Gains:
- Being able to distinguish where in the UX workflow (research synthesis, persona, wireframe, interface text, test analysis) artificial intelligence saves real time, and where responsibilities such as listening to the user and the final design decision remain with the designer, according to the task risk level
- Ability to develop a reflex that connects each AI output to real user data and evidence, and verifies hallucinations and bias through probing steps
- Ability to anonymize user conversation records and personal data within the scope of KVKK and transfer them to artificial intelligence within the limits of consent
User experience (UX, that is, the experience of the person using a product from start to finish) design has been proceeding with the principle of "first understand the person, then solve it" for years. When artificial intelligence enters this cycle, most people's first question is: "Will artificial intelligence do the design for me?" Short answer: no. Artificial intelligence is a design assistant; It produces drafts, summarizes data, multiplies options and speeds up repetitive tasks. But it always belongs to the designer to really listen to the user, decide what is the right solution, and bear the responsibility for this decision. This first unit clarifies where to put AI in the UX workflow, where you should stop, and how to move user data safely.
Where does AI come in handy in the UX workflow?
A UX project progresses through roughly five stages: discovery (research), define (persona and problem), produce (flow, wireframe, prototype), test (usability), and deliver (design system, documentation). Artificial intelligence plays a different role in each of these stages.
- Discovery: Summarizes dozens of interview transcripts (transcribed recordings) in minutes instead of hours, extracting initial themes. But it is up to the individual to decide which theme is important.
- Description: Produces Persona (fictional profile representing a specific user type) and journey map drafts. But it is your job to ensure that these drafts are based on real data.
- Generation: Wireframe (the box-line skeleton of a screen), flowchart, and interface text reproduce variations.
- Testing: Obtains findings from test records and ranks them in order of importance.
- Presentation: Writes component documentation and usage rules.
The common thread is this: AI generates drafts and options; Judgment and approval remain with you.
Tip: Use artificial intelligence not to "give me the answer" but to "give me options quickly and let me choose." The greatest efficiency comes when you delegate the preparation work, not the decision.
Mission risk level: where free, where cautious?
Not every task carries the same risk. A simple rule of thumb: the more directly the output affects a user, the higher the risk.
Quest
Risk level
The role of artificial intelligence
human approval
Meeting note summary
low
Almost fully transferred
light skimming
Generating a wireframe variation
low-medium
Generates draft
Designer selects and fixes
Creating a research theme
medium
Pre-synthesis
Verification by quotation is required
Interface text (error message, etc.)
medium-high
produces variation
Tone, accuracy, law check
Accessibility confirmation
high
pre-scan
Real test mandatory
Determine the real needs of the user
highest
can't help
All in the designer
This chart is a compass: as you move up, you decrease your trust in the AI and increase your own validation.
Hallucination and bias: two fundamental limits
You must internalize the two blind spots of artificial intelligence from the very beginning.
Hallucination (fabrication): A large language model (possibly word-guessing AI trained on large amounts of text) can produce fluent but actually inaccurate information. For example, it says "Users asked for speed in 20 conversations", but in reality there is no such quote. No matter how confident the output appears, it is not accurate without linking to the source.
Bias: The model carries biases inherent in the data on which it is trained. Clichés such as "older users are afraid of technology" can creep into the persona or script. These sentences are unsupported generalizations and should be tested with real data.
Caution: Fluency and accuracy of AI output are not the same thing. The most dangerous output is the one written in wrong but perfect sentences.
User data, KVKK and consent
UX research works with personal data: conversation logs, name, workplace, screenshots, even audio. Uploading this data to an arbitrary artificial intelligence tool may be a violation of KVKK (Personal Data Protection Law). The rule of thumb is three steps: anonymize, stay within consent, use agency-approved tool.
- Mask direct identifiers such as name, phone, email, company name (e.g. “Participant 3”).
- Do not exceed the scope of consent given to the participant during the research; If you didn't say "it will be processed into artificial intelligence", be cautious.
