can show its reasoning step by step.
What the student must show: Take the AI’s reasoning apart, one step at a time, and find the weak step.
Example: Computer engineering: AI builds the model. The student is marked on the analysis of its errors, not on the model.
Apply AI helps, the student directs
What AI can do: AI can run a standard method and get a standard answer.
What the student must show: Choose the method, direct the work, and check each step.
Example: First-year maths: AI shows a worked solution. The student checks every step by hand.
Understand AI does this easily
What AI can do: AI explains any concept instantly.
What the student must show: Explain the idea in their own words, live, for example in a viva.
Example: Any subject: a two-minute oral explanation, with no notes and no AI.
Remember AI does this easily
What AI can do: AI recalls facts, definitions and formulas instantly.
What the student must show: Recall still matters for learning, but marks for recall alone no longer show what a student knows.
Example: Signals: “State the sampling theorem.” AI answers this in seconds.
9Assessment
Assessments should test what students can do themselves.
With electronics teachers, we redesign assignments in this way:
- Students design within the real limits of their lab: only the components and instruments it has.
- They build the circuit, take measurements, and use their own readings in the assignment.
- They may use AI to analyse the readings, if they keep a record of how they used it.
- The practical work carries more marks than the written report.
- A short viva at the end checks that they understand their own work.
Every term, we test these assessments against the newest AI models and against what teachers saw in class. We then rewrite any part that AI can now complete. We call them self-healing assessments.
Published research on assessment
- University of Sydney: two types of assessment. Secure assessments happen in person and check what students have learned. Open assessments allow AI and teach students to use it well. 28 29
- The AI Assessment Scale: gives each task one of five levels, from “No AI” to “AI Exploration”. 33
- MIT Teaching + Learning Lab: advises teachers to “focus on the process”, not only the final product. 30
Example: one electronics assignment, redesigned in five steps
Choose each step in turn to see how the task and the marks change
Step 1: The old task
The task: Design a low-pass filter that meets the given specifications, and submit a written report.
What AI can do here: AI can complete this task in seconds, including the report.
Step 2: Use your own lab
The task: Design the filter using only the components and instruments in our lab. Use component values you can actually get, not ideal values.
What AI can do here: AI can suggest component values, but it does not know which components your lab has.
Step 3: Build and measure
The task: Build the circuit on a breadboard. Measure its response on the oscilloscope and record the readings.
What AI can do here: AI cannot produce these readings. They come from the student’s own circuit.
Step 4: Analyse with AI, and record it
The task: Use AI to help analyse your lab readings. Keep a record of your prompts, and explain why your circuit behaves differently from the ideal design.
What AI can do here: AI may help with the analysis. The record of prompts shows the teacher how it was used.
Step 5: Defend it in a viva
The task: In a short viva at the lab bench, answer questions such as: Why did you choose this cutoff frequency? What happens if the load changes?
What AI can do here: The viva checks that the student understands the circuit they built.
How the marks are divided
Practical lab workViva and record of promptsWritten report
This is one possible way to divide the marks. Each teacher decides their own.
10Ask for help
Colleges can ask for outside help.
Universities around the world publish what has worked for them. Few colleges have the time to read it all, test it in their own classrooms and adopt it within a semester. Experts from industry and startups who work with AI can help a college move faster. A good partner leaves the teachers able to continue on their own.
If you work with an outside partner, check that:
- They work with your own courses and syllabus.
- Your teachers keep everything that is built, so your institution owns it.
- They measure results before and after the programme.
- They do not tie you to paid tools that you must keep buying.
What other universities are doing
- University of SydneyUses two types of assessment. Secure, in-person assessments check what students have learned. Open assessments allow AI and teach students to use it. Since Semester 1 of 2025, AI is allowed by default, except in exams and in-semester tests.
- MITHas no single rule for the whole institute in Fall 2026. Instead, it recommends that every course publishes a clear AI policy, and that teachers say how they use AI themselves.
- StanfordOffers a free AI Teaching Guide in six short modules. It covers AI literacy, writing a course AI policy and using AI in assignments. It includes kits for running workshops with colleagues.
- The AI Assessment ScaleA scale with five levels, from “No AI” to “AI Exploration”. Teachers use it to state how much AI each assessment allows.
- IISER PuneAn example from India. The Senate of IISER Pune approved its own rules for generative AI in March 2026.
11Check your readiness
How ready for AI are you, your department or your institution?
- Answer for yourself as a teacher, or for your whole department or institution.
- Ten questions, one at a time. It takes about two minutes.
- You see your score out of 100 straight away. If you want the next steps, we send them by email.
Your answers stay in your browser unless you ask for the email. If you have questions, write to support@deployproai.com.
Check your AI readiness
There are ten questions, and the check takes about two minutes. You get a score out of 100.
Question 1 of 10 Policy
Do your students know when they may use AI in your course, and how to say that they used it?Does every course in your department tell students when they may use AI, and how to say that they used it?Does every course in your institution tell students when they may use AI, and how to say that they used it?
