The IKEA effect in human-AI collaboration: Does the effect exist for non-physical products? Part II.

According to the IKEA effect, people are willing to pay more for a product they have created through their own perceived effort than they would for an off-the-shelf product. In our research, we investigated whether the IKEA effect would exist (1) if ChatGPT were also involved in creating the product and (2) if the final product is textual content. We conducted a randomized controlled trial that included all the background factors known to jointly trigger the IKEA effect.
Our results show that the IKEA effect can be detected in human-AI collaboration when the product is non-physical content. We have demonstrated that participants (1) produced a superior product based on their subjective preferences and (2) would purchase it at a higher price. However, in our research, (3) the IKEA effect applied not only to the end product, but also to the instrument: members of the IKEA group were more satisfied with ChatGPT and would pay more for the application in terms of the product they created.
Thus, by including all known background factors that jointly trigger the IKEA effect, we have successfully refuted previous studies that were unable to prove the IKEA effect in the field of text generation.
Artificial intelligence is a technological opportunity that allows shoppers to personalize products or select and access them more quickly. The IKEA effect could be exploited in this area. Our paper contributes to the practical identification of the boundary conditions necessary to trigger the effect.

2 The research method

2.1 Task used in the experiment

In our research, we use ChatGPT, a freely available tool developed by OpenAI, which is essentially an advanced chatbot that can handle almost any text-based request. Since its three years of existence, the tool has developed a large number of uses, e.g., assisting text writers in repetitive tasks, avoiding bias, improving clarity, assisting proofreaders, metadata annotation of texts, language translation, literature search (Lund et al. 2023). We will not address here the ethical issues related to the use of the tool, as they are not related to our subject, but only mention their existence, such as the existence of biases encoded in the texts used to teach the model or the issue of plagiarism (on the dilemmas raised by general ethical issues related to Ai, see for example Somosi and Hajdú, 2023).
The participants in our research were asked to create a text for a brochure of up to one page about a foreign destination of their choice using ChatGPT. Travel is a popular topic, and there is hardly anyone who has not searched the internet for travel information, so participants are expected to have a basis for comparison when judging the information gathered using ChatGPT. We expect that by relying on AI, more focused and structured information can be collected. Therefore, we believe that the task is suitable for measuring the IKEA effect because (1) travel is a popular and well-known topic and easy to work with, (2) the task does not require any special knowledge because the text can be written based on the research participant’s own preferences, ideas and knowledge, (3) the subjects can also evaluate the appropriateness of the product.
It was important to us that the task involved collaboration between the individual and ChatGPT, and we also know that to trigger the IKEA effect, the individual needs to perceive their own effort. Therefore, we structured the task to emphasize their own work: they were asked to (1) first come up with a destination, (2) design the structure, (3) style and tone of the brochure, (4) write the prompts based on these (see Online Appendix for the exact wording of the tasks). They were also asked to share their solutions to these four subtasks in a questionnaire, and only then could they turn to ChatGPT. Here, they were free to write the text of the brochure: they could work in several steps, clarify their request, