Wraz z dynamicznym rozwojem sztucznej inteligencji, przed badaczami UX stają nowe, prowokujące pytania. Czy tworząc produkty dla ludzi, wciąż musimy badać wyłącznie ludzi? Czy modele językowe, symulujące ludzkie zachowania i odpowiedzi, stają się (lub mogą stać się) naszym nowym narzędziem badawczym? Czy rozmowa z AI może być cenniejsza niż z rzeczywistym człowiekiem? A może to wszystko bujda?
Podczas tej prelekcji opowiem Ci o syntetycznych personach i cyfrowych bliźniakach. Pokażę w jakich sytuacjach i na jakich etapach projektu możemy je wykorzystać, jakie dają efekty, a także gdzie (na razie) leżą granice tej technologii.
Dowiesz się:
- Co to są i jak tworzyć wiarygodne, syntetyczne persony oraz kiedy warto ich używać, w tym jakie są do tego narzędzia
- Jak wypadły syntetyczne persony na tle prawdziwych użytkowników XTB (ludzi) pod kątem ich odpowiedzi oraz co z tego wynika
- W jakich sytuacjach AI jest świetnym uzupełnieniem badań, a w jakich nie zastąpi tradycyjnych, pogłębionych wywiadów.
Wpadnij posłuchać i zdecyduj czy syntetyczne persony mogą być przyszłością badań oraz zastosowań w Twoich projektach. Gorąco zapraszam!raw
Watch this WaysConf session, then continue with related talks or explore the current programme.
This Polish-language talk tests whether synthetic personas can replace human participants in product research. Using customer segments and prior evidence from an investment platform, the speaker compares language-model interviews with what real customers said about motivation, chatbots, and interface placement. The synthetic responses are coherent but often generic, overly rational, and weak at revealing context, emotion, contradiction, or product-specific behavior. The conclusion is deliberately bounded: synthetic personas can generate hypotheses and support quick, broad exploration, but decisions with product consequences still require human evidence.
Useful simulations require credible segments, research evidence, and contextual inputs; a prompt cannot recover customer reality that the organization never captured.
Watch from 6:30The synthetic answers sounded orderly and reasonable but missed the personal stories, tensions, and unexpected motives that made human interviews informative.
Watch from 10:20Real participants often say one thing and describe behavior that complicates it, while synthetic personas tend to produce internally consistent, sanitized explanations.
Watch from 14:09Questions about chatbot expectations and placement showed that general model knowledge could not reliably substitute for observing how customers use a specific product.
Watch from 18:00Synthetic participants are most defensible for fast exploration of broad needs and motivations, followed by verification with people before making product decisions.
Watch from 21:59These concise notes were edited from automatic captions and checked against the talk structure. They are not a verbatim transcript.
The experiment began with existing customer segmentation and organizational knowledge rather than an invented character description. That foundation mattered because a language model can only simulate the context it receives or has generalized from elsewhere. Teams with weak segmentation and little primary evidence should be especially cautious: fluent answers may conceal how little product-specific knowledge supports them.
Watch from 6:30When asked why they invested, synthetic customers offered sensible statements about growing capital and reaching financial goals. Real participants supplied more distinctive reasons, experiences, and emotional texture. The comparison showed that plausibility is not the same as insight. A response can sound correct while contributing little that changes the team’s understanding.
Watch from 10:09Synthetic profiles tended to explain their needs in a stable, orderly way. Human accounts were less clean: declared preferences, remembered behavior, and immediate reactions could conflict. Those contradictions are not noise to remove. They often point to unmet needs, situational constraints, or language that a product team would never discover from a perfectly consistent simulated respondent.
Watch from 14:00The question of where a chatbot should appear moved the conversation from broad attitudes to a concrete interface. Here, the synthetic rationale remained generic and could not establish how actual customers would notice, interpret, or use the control in context. The exercise illustrates why a model-generated suggestion can inspire a test but should not serve as the test result.
Watch from 18:00The speaker’s conclusion is not that synthetic personas are useless. They can support quick exploration of general motivations, help teams find questions, and offer starting points when evidence is limited. Their responsible role ends before high-confidence product decisions. Once consequences depend on context, behavior, and emotion, the organization must return to research with people.
Watch from 21:50