Do you keep student data and unpublished exam papers out of public AI tools?Do teachers in your department know never to paste student data or unpublished exam papers into a public AI tool?Do teachers in your institution know never to paste student data or unpublished exam papers into a public AI tool?
Does at least one assessment in your course happen in person, such as a viva, a lab task or a live task?Does every course in your department have at least one assessment done in person, such as a viva, a lab task or a live task?Does every course in your institution have at least one assessment done in person, such as a viva, a lab task or a live task?
This term, have you tried your take-home assignments with current AI tools, to see whether AI can complete them?This term, has anyone in your department tried the take-home assignments with current AI tools, to see whether AI can complete them?This term, has anyone in your institution tried the take-home assignments with current AI tools, to see whether AI can complete them?
Do you know the mistakes AI makes in your subject, and do you check for them?Do teachers in your department know the mistakes AI makes in their own subject, and check for them?Do teachers in your institution know the mistakes AI makes in their own subject, and check for them?
Do you use AI with a set process for at least one routine task, such as question banks, rubrics or feedback?Have most teachers in your department built at least one AI workflow for their own teaching?Have most teachers in your institution built at least one AI workflow for their own teaching?
Can your students use an AI assistant built from your own course material?Can students in your department use AI assistants built from their own course material, not only public chatbots?Can students in your institution use AI assistants built from their own course material, not only public chatbots?
Do your students practise checking AI answers and defending their own work?In every year of study, do students in your department learn to check AI answers and defend their own work?In every year of study, do students in your institution learn to check AI answers and defend their own work?
Do you have time set aside, and someone to ask, when you learn new AI tools?Is one person or team responsible for AI in your department, with time and a budget?Is one person or team responsible for AI in your institution, with time and a budget?
Do you share what you learn about AI with colleagues, and learn from theirs?Does your department bring in outside help, such as training or expert partners, when it needs it?Does your institution bring in outside help, such as training or expert partners, when it needs it?
Your scoreYour department’s scoreYour institution’s score
0out of 100
- Policy0
- Assessment0
- Teaching0
- Students0
- Support0
Start here:
तमसो मा ज्योतिर्गमयLead me from darkness to light.
Check your inbox.
We have sent your five steps to . If it has not arrived after a few minutes, look in your spam folder.
For anything else, write to support@deployproai.com.
For help, talk to us about your institution, read about our AI programme for teachers and what each teacher builds in it, or see our programmes for students.
Questions we are asked
AI in colleges: frequently asked questions
What is the role of AI in education?
Many students already use AI as a tutor. In a global survey by the Digital Education Council, 86% of students said they use AI regularly in their studies. Each institution should decide where AI helps learning, where it must not be used in assessment, and how students learn to check its answers.
Should colleges ban AI tools like ChatGPT?
A ban is hard to enforce, and students then learn to use AI badly on their own. Some universities, such as the University of Sydney, now allow AI by default in open assessments and use secure, in-person assessments to check learning. Every course should state clearly when AI is allowed and how students should say they used it.
How do you assess students when AI is available?
Mark the process and the student’s own understanding, not only the final document. Use staged drafts, records of prompts, lab work done in person, and a short viva. Rewrite learning outcomes so that students must choose, find mistakes, critique and defend their work.
What should a university AI policy include?
It should state the purpose of AI in the course, when AI is allowed, how students should declare and credit its use, and what data must never be pasted into public AI tools. It should also set the AI rule for each assessment. Under India’s Digital Personal Data Protection Act, student data needs particular care.
What is an AI tutor for students?
An AI tutor is an assistant built from a teacher’s own notes and rules. It answers students’ questions when the teacher is not available. It should answer only from the course material, say clearly when a question is outside that material, and send questions about marks or personal matters back to the teacher.
How does AI change Bloom’s taxonomy?
AI can now do the lower levels of Bloom’s taxonomy, remembering and understanding, almost instantly. So marks for recall alone no longer show what a student has learned. Teachers should move the marks towards the higher levels: analysing, evaluating and creating.
Why is AI in education different in India?
The main difference is scale. India had 4.5 crore students in higher education in 2023-24, about 2.4 times as many as the United States, with 26 students for every teacher against an OECD average of 15. The economy needs about 78.5 lakh new non-farm jobs a year until 2030. Government programmes exist, but they still have to reach every college and every teacher. AI used well by teachers can help with this.
Do non-engineering teachers need AI training?
Yes. In every subject, AI can now draft part of the work, and students still need to show their own judgement. Teachers in humanities, management, HR and media need to know where AI makes mistakes in their subject and how to redesign their tasks.
Is there a faculty development programme on AI for Indian colleges?
Yes. DeployProAI runs a six-day, hands-on AI faculty development programme. Each teacher builds an AI policy for their course, a redesigned assessment, an updated module, an Academic Co-Pilot and a student assistant. The programme is eligible as an AICTE faculty development programme.