Faculty Development Program · Day 1 · Afternoon

Using AI for Teaching
& Research

Foundations and prompting. Fewer tools, used well — a method you can repeat on your own course, not a tour of everything.

Srikanth Doddi · Architect, Digital Engineering · OSI Digital M.Tech, Data Science — BITS Pilani

AI Tools for Teaching, Learning & Research · JNTUK Kakinada & Sri Vasavi Engineering College

The promise

What you'll walk out with

01

One usable artifact

A quiz, an in-class activity, or an explanation built from your own course — not a demo.

02

A repeatable method

A four-step way of working with AI that holds up across any subject you teach.

03

The judgment to trust it

When to rely on AI, when to verify, and what never to paste into a free tool.

Today is not a tour of every AI tool. It is how to use a few of them well.

Roadmap · two sessions, one thread

The path through the afternoon

Session 3 · 2:00–3:30

Foundations & tools — what these tools are, where they fail
The landscape — which tool for which job
Prompting — the skill that makes any of them work
Live demo + the verification habit

Session 4 · 3:45–4:30

Innovative teaching — class worth showing up for
Three classroom moves you can steal
You build one for your own course
Feeds today's post-session test

High tea 3:30–3:45 · The thread: the method you learn before tea is the method you use after it.

Orientation

AI, ML, GenAI — and what changed

Artificial Intelligencethe whole field

⊃

Machine Learninglearns from data

⊃

Generative AIChatGPT · Claude · Gemini

Core AI / Machine Learning

Predicts & classifies. Learns from labelled data to do one narrow job — spam filters, recommendations, forecasts, face unlock.

Generative AI · 2022–23 →

Generates new content. One big model, many tasks — just asked in plain language. The chatbots on your screen.

Models got big enough to follow plain-English instructions — that's when "AI" became something you could just talk to.

Orientation · how we got here

One idea in 2017 turned AI into ChatGPT

2017 · The paper

Attention Is All You Need

Vaswani et al., Google introduce the Transformer — a model that reads a whole sentence at once.

The idea

It learns what to pay attention to

Every word looks at every other word at the same time. This one mechanism sits under all of today's models.

2018–2020 · Scaled up

Trained on the whole internet

GPT · BERT · GPT-3. Stack the Transformer huge and the model becomes general — not built for one task.

2022 · The chat box

Then OpenAI put a chat box on it

Tuned to follow instructions and converse — the same idea, now something you can just talk to. ChatGPT ships.

Five years later

100 million users in two months

The fastest-adopted app in history. The research was public since 2017 — it finally reached everyone.

Attention Is All You Need

Vaswani · Shazeer · Parmar · et al. — Google, 2017

ChatGPT

Explain photosynthesis to a 10-year-old.

Sure! Think of a plant as a tiny kitchen…

The real breakthrough was 2017 — not 2022.

2017

2022 01 / 05

→ or click anywhere to reveal each step

Machine learningHow an LLM predicts the next word

Shared vocabulary · every branch

The words you'll hear — in plain English

LLM

A giant auto-complete: predicts the next word, over and over. Fluent — but not a fact database.

Token

The chunk of text models read & write in. Limits and cost are counted in tokens, not words.

Prompt

What you type — your question plus instructions, examples and pasted material.

Context window

Its short-term memory — how much text it can hold open at once.

Hallucination

A confident, plausible answer that is simply false — invented citations, numbers, formulas.

Grounding

Tying the answer to real sources you give it, instead of its own memory.

RAG

“Chat with your PDF”: fetch the right passages first, then answer from them.

Training cutoff

The date after which it knows nothing new — switch on web search for recent things.

Multimodal

Handles images, diagrams, audio — not just text. Photograph a derivation and ask.

Agent

AI that takes multi-step actions — searches, runs code, uses tools — not just answers.

Reasoning model

A “think step by step” mode — much better at maths, proofs, multi-step problems.

Temperature

A randomness dial: low = focused & repeatable, high = varied & creative.

You don't need the maths — just these words to follow along and ask better questions.

The bigger landscape

What people actually do with GenAI

Everyday assistant

Draft, explain, brainstorm, plan, in a chat. ChatGPT · Claude · Gemini

Research & reading

Summarise papers, query your own PDFs, find sources. NotebookLM · Elicit · Perplexity

Writing software

Generate, explain and debug code, fast. Copilot · Cursor · Claude Code

Building products

Apps on top of AI — chatbots over your own data (RAG), copilots, agents.

Content & media

Slides, images, video and voice from a prompt. Gamma · Midjourney · HeyGen

Automating work

Agents that take actions across your tools. agents · workflows

For teaching and research, the top row is almost all your value. The rest is good to recognise, not to master.

