What AI Can and Cannot Do for UX Research

The biggest mistake I see teams make with AI is treating it like a magic black box. They throw unstructured data in and expect coherent, reliable insights to come out. This is dangerous. Not because the technology is bad, but because the expectation is wrong.

Large Language Models like ChatGPT, Claude, and Gemini have become standard tools in the research stack. The hype cycle has settled. What remains is a more important question: do you actually understand what you are working with?

The Current State

The investment phase is over for teams that started early. They have working workflows, established validation practices, and are realizing measurable ROI. For those still experimenting or sitting on the sidelines, the gap is becoming a career risk.

The divide is no longer "AI enthusiasts vs. skeptics." It is "AI-fluent researchers vs. researchers falling behind." This is not hype. It is observable in hiring patterns, team structures, and the expectations placed on individual researchers. One researcher with strong AI fluency now covers ground that used to require two (see Building a Research Career in the Age of AI for the implications).

The loudest evangelists and skeptics have mostly moved on. The real conversation is now about reliability, governance, and workflow integration, the unglamorous work of making AI consistently useful under real-world conditions.

For a structured framework to evaluate AI tools based on these capabilities, see Evaluating AI Research Tools.

What LLMs Actually Are

The "T" in GPT stands for Transformer ^[2]. This is not just a technical term. It is the most useful description of the technology's core function.

An LLM is not a knowledge-generation machine; it is a concept-transformation engine. It is exceptionally good at taking information in one format and structuring or rephrasing it into another. It is less a generator of new facts and more a manipulator of existing concepts.

At its core, an LLM predicts the most probable next token based on patterns learned from vast amounts of text. This predictive nature is why models can "hallucinate": they generate plausible-sounding text, not verified facts. Hallucinations are still present in 2026, but they are better understood and manageable with proper workflows: structured prompts, validation steps, and cross-model verification.

What Has Changed Since 2024

The core architecture is the same, but the capabilities have expanded significantly:

• Multimodal input: Models now process images, video, and audio natively, not just text. This opens research applications from screenshot analysis to session recording review.
• Agentic AI: Models can now orchestrate multi-step tasks autonomously: running analysis pipelines, calling tools, and making decisions within defined constraints.
• Larger context windows: Models routinely handle hundreds of thousands of tokens, making it feasible to process full transcripts, entire codebases, or complete datasets in a single pass.
• Improved reasoning: Chain-of-thought and structured reasoning capabilities have improved substantially, making complex analytical tasks more reliable.

These are real advances, not incremental. But they do not change the fundamental nature of the technology: it remains a transformation engine, not a judgment engine.

What the Research Shows

A study presented at CHI 2024 ^[1] provides a useful snapshot based on early-2024 models:

When experienced UX professionals evaluated usability issues suggested by an LLM:

• They agreed with 77% of the issues the AI found
• But the AI missed around 60% of the unique problems that human experts identified

Models have improved since this study was conducted. The exact numbers have likely shifted. But the fundamental pattern holds: AI is strong on known-pattern issues and weak on contextual, novel problems. The directional insight matters more than the specific percentages.

This is a clear signal of AI's proper role: assistant, not replacement.

AI excels at identifying common, pattern-based issues because it has been trained on vast amounts of data reflecting these known problems. It is less effective at:

• Uncovering novel issues
• Understanding deep contextual nuance
• Identifying subtle emotional reactions that a human observer would catch

The Evolving Role of the Researcher

These limitations highlight the evolution of the researcher's role. The value of the human researcher shifts away from the tedious work of raw analysis and toward the strategic work that AI cannot do:

Strategic Framing

Asking the right questions and designing sound research in the first place. AI cannot tell you what questions matter. That requires understanding the business context and user landscape.

Critical Validation

Questioning the AI's output, spotting its biases, and separating signal from noise. The AI produces drafts; you produce judgment.

Foundational models are often trained to be helpful and agreeable, a trait known as "sycophancy". To get objective results, you must turn an agreeable assistant into a critical sparring partner.

Influential Communication

Translating findings into clear, actionable recommendations that drive business decisions. The political and organizational skill of getting insights implemented remains distinctly human.

For how these changes shape career development in research, see Building a Research Career in the Age of AI.

Best Use Cases for LLMs in Research

Based on experience, here are the tasks where LLMs provide the most reliable value:

For advanced prompting, RAG, and fine-tuning techniques, see Advanced AI Techniques for Research.

Practical Workflow: Thematic Analysis with an LLM

Here is a concrete, step-by-step workflow for one of the most common AI-assisted research tasks: thematic analysis of qualitative data.

