How people really use AI
Eighteen peer-reviewed studies on what people actually type, where they give up, and what makes them trust a result — and what each one implies for a reference like this one.
The short version
People do not compose prompts. They collect phrases, pick from options, and stop after one bad result because they cannot tell whose fault it was. The most consistent finding across this literature is that stating a system's limits earns more trust than hiding them, and that the effect compounds over weeks rather than showing up on a first visit.
Every DOI was resolved against the Crossref API and cross-checked against OpenAlex on 2026-09-01: title, venue, year, authors and citation count. Abstracts were read where available to check that each claim matches what the paper actually studied.
All 14 DOIs from the brief resolve to real, peer-reviewed work. Four corrections follow.
What did not survive checking
Recorded rather than quietly fixed, because the last of these concerns the finding this site would most like to be true.
| What was wrong | The correction |
|---|---|
| A misattributed author | The interface-and-exploration paper was credited to 'Sanchez'. Crossref and OpenAlex both give the authors as Torricelli, Martino and Baronchelli. Corrected here. |
| A year that has moved | The voice-assistant study was cited as IJHCS 2024. Crossref records the issue year as 2025. Not a material change to the finding, but the citation was wrong. |
| A citation count that depends on who is counting | The brief gives 875 citations for the most-cited paper. OpenAlex agrees; Crossref says 839 for the same DOI. Neither is wrong — they index different corpora. The figure on this page is OpenAlex's, read on 2026-09-01, and it is labelled as such rather than presented as the citation count. |
| Two trust findings rest on studies of a different kind of tool | The claim that honest limits build trust is supported here by a longitudinal study of gig drivers using decision aids and a trip-planning experiment. Both are about AI decision-making, not creative or generative tools. That does not make the claim wrong, and it is the finding this site most wants to be true, which is exactly why the population is stated rather than glossed. The second of the two does not test disclosure of limits at all; it tests task complexity and uncertainty. |
How to read this page
- These are other people's studies. Each entry states what the researchers actually did and with whom, because a finding from 129 crowd workers writing short texts is not the same kind of fact as one from 145,000 platform logs.
- Where a study's population differs from this site's audience, that is recorded in a scope note rather than left for the reader to discover in the paper.
- The implications are ours and are argued, not measured. Nothing in this literature was conducted on this site.
The findings
Nobody knows how to prompt, including people who use these tools daily.
What the study did
A design probe put non-experts in front of an LLM prompting tool and watched. They explored prompt designs opportunistically rather than systematically, and two things got in the way: expectations carried over from instructing another human, and a tendency to overgeneralise from a single result.
Source
Why Johnny Can't Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts
LLM text prompting, not image or video. The mechanism transfers; the specifics may not.
What it means for a site like this
It is the reason a prompt reference is worth writing at all, and it argues against the format most of them use. If people overgeneralise from one result, a page of examples without explanation teaches the wrong lesson — which is why every vocabulary entry here says what decision it controls rather than only what to type.
Given something to choose from, people choose. Given a box, they hesitate.
What the study did
129 crowd workers wrote short texts with GPT-3 under different interfaces: one suggestion versus three, with and without the ability to write instructions. When the interface offered multiple suggestions *and* an instruction box, people preferred picking from the suggestions. Where they did write instructions, it was to ask for inspiration, topics or facts rather than to steer.
Short-form creative writing with a text model. The finding is about interface shape rather than medium, but it has not been replicated for image or video.
What it means for a site like this
A reference should hand you something to pick, not a blank page. It is the argument for the fragment lists, the mood recipes that already agree with themselves, and the worked prompt sets — and against a page that only explains principles.
Prompt skill spreads by copying phrases, not by composing them.
What the study did
A three-month ethnographic study of online text-to-image communities found that practitioners control output by appending short key phrases — 'modifiers' — that they collect from each other. The paper builds a taxonomy of six types of them.
Source
A taxonomy of prompt modifiers for text-to-image generation
Text-to-image community practice, observed rather than experimental.
What it means for a site like this
This is the empirical case for the shape of the prompt-craft section: lists of terms grouped by the decision they control, phrases you can lift, and printable sheets. People are already working by collecting vocabulary, so a reference that hands them vocabulary is meeting the practice rather than correcting it.
After a bad result people blame themselves, and then they stop.
What the study did
Interviews with text-to-image users traced a write, judge, refine loop. The loop breaks at the point where someone cannot tell whether the failure was their prompt or the model's limit — and when they cannot tell, they assume it was them.
Source
Is It AI or Is It Me? Understanding Users' Prompt Journey with Text-to-Image Generative AI Tools
What it means for a site like this
The strongest argument on this page for the concepts section existing. Pages that say plainly what models cannot do — physics, permanence, identity across generations — exist to answer the question this study identifies as the breaking point. It also explains why the most useful single line in the VFX prompt set is the one recording that the model silently dropped a beat: a failure you can attribute is a failure you can work around.
How widely people explore is decided by the interface, not only the model.
What the study did
An analysis of more than 145,000 real prompts from the logs of two platforms, Stable Diffusion and Pick-a-Pic, found users tend toward exploring new concepts rather than reusing old ones — but that the two platforms, which differ in scope and features, produced starkly different exploration behaviour.