Everyday assistant

Ask in plain language — then iterate

Claude Opus 4.8

Explain recursion with an analogy

Picture two mirrors facing each other — each reflects the next, smaller and smaller, until the image is too tiny to see.

ChatGPT GPT-4o

Draft a polite leave email

Subject: Leave request Hi [Manager], I'd like to request leave on the 14th–15th. Happy to hand over beforehand.

Gemini 2.5 Pro

Summarise this page in 3 points

1. Core thesis stated up front. 2. Three pillars of evidence. 3. A clear call to action.

Research & reading

Answers grounded in real sources

What does the evidence say about active learning?

Research

Searching sources

Grounded answer 4 sources

Across randomized studies, active learning cuts failure rates in STEM courses by roughly a third versus traditional lecture1.

Gains hold across class sizes and disciplines, with the strongest effect in small-group, problem-driven settings2.

Retrieval practice and spaced review further improve long-term retention of concepts3.

Sources

1

Active learning increases student performance in STEM

Freeman et al. · PNAS · 2014

2

Meta-analysis of small-group instruction outcomes

Springer, Stanne · Rev. Educ. Res. · 1999

3

Retrieval practice produces durable learning

Roediger, Karpicke · Psych. Science · 2006

4

Engagement and outcomes in large lectures

Deslauriers et al. · PNAS · 2019

Writing software

Code, generated and explained

quicksort.py ×

Claude Code

1 2 3 4 5 6 7 8 9 10 11 12

# quicksort — sorts a list in place

def quicksort(arr, lo=0, hi=None):

if hi is None: hi = len(arr) - 1

if lo >= hi: return arr

pivot = arr[(lo + hi) // 2]

i, j = lo, hi

while i <= j:

while arr[i] < pivot: i += 1

arr[i], arr[j] = arr[j], arr[i]

quicksort(arr, lo, j)

return quicksort(arr, i, hi) or arr

# sorted = quicksort([5, 2, 9, 1, 7])

generated quicksort.py 12 lines · 3 added Done

Content & media

Slides, images & voice from a prompt

Content & media studio

Generating

Image 1024 × 1024

Slides 3 frames

Voice Neural TTS

Video 0:08 / 0:12

The honest part

Strong at language. Weak at facts.

Reliably good at

Drafting, rephrasing, summarising
Explaining a concept five different ways
Structuring: outlines, rubrics, question banks
Brainstorming and changing tone or level

Unreliable at

Facts, dates, numbers, prices — often confidently wrong
Citations — it invents plausible papers that don't exist
Anything after its training cutoff
Knowing when it's wrong — it never sounds unsure

Rule of thumb: trust it for shape and language, verify it for facts.

The landscape

Three kinds of tool, not thirty

General assistants

Your everyday workhorse: draft, explain, restructure, brainstorm.

Claude ChatGPT Gemini Copilot Perplexity Grok

cite

Document-grounded · research

Answers only from sources you give it — the honest fix for made-up citations.

NotebookLM Elicit Consensus SciSpace

Production · media

Turn ideas into slides, images, video or voice.

Gamma Canva Midjourney HeyGen ElevenLabs

Dozens exist in each box — the highlighted few are all you need. Learn one general assistant deeply.

Tools that fit · general assistants

The everyday assistants — and the catch

ChatGPT

The everyday workhorse — notes, MCQs, explaining a concept at three levels.

Catch — free flagship (GPT-5.1) is capped, then silently drops to a weaker “mini”.

Claude

Long documents — read a whole paper or spec, clean structured output.

Catch — tightest free limits of the big three; the top Opus model is paid.

Gemini

Inside Google Workspace — Docs, PDFs & images without leaving Google.

Catch — free Gemini 3 Pro is throttled hard, then drops to a lighter Flash model.

Copilot

Quick web-grounded answers; Office drafting if your institution has M365.

Catch — the useful Office integration needs a paid Microsoft 365 licence.

Perplexity

Sourced answers with clickable citations you can actually verify.

Catch — deep “Pro Search” ~5/day on free — and still open the source.

Grok

Current, conversational, tied to X / live news.

Catch — ~10 messages per 2h on free; weak for citable academic facts.

DeepSeek

Strong free maths & coding; open weights you can self-host.

Catch — China-hosted app — keep private data out; self-host if sensitive.

Meta AI

Zero-setup answers & images right inside WhatsApp.

Catch — casual only — not for research or anything confidential.

Free tiers got genuinely strong in 2025–26 — reasoning models reached everyone. But flagship access is metered and silently drops to a weaker model — learn to notice when it does.

Tools that fit · content & media

Prompt → slides, images, voice, video

Gamma

Notes → a presentable, on-brand deck in minutes.

Catch — one-time 400 free credits; a “Made with Gamma” watermark on exports.

Canva Magic

Everyday visuals — handouts, posters, slides — zero design skill.

