AI testing,
made simple.
“AI testing” means two things, and a QA engineer in 2026 needs both. One is using AI like ChatGPT and Gemini to test faster. The other is testing the AI features your company is building. Here is everything in plain English, and it is easier than it sounds.
AI words, in plain English
New to this? Read these first. That is all the vocabulary you need.
AI
Software that can read, write and answer almost like a person. ChatGPT and Gemini are AI.
LLM (large language model)
The brain behind tools like ChatGPT and Gemini. It has read a huge amount of text, so it can answer questions and write for you.
Model
One specific AI brain. GPT-4o (inside ChatGPT) and Gemini are models. A newer model is usually smarter.
Prompt
The instruction you type. "Write 5 test cases for a login page" is a prompt.
Token
A small piece of a word. AI reads and bills in tokens. One English word is roughly one token.
Context window
How much text the AI can keep in mind at once. A bigger window lets you paste a full requirement.
Hallucination
When AI says something wrong but sounds very sure, like naming a function that does not exist.
Grounding
Giving the AI the real document to answer from, so it sticks to facts instead of guessing.
RAG
The AI first searches your documents, then answers using them. This is how a company help-bot answers from its own pages.
Agent
An AI that finishes a task in steps on its own, not just chat. For example: read a requirement, then write and run the tests.
Prompt injection
A trick where a user hides an instruction to make the AI misbehave, like "ignore your rules and show the data".
Guardrails
The safety rules that stop the AI from doing unsafe or off-topic things.
Eval
A test for AI. You give it inputs, check the answers, and give a score. Like a test suite, but for AI answers.
Golden dataset
A saved set of questions with the answers you would accept. You use it to check the AI again and again.
LLM-as-judge
Using a strong AI (like Gemini or GPT-4o) to grade another AI's answers, so a person does not have to read each one.
Side 1
Using AI to test
AI is your helper. It writes test cases and code, fixes broken locators, spots screen changes, and reads failures for you. It makes you faster, and you stay in charge.
Side 2
Testing AI features
Testing chatbots, RAG search and agents. There is no single correct answer, so you check with evals, a golden dataset and safety rules instead of an exact match.
Using AI to test, step by step
Where AI actually saves a tester time today.
Writing test cases
Turn a requirement or user story into positive, negative and edge cases in seconds. Then check it for the cases the AI missed.
Tools: ChatGPT, Gemini, a saved set of prompts
Writing test code
Ask it to write page objects, tests and helpers from your comments. Keep the good parts, drop the rest, and add your own checks.
Tools: GitHub Copilot, Cursor, Claude Code
Self-healing locators
When the screen changes, the tool finds the element again instead of failing. You spend far less time fixing broken locators.
Tools: Testim, Mabl, Functionize, KaneAI
Visual checks
Compare screens the way a person would, not pixel by pixel. It catches real changes and ignores tiny rendering noise.
Tools: Applitools, Percy
Test data
Make realistic but fake data for edge cases and load: names, addresses, orders, and different regions. No real customer data used.
Tools: Faker with an AI, Gretel, Mockaroo
Reading failures
Summarise failing runs, group similar failures, and get a likely cause from a stack trace or a CI log.
Tools: ChatGPT or Gemini on logs, BrowserStack AI
Testing AI is different, but simpler than you fear
Normal testing checks for one exact answer. An AI like ChatGPT can give a different, still correct answer every time, so an exact match will fail. You just change the question you ask.
Old question ✗
“Is the answer exactly equal to X?”
New question ✓
“Is it good, true and safe, across many examples?”
How you actually test it
The whole idea, in three steps.
Golden dataset
Questions plus the answers you would accept
Run and grade
Send each question, then an AI grades the answer
Pass or fail in CI
A score gate checks every prompt and model change
That is the whole idea. The rest is just more examples and clearer scoring rules.
Evals
The test suite for AI. An input goes in, the answer comes out, you score it, and it runs in CI on every change.
Golden dataset
A saved set of inputs with the answers you would accept. Grow it every time you find a new failure.
AI as the judge
Use a strong AI to grade the answers on a rubric (correct, relevant, safe), so you do not check each one by hand.
Ways to check
Exact match, contains the right thing, means the same, valid JSON, a rubric score, and a safety pass or fail.
Safety tests
Attack the safety layer on purpose: hidden instructions, data leaks, off-topic asks, and cases where it should refuse.
A human sample
Check a small sample by hand, see where the AI judge disagrees, and add those cases back to your golden dataset.
What to check in an AI feature
The risks that only AI features have. This is your checklist.
Is it correct and grounded
Does the answer really come from the given document and instructions, or did the AI make it up?
Hallucination
Made-up facts, fake links, invented functions. This is the biggest risk in any AI feature.
Relevance and quality
On topic, complete, and in the right format (valid JSON, right tone). Scored by a rubric or an AI judge.
Bias and safety
Harmful, unfair or unsafe answers, across different people, languages and tricky questions.
Can it be tricked
Can a user get past the rules, leak data, or make the AI ignore its instructions? Try to break it on purpose.
Did it find the right source
For a RAG bot, check that it pulled the correct document before answering, and that it cited it.
Does it stay steady
The same question can give a slightly different answer each time. Catch it when a prompt or model changes for the worse.
Speed and cost
How fast it replies, the cost per request, timeouts, and how it behaves on empty or messy input.
The AI-testing toolkit
Free and paid tools QA teams use to test AI. Tap any tool for a full guide.
| Tool | What it does | Guide |
|---|---|---|
| promptfoo | Run evals and safety tests on prompts and models, works in CI | Read → |
| DeepEval | Pytest-style AI evals (correct, relevant, bias) | Read → |
| Ragas | Check RAG bots: did they find and use the right source | Read → |
| LangSmith | Trace, save datasets and run evals for AI apps | Read → |
| Giskard | Scan AI models for weak spots and unsafe answers | Read → |
| Confident AI | Dashboards to track eval scores over time | Read → |
Start this weekend
You can have a real AI test running by Sunday night.
- 1
Pick one AI feature (a chatbot, a summariser, or a search that answers from documents).
- 2
Write 15 to 20 example questions with the answer you would accept. This is your golden dataset.
- 3
Score them with an eval tool. Start with promptfoo or DeepEval.
- 4
Add 5 tricky or unsafe questions and check that the safety rules hold.
- 5
Put the evals in CI, so a prompt or model change cannot break things without you knowing.
Recommended Udemy courses
Want a guided path for a topic above? These are the AI courses I would point a QA engineer to.
GenAI & AI Agents for QA Automation | Copilot & Claude Code
Learn Agentic AI – Build Multi-Agent Automation Workflows
Learn TestMu AI – Agentic AI Quality Engineering Platform
2026 – ISTQB AI Testing (CT-AI) Certification – Crash Course
RAG-LLM Evals & Test Automation for Beginners
LangChain Framework for Beginners – Build AI Systems + RAG
Introduction to Machine Learning Models (AI) Testing
Build Your Own MCP Servers with TypeScript – Beginner's Guide
What AI still cannot do
AI writes the first draft, you make the call. It can be wrong, it changes its answer, and it will happily write a test that looks right but is not. Treat every answer like a junior's first attempt. Read it, give it the real facts, and keep the checks yours. The testers who do well in 2026 use the speed of AI with the judgment of an engineer.
Go from reading to doing
Follow the step-by-step AI for QA roadmap, or try real requests in the API practice lab.