Source
The Role of Interface Design on Prompt-mediated Creativity in Generative AI
What it means for a site like this
A caution about our own facet pages and filters. What we put in front of a reader narrows or widens what they try, so a category page that lists only the obvious models is not neutral — it is a decision about what gets used.
A short explanation raises trust. A long one lowers it.
What the study did
Two between-subjects experiments varied explanation detail (low versus high) and control (automatic versus on demand) for AI-based recommendations, and measured trust and acceptance. More detail improved perceived understanding, but the relationship with acceptance is not simply more-is-better; the literature this draws on treats information overload as the limiting factor.
Source
Why should I trust you? Influence of explanation design on consumer behavior in AI-based services
AI-based recommendations, not generative media. Read it as a finding about explanation format rather than about creative tools.
What it means for a site like this
It is an argument for the shape this site already uses on model pages: one plain sentence about where a figure came from, with the full provenance one click away in a disclosure rather than in the paragraph. It is also a warning about our own long caveat blocks.
Saying what a system cannot do earns more trust than leaving it unsaid, and the effect grows over weeks.
What the study did
A longitudinal in-situ study of gig drivers using AI decision aids examined designs that expose uncertainty and how they affect trust over time, rather than in a single session. A separate CHI study of 258 participants on a trip-planning task found task complexity and uncertainty significantly shape trust and reliance.
Both studies are decision-aid research, with gig drivers and a trip-planning task. Neither studied creative or generative tools. The direction of the effect is what transfers; the magnitude is not established for this domain, and anyone citing these for creative AI — including us — should say so.
What it means for a site like this
This is the closest thing to external evidence for how this whole site is built: "not established" instead of an estimate, per-zone confidence levels, recorded disagreements between sources, six editorial flags withheld for want of a primary source, and a commercial-use block that refuses to tell you that you are cleared. That discipline was adopted on principle. It appears to also be the thing that earns trust — and specifically over weeks rather than on a first visit, which is the timescale a reference site lives on.
People arrive for different reasons, and behave differently once they do.
What the study did
Survey and forum research into why people use generative conversational AI identifies distinct motivations — getting work done, making things, learning, and play — each with different usage patterns.
Source
Why do people use ChatGPT? Exploring user motivations for generative conversational AI
Also cited
Exploring Applications and User Experience with Generative AI Tools
What it means for a site like this
The site is already split three ways — craft, directory, reference — which maps loosely onto learning, working and making. What it does not do is help someone identify which they are on arrival, and the home page currently asks them to work that out from three equally weighted panels.
Four studies the brief did not include
Found while verifying the others, and closer to this site's subject than most of the originals.
Design Guidelines for Prompt Engineering Text-to-Image Generative Models
The most directly useful paper for this site's subject, and absent from the brief. 5,493 generations across five experiments, covering 51 subjects and 51 styles, producing actual design guidelines for text-to-image prompts structured as subject plus style keywords.
Source
Design Guidelines for Prompt Engineering Text-to-Image Generative Models
Its subject-plus-style structure is the same shape as the aspect ordering used in the prompt guides here, arrived at independently. Worth reading before anyone rewrites those pages.
Generative AI for Film Creation: A Survey of Recent Advances
A survey of generative AI in actual filmmaking workflows: character consistency, stylistic coherence, motion continuity, and the integration of real footage with generated elements.
Source
Generative AI for Film Creation: A Survey of Recent Advances
It names the three problems this site's concept pages are built around, and it describes the real-footage-plus-generated-element technique that the hand-drawn VFX set on this site is an instance of.
Sound Designer-Generative AI Interactions: Towards Designing Creative Support Tools for Professional Sound Designers
Nine professional sound designers used two generative tools in their own practice, with semi-structured interviews. Most research in this area studies novices; this one studies practitioners.
The only study found here that looks at generative audio in professional practice, which makes it the best external check on the audio prompt and audio tag pages.
User Experience Design Professionals' Perceptions of Generative Artificial Intelligence
How UX design professionals perceive generative AI in their own work.
Source
User Experience Design Professionals' Perceptions of Generative Artificial Intelligence
Relevant to the craft pages: it is evidence about how practitioners in a design discipline actually place these tools, rather than speculation about it.
Common questions
Why do people struggle to write AI prompts?
A CHI 2023 study of non-experts found they explore prompt designs opportunistically rather than systematically, carry over expectations from instructing another person, and overgeneralise from a single result. They also cannot reliably tell whether a prompt worked.
Do people prefer writing prompts or picking from options?
Picking. In a CHI 2023 study, 129 people writing with GPT-3 preferred choosing from multiple suggestions over writing instructions to steer the model, and where they did write instructions it was to ask for inspiration rather than to control output.
How do people actually learn to prompt?
By collecting phrases from each other. A three-month ethnography of text-to-image communities found practitioners control output with short modifier phrases they take from other people, not phrases they compose themselves.
Does admitting what an AI tool cannot do reduce trust?
The research points the other way: designs that expose uncertainty appear to improve trust, and the effect compounds over weeks rather than appearing in a single session. The strongest study on this looked at gig drivers using decision aids rather than creative tools, so the direction transfers more safely than the magnitude.
Assembled from a contributed research brief covering fourteen studies, plus four further works found during verification. The brief's own framing, product recommendations and internal references are not reproduced; the implications on this page are written for this site.