Catch — ~50 AI uses/month on free, shared across all its AI tools.

Napkin AI

Turn text into editable concept diagrams, fast.

Catch — template-driven; not for to-scale technical schematics.

Nano Banana Pro

Figures with readable text inside the image — the 2025 leap.

Catch — free but rate-capped; still check every label it writes.

Ideogram

Posters & title cards with correctly spelled text.

Catch — small free quota; free images are public by default.

HeyGen / Synthesia

A script → a presenter or translated lecture video.

Catch — free = a few minutes/month, watermarked; disclose AI use.

ElevenLabs

Professional voiceover without recording yourself.

Catch — ~10k chars/month free, non-commercial; clone voices only with consent.

Suno

Original background music for lecture videos.

Catch — free runs an older model; outputs non-commercial, attribution required.

Prompt to a deck, a labelled figure, or a voiceover in minutes — but free outputs carry watermarks, capped credits, and no IP guarantee. Check licensing before you reuse.

Start from the work, not the tool

If you're doing this, reach for that

If you're…	Reach for	And the move is
Doing research — literature, staying current	NotebookLM · Elicit · Consensus	Summarise & query papers, find sources — then verify every citation
Teaching — explaining a concept	Claude · ChatGPT · Gemini	Analogies, three levels of difficulty, worked examples, quiz banks
Making slides or handouts	Gamma (+ an assistant for the outline)	Outline → first-draft deck in minutes, then edit
Grading — giving feedback	A general assistant, anonymised	Draft rubrics & comment banks — never paste student records
Writing emails, reports, admin	Any general assistant	Fast first drafts you tighten yourself

Start from what you're doing this week — then pick the row. The names change; the jobs don't.

AI for research

Which AI tool fits which research job

★ Start here — NotebookLM: answers only from sources YOU upload, with click-through citations — it can't invent references like open-web chatbots (~13% vs ~40% hallucination). Best for your own corpus.

For this research job	Reach for	Why — and the catch
Literature review + data extraction	Elicit	Pulls methods, samples & findings into tables across ~125M papers — verify (2025: wrong counts)
Yes/no evidence question	Consensus	Synthesises 200M+ papers; the "meter" is vote-counting, not formal synthesis
Quick current scan, cited	Perplexity (Academic)	Inline citations, peer-reviewed filter — but ~37% had inaccuracies; click through
Map a field / seminal work	Semantic Scholar · Connected Papers	Citation graphs surface what keyword search misses — they map, not judge
Read a dense paper	SciSpace	Explains math, tables & jargon inline as you read — sanity-check hard detail
Draft / analysis code	Claude · ChatGPT · Gemini	Great for writing & code — NOT for finding real papers (fabricate ~12–50% of cites)

Grounded tools find & explain; general assistants write. Never trust an AI-supplied citation without checking the source.

AI for research

What faculty do — and where AI helps

Literature review & staying current

Surfaces relevant papers fast; maps a field's seminal work. Limit · coverage gaps — verify at source

Understanding dense papers

SciSpace & NotebookLM explain methods, math, tables, jargon. Limit · subtly wrong on hard detail

Finding gaps & framing questions

Citation-graph tools + grounded chat spot under-studied areas. Limit · the researcher judges novelty

Drafting & polishing manuscripts

Restructure, tighten, clarify — a leveller for non-native writers. Limit · keep your voice; check policies

Data analysis & coding

Generate & debug Python/R for cleaning, stats, plotting. Limit · validate logic — code can be wrong

Grant / proposal writing

Outline, draft and tighten aims, significance, impact. Limit · facts & budgets must be verified

Models invent fake citations (2025: GPT-4o ~50%+; Claude/Gemini ~12–20%; Perplexity ~37% off). Verify every reference; prefer grounded tools; never paste unpublished work into free accounts.

The skill that makes all of them work

Most people stop at the first answer.

That is exactly why AI training so often fails to stick.

What usually happens

Type one prompt
Take whatever comes out
Edit it by hand
Conclude "AI isn't that useful"

What actually works

Give it context: who, what level, how long
Read, then refine — "make it simpler"
Iterate two or three times
The third answer is the useful one

Live demo

Teaching IaaS vs PaaS vs SaaS

Course outcome CO2, Cloud Computing (V23AITPE06) — the concept students most reliably confuse.

01 · WEAK PROMPT

"Explain IaaS, PaaS, SaaS."

Output = textbook. Stop here and AI added nothing.

02 · ADD CONTEXT

"…for 2nd-year B.Tech, one running analogy, under 200 words."

Now it's a teaching tool.

03 · GENERATE

"5 MCQs on the PaaS/IaaS mix-up + a 10-min sort-the-services activity."

An evening of prep becomes three minutes.

