Two machines make the same kind of drawing. A viewer first sees only the finished image, then sees how a machine made it, then sees the machine itself. The image has not changed, yet the viewer may change their judgment about whether the act was creative. This is the question at the center of Pennanen and colleagues’ 2025 study, From Product to Producer. It examines the human assessment of AI creativity, not a machine’s inner experience or a final philosophical verdict about creativity.
That distinction matters. A polished picture can prompt admiration without telling us how it was produced. A visible procedure may suggest deliberate selection or skill. A robot’s body may evoke expectations about agency and intention. Pennanen and colleagues designed experiments to separate some of these cues and ask how they affect ratings. Their findings make a useful research brief for the Learning Hub because they show that what observers can see is part of a creativity judgment.
The question behind the paper
The researchers asked whether people rate a robot’s creativity differently as they receive more perceptual evidence about an act. Their levels were the product alone, the product with the process, and the product with the process and producer. They also compared two robot forms, expecting that embodiment morphology might alter the evaluation. The study therefore tested two related ideas: whether more context changes judgment and whether the machine’s visible form matters in its own right.
This was not a competition in which an AI system tried to outperform human artists. Nor was it an experiment that changed the machine’s creative process while holding audience information fixed. The same act could be judged with different evidence available. That design helps isolate the observer side of a question often discussed as if the artifact alone contained the answer. A change in rating under these conditions is evidence about attribution, not about a newly created machine capacity.
The paper belongs to a longer tradition of studying creativity as something people assess. Amabile’s consensual assessment technique asks judges to evaluate creative products without reducing creativity to one mechanically counted feature. Such evaluations can be reliable in a defined domain, but they depend on task, judges and information. In a robot experiment, the object of judgment expands from a finished drawing to an act with a visible process and producer. The shift is precisely what this paper makes observable.
What the researchers actually did
The published article reports a smaller pre-study with 30 participants and a main within-participants laboratory study recruiting 60 participants. The work involved drawings by two physical robots with distinct morphologies, a plotter and a robot arm. Participants rated what they saw under increasing levels of perceptual evidence. The design is often summarized as three evidence levels by two embodiments. It is a controlled visual-art setting, not a representative survey of all creative occupations or all AI systems.
In the main study, the evidence levels accumulated rather than asking separate groups to see unrelated artifacts. A participant could see the product, then additional information about the process, then the producer. That sequence is useful because it reveals how judgment changes as a particular act becomes more visible. It also means order and carryover need consideration: once an observer has seen more, they cannot unsee it. The design supports a within-person comparison under the authors’ procedure, not a claim that the effect would be identical in every presentation order or real-world gallery.
The authors adapted creativity-rating questions in the spirit of the consensual assessment technique and included related judgments such as originality, liking, aesthetic appeal, technical artistic ability and detail. These are not interchangeable qualities. A viewer can admire technical execution while finding a work predictable, or find an idea original while disliking the result. Asking about them separately helps identify what a broad creativity judgment may be tracking.
The main research question was about observers. The robots were held sufficiently comparable in their creative operation for the comparison to focus on what people perceived about the product, process and embodiment. This is a strength for causal interpretation of the rating change. It also narrows the conclusion: the study does not test whether a more capable model actually makes more creative decisions or whether a machine can have intentions.
More visible process, higher perceived creativity
The paper reports that creativity ratings rose as viewers saw more than the finished product. Ratings were higher when the process was visible, and higher again when the producer was also visible. The authors’ discussion says the perceptual-evidence hypothesis was supported with strong statistical evidence in this controlled setting. What changed was the evidence available to participants about the same creative act. That is the central result.
A tempting but mistaken inference is that transparency itself made the robot more creative. It did not change the act being rated. The result instead suggests that observers use process and producer cues to interpret an artifact. They may infer effort, complexity, agency or skill from seeing how the drawing unfolds. The study did not establish one unique psychological mechanism for the rating shift. Its exploratory measures can guide later tests, but they should not be promoted into a settled causal explanation.
There is another practical interpretation. A product-only test can systematically differ from an evaluation in which process is visible, even when the artifact is unchanged. A benchmark or public demonstration that presents one system’s steps and another’s output alone may therefore compare unlike experiences. This does not mean more disclosure is always better; selective disclosure can also persuade. It means evaluation protocols should state what judges saw, in what order, and why that evidence was available.
What robot form did and did not change
The researchers also expected morphology to affect judgments. One robot was more like a plotter, while the other had a robotic-arm form. In the reported analyses, embodiment morphology did not produce a statistically significant main effect on creativity ratings, nor a significant effect within the evidence levels. This is a result against a particular hypothesis in a particular design. It is not proof that a robot’s appearance never matters for perceived creativity.
The exploratory analyses showed differences in how the robots were perceived on some other dimensions, including animacy and likeability. That is useful because it separates two claims that are easily conflated. Viewers may experience a robot’s presence differently without translating that difference into the creativity rating measured here. Different art forms, robot motions, social settings or audiences could produce other patterns, but those possibilities remain to be tested.
A null result also needs care. Statistical non-significance is not a guarantee of exactly zero effect. Sample size, measurement sensitivity and the range of morphologies can affect what a study detects. The appropriate conclusion is narrower: the predicted morphology effect was not supported in this experiment. The paper’s evidence for the perceptual-evidence effect is stronger than its basis for declaring morphology irrelevant everywhere.
Creativity ratings are not a test of machine consciousness
The language of “AI creativity” invites ontological debate: does the system truly create, or does it only produce outputs that people judge creative? Pennanen and colleagues focus on the second question. A person can rate a work creative while holding different beliefs about the system that produced it. A high rating shows an attribution under given information; it does not show that the producer has subjective experience, human-like motivation or independent intentions.