Step 1: Prepare Tidy Data

The biggest mistake is feeding unstructured transcripts into an LLM. Instead, use "Tidy Data" principles. Create a simple table where every row is a participant quote and columns represent metadata (participant ID, task context, timestamp). Anonymize all PII (Personally Identifiable Information) before upload.

Step 2: Engineer a Structured Prompt

Do not ask the AI to "find insights." Give it a mechanical task with explicit constraints:

• Role: "You are a meticulous UX Researcher."
• Task: "Categorize each user quote based on the taxonomy provided below."
• Taxonomy: Provide strict definitions (e.g., "Usability," "Feature Request," "Trust/Security").

Step 3: The Committee of Raters

To increase reliability, use multiple models (e.g., GPT-4 and Claude) as a "Committee of Raters." Feed them the same data and prompt.

• Where they agree, you have high confidence.
• Where they disagree, you have a signal for nuance that requires human review.

This approach mirrors traditional inter-rater reliability practices in qualitative research, using AI disagreement as a flag for human attention rather than a failure.

For the detailed four-step workflow that operationalizes this approach, see AI-Assisted Thematic Analysis: A Practical Workflow.

Step 4: Human Validation (The Nuance Check)

The AI sees text; you saw the session. Perform a "Nuance Check" on the output:

• Sarcasm: Did the user say "Great job" with an eye-roll? AI will tag that as "Positive Sentiment." You must correct it.
• Silence: Did the user hesitate before clicking? AI cannot see silence.
• Context: Did the user's frustration stem from the interface or from an unrelated interruption during the session?

Beyond Manual Pipelines: Agentic Workflows

In 2026, the frontier is agentic workflows: the model orchestrates multi-step analysis pipelines: ingesting data, applying taxonomies, flagging disagreements, generating draft reports. This is powerful and real. Teams are building pipelines where an AI agent processes a full dataset end-to-end, from raw transcripts to structured findings.

But here is the critical point: the more autonomous the workflow, the more important human validation becomes, not less. Autonomy without oversight is not efficiency. It is risk accumulation. Every step the model executes without a human checkpoint is a step where errors compound silently.

The four-step workflow above is still the right mental model. Agentic tools execute it faster, but they do not change the logic. You still need tidy data, structured prompts, cross-validation, and human judgment. The difference is speed, not substance.

What AI Cannot Do, And Probably Will Not

Some limitations are temporary: better models, more data, improved architectures. But the following are not gaps that future models will close. They are structural features of what research actually is: a human practice that involves judgment, ethics, and relationships.

Ethical Judgment in Consent and Risk Decisions

When a participant seems distressed. When data feels too identifiable to use even after anonymization. When the study design creates risk that the protocol did not anticipate. These require human moral reasoning, the kind that weighs competing values, not competing probabilities.

Navigating Organizational Politics

Getting research funded. Making findings stick. Building the relationships that turn insight into action. These are influence skills exercised in hallways, meetings, and one-on-one conversations. AI does not navigate power structures.

Knowing What NOT to Research

Deciding which questions do not deserve investment right now requires business context, strategic judgment, and an understanding of organizational capacity that AI does not have. The ability to say "not now" is as valuable as the ability to say "here is what we found."

Building Trust with Stakeholders

Trust is built through repeated human interaction, demonstrated credibility, and track record. Better outputs help, but they are not the mechanism by which trust is earned. Stakeholders trust people, not models.

Reading the Room During Live Research

The pause. The eye-roll. The group dynamic shift. The participant who says "it is fine" while their body language says the opposite. AI processes text and audio with increasing sophistication, but it does not truly observe in the way a present human researcher does.

The Brand Connection: While AI can measure "ease of use" and even "usefulness" if fine-tuned correctly, the genuine emotional connection between a consumer and a brand remains elusive to algorithms. Brand loyalty isn't just clicks and sentiment scores; it's often a visceral reaction that relies on the researcher's "gut feeling"—a synthesis of micro-expressions and unstated history—that an LLM cannot yet reach.

What This Means for Practice

The goal is not to replace your judgment with AI but to use AI to amplify your judgment. The most effective researchers will be those who:

1. Understand what LLMs are actually good at (transformation, not generation)
2. Provide structured inputs that play to those strengths
3. Maintain rigorous human oversight of all outputs
4. Focus their own energy on the strategic work AI cannot do

This is not about learning a specific tool. Tools will change. It is about learning a way of thinking about human-AI partnership that will outlast any particular model or platform.