04 · VERIFY & PROTECT

Ask for live AWS pricing → check the docs.

It answers confidently. Never paste student data.

Open the interactive explainer

Trust structure. Verify facts.

CSE · Cloud computingDive the cloud stack — SaaS → PaaS → IaaS

Civil · StructuresGolden Gate rigidity — a 3D walk-through

ECE · MicroprocessorsInside the 8086 — fetch · decode · execute

Mechanical · MachinesGear trains — trading speed for torque

The method travels

Same four steps — any discipline

Cloud / CS

"One analogy for IaaS vs PaaS vs SaaS, 2nd-year level" → quiz + sort activity.

Languages / Humanities

"Give three counter-arguments to this essay's thesis" → live debate students can't pre-fake.

Core engineering

"Three real-world failures caused by this concept, with the physics" → case discussion.

Context → refine → generate → verify. The subject changes; the method doesn't.

The verification habit

If it matters, check it at the source

If a fact, date, number, or citation matters — check it before it reaches a student.

Citations

Open the paper. If you can't find it, it probably doesn't exist.

Live data

Prices, specs, versions — confirm on the official page.

Surprising claims

If it sounds too neat, that's a prompt to double-check, not to trust.

One privacy rule

What goes in, and what never does

This program has no institutional AI accounts — so everyone is on a free, personal tool.

Safe to paste

Your own lecture notes and drafts
Public textbook concepts
Anonymous, made-up examples

Never paste

Student names, marks, or records
Unpublished research or others' work
Anything confidential to the institution

Prompting · the techniques

Eight moves that change the answer

Give it a role

“You are an examiner for 2nd-year ECE…”

Add context

Who it's for, their level, and how long you want it.

Show an example

Paste one MCQ: “make five more like this.”

Think step by step

For derivations & numericals, ask it to reason first.

Constrain it

“Under 150 words, one analogy, no jargon.”

Ground it

Paste the source: “answer only from this.”

Ask for structure

“As a table / a rubric / three bullets.”

Critique & refine

“What's weak here? now fix it.” — the third answer wins.

Most people use none of these and conclude “AI isn't that useful.” Stack two or three and the same model gives a far better answer.

Prompting · anatomy

Weak prompt → strong prompt

Weak

“Explain sampling.”

A generic textbook paragraph. Stop here and AI added nothing.

Strong

“You're teaching 2nd-year ECE who just met sampling. Explain aliasing with one real-world analogy, under 150 words — then 3 MCQs (4 options, answer marked, one numerical).”

A ready-to-use teaching asset, in one shot.

The five parts: Role · Context · Task · Format · Constraint. Add the ones you're missing and the answer transforms.

Make it reusable · skills

From prompts to skills — build it once

Re-prompt every time

Paste your course & level again
Re-explain your format and tone
Fix the same things every single week

Build a skill once

Save the role, rules, examples & files once
Invoke with a single line — your setup is baked in
Same quality, every time, in seconds

A skill turns your best prompt into a button you press again and again.

Make it reusable · build your own

Build your own — free, today

Gemini Gems

Save instructions + reference files as your own named assistant.

Free — free accounts can build them. The best free pick.

ChatGPT Custom GPTs

The largest ecosystem of shareable custom assistants.

Free — free to use; building your own needs Plus.

Claude Projects

A workspace that holds your instructions + documents for one recurring task.

Free — on paid Claude plans.

Claude Skills

The power-user version — a packaged procedure (with files/scripts) Claude follows.

Free — via Claude Code & the API; more technical.

Name it“MCQ Maker — DSP”

→

Role + rulescourse, level, tone

→

Your examplesyour exact format

→

Attach filessyllabus, rubric

→

Invokeone line, forever

Free tiers shift — confirm current limits. And a skill makes your setup reusable, not your facts correct: still verify, still no private data.

Make it reusable · for your course

Skills worth building this week

MCQ / Quiz Maker

Your format, difficulty and CO mapping — five questions on demand.

Concept Explainer

Your level and your running analogy — three ways to explain anything.

Rubric Grader

Anonymised work + your rubric → consistent comments and a draft mark.

Lab / Notes Drafter

Your template and depth → a first draft of a manual or unit notes.

Question-Bank Builder

By unit and Bloom level — grow a bank a few questions at a time.

Feedback & Email Writer

Your tone → fast first drafts of feedback, notices and replies.

Start with the task you repeat most. Build it once this week — reuse it all semester.

Close · if you remember only three things

Three things to carry out the door

01

Don't stop at the first answer

Context, then refine two or three times. That's where the value is.

02

Trust shape, verify facts

Great for structure and language. Always check numbers, dates, citations.

03

Protect what's private

No student data, no unpublished work, into any free tool.

Thank you — Srikanth Doddi · OSI Digital

Using AI for Teaching& Research