Runco’s 2024 conceptual paper takes a more restrictive view. It argues that generative AI may be innovative or discover combinations while lacking the processes that make human creativity authentic in his account. Gilhooly’s later analysis emphasizes how definitions of creativity affect whether computed responses can qualify. These papers do not cancel Pennanen’s experiment. They answer a different kind of question: which criteria should the concept include? The experiment answers how observers respond to selected evidence.
Likewise, Boden’s earlier discussion of creativity and AI examines whether computational systems can generate novelty in meaningful senses. Her framework makes it possible to distinguish a useful computer-produced idea from a claim that the computer is a human artist. The categories remain debated. The empirical rating result should not be used to close a conceptual argument that the study did not set out to settle.
This distinction protects human contributors as well. Behind many AI outputs are people who select prompts, curate data, set constraints, choose candidates, judge quality and decide where a work belongs. A robot drawing experiment isolates viewer judgments under controlled conditions, but real creative production often involves teams, tools and institutions. Attribution should track the actual process, not assume a single producer because an interface offers one visible object.
How the result changes evaluation practice
Suppose a museum invites visitors to rate drawings made by people and machines. If visitors see only final works in one condition and a studio video in another, differences in ratings may reflect information about process as much as differences in products. A fair comparison must decide whether the question is “How does the artifact strike a viewer without context?” or “How does the act appear when its making is visible?” Both are legitimate, but they are not the same measure.
In education, a teacher often assesses both final work and evidence of learning. A student can produce a strong result through a tool without understanding the underlying method. Conversely, a rough result can show meaningful experimentation. Pennanen’s result does not dictate school grading, yet it illustrates why process disclosure can change judgments. A school should say which aspect it values and avoid using a creativity label when it is really evaluating compliance, technique or presentation.
In product design, a company may reveal generation steps to increase perceived creativity. That could make the system easier to understand. It could also stage a process selectively to cultivate an impression of agency. The study gives a reason to examine this design choice; it does not establish that every process visualization is informative or ethical. Evaluators should compare visible cues with the actual causal process and ask what users infer from them.
A good replication would vary the artistic domain, participant expertise and cultural context. It would also test whether viewers distinguish informative process evidence from decorative animation. Does a truthful account of how alternatives were generated change ratings differently from a theatrical display that merely appears laborious? That question follows from the result but remains a proposal. The original experiment did not answer it.
Why the paper matters to Inner Technology
Inner Technology asks what capacities people bring to technological life. One is evaluative judgment: the ability to notice which evidence supports an impression and what the impression leaves out. The paper shows that an observer’s creativity judgment is sensitive to contextual information. A practical habit is to ask, “What did I see: the result, the making, or an account of the maker?” The answer can reveal why two apparently conflicting assessments differ.
Another capacity is epistemic humility. If seeing a process raises a rating, the rating may still be sincere. Yet the observer should be able to revise it when the process is better understood. A machine-generated drawing can be aesthetically valuable without a settled claim about machine personhood. A human contribution can be substantial even when the final image was partly produced by an automated system.
The Institute’s suggested evaluation practice is to identify the target of a judgment, record the information presented and keep separate scores for qualities such as originality, value, technical execution and process understanding. This is an editorial synthesis of the study and broader creativity-assessment literature, not a validated universal scoring rubric. It should be adapted to the domain and to the people whose work is being judged.
What the evidence does not show
The study does not show that robots possess consciousness, intentions or a human creative identity. It does not show that revealing a process improves the quality of the artwork. It shows that ratings changed as additional evidence was presented. The ratings cannot tell us whether a participant inferred effort, intelligence, novelty or social presence without further tests designed for those mechanisms.
The experiment does not generalize automatically from a controlled drawing task to writing, music, scientific discovery or collaborative design. Its pre-study and main study were limited in size and used selected machines, tasks and audiences. The null morphology finding should be read as a failure to support a specific prediction in this setting, not a universal law of embodiment.
Finally, the paper does not settle whether an AI output should be called creative under every definition. Runco and Gilhooly make clear that this depends partly on what one means by creativity. The experimental and conceptual questions can inform each other only when they remain distinct. A public claim should say whether it concerns an output, a human rating, a process or a producer.
Return to the drawing
The viewer begins with an image and ends with a richer account of its making. The changed judgment is not a measurement error to dismiss. It is evidence that the context of production belongs to how people assess creativity. But the increased rating remains a human response, not a certificate of what the machine is.
From Product to Producer invites a more exact conversation about AI creativity. When a work appears creative, ask what is being judged, which parts of its production are visible and which human choices remain outside the frame. That question can make evaluation fairer without requiring an easy verdict on the deepest conceptual debate.
Research references
- Pennanen, N., Linkola, S., Kantosalo, A., Hiillos, N., Männistö, T., & Guckelsberger, C. (2025). From product to producer: The impact of perceptual evidence and robot embodiment on the human assessment of AI creativity. ACM Transactions on Human-Robot Interaction, 14(3), Article 41. View source
- Amabile, T. M. (1982). Social psychology of creativity: A consensual assessment technique. Journal of Personality and Social Psychology, 43(5), 997–1013. View source
- Runco, M. A. (2024). The discovery and innovation of AI does not qualify as creativity. Journal of Cognitive Psychology. View source
- Gilhooly, K. J. (2025). Creativity: Definitions and computability. Journal of Cognitive Psychology. View source
- Boden, M. A. (1998). Creativity and artificial intelligence. Artificial Intelligence, 103(1–2), 347–356. View source