- In a corporate environment, choose tools that are clear where the data is stored and whether it is used for model training.
three mini cases
Case 1 — Time saver, in the right place. A designer had the AI summarize 18 interview transcripts (62 pages in total). Pre-reading, which would have taken 2 days by hand, was reduced to 40 minutes. But the designer matched each theme in the output to the actual quote and found that 3 themes had no basis. Result: speed + reliability.
Case 2 — Unverified output entered the presentation. Another team included the sentence "70% of users prefer mobile" produced by artificial intelligence in the report. When asked about the source, it was revealed that there was no such data; the model had made up the number. Stakeholder trust has been damaged. Lesson: numbers and claims do not spread without evidence.
Case 3 — Risk of personal data leakage. An intern uploaded the audio recording, along with participant names, to a free site that automatically transcribes it. The institution's security team marked this as a KVKK risk and the process was stopped. Lesson: tool selection and anonymization are done from the beginning.
Copiable prompts
Your role: senior UX research assistant. Task: Extract up to 6 themes from the anonymized interview transcript below. Rules: Add AT LEAST one verbatim quote next to each theme, with the tag [Participant X]. Do not produce a theme for which you cannot find a quote. When in doubt, write "weak support". Transcript: <<text>>
Check this AI output: Check for a factual basis in the text for each item in the theme list below. Mark those that have no basis as "NOT VERIFIED" and write why you are suspicious.List: <<themes>> Source text: <<transcript>>
Find all personal/identifying data (name, phone, email, company, address) from the text below and replace with [MASKED]. Also provide the change list as a table. Don't delete any information, just mask it.Text: <<dump>>
Classify a design task: Determine the risk level (low/medium/high) for "<<task>>", write the role of AI and the human approval required in one sentence. Explain your reasoning.
Weak prompt / Strong prompt
Weak: "Summarize these conversations."
Result: An unsourced, generic and unverifiable text; It is not clear which sentence came from where.
Strong: "Extract up to 6 themes from this anonymized transcript; add at least one verbatim quote tagged [Participant X] to each theme; do not produce a theme for which you cannot find a quote."
Result: An evidence-based, auditable, decision-friendly outcome.
The difference is in one sentence: strong prompt gives context + rule + proof requirement.
Common mistakes
- Moving the output from validation to decision. The most common and most expensive mistake; One contrived theme can distort the entire design direction.
- Uploading personal data without anonymizing it. KVKK violation and loss of corporate trust.
- Making artificial intelligence "decide". "Which design is better?" The answer to the question lies in the product, the user and the data; not in the model.
- Not reading the vehicle's data policy. Entering critical data without knowing whether your data goes to model training or not.
- Mistaking fluency for accuracy. A well-written sentence does not mean a correct sentence.
In summary
AI is an accelerator and blueprint partner in the UX workflow; It undertakes the preparation work at every stage, from discovery to presentation. But judgment, decision and responsibility remain with the designer. "How much does the risk of the task affect the user?" Measure with the question; increase your verification as the risk increases. Hallucination and bias are two permanent limits: back every claim with evidence, test every cliché with data. Transport user data by anonymizing it within the framework of KVKK and consent. These reflexes are the safe basis for every technique in the rest of the module.
Application task
- List 5 design tasks from your own current project (or a fictional project).
- Place each task on the risk table above: low, medium, high.
- For each task, write one sentence “role of AI” and “human approval required.”
- If you have an interview transcript, anonymize it with the third prompt and review the masking table.
- Plot the result in a table and explain in a paragraph why the highest risk task requires human approval.
checklist
- [ ] I positioned the AI as a draft generator, not as a decision maker.
- [ ] I determined the risk level of each mission.
- [ ] I have adopted the habit of linking outputs to evidence.
- [ ] I have not uploaded personal data to any tool without anonymizing it.
- [ ] I checked the data policy of the vehicle I am using.
- [ ] I added hallucination and bias checking to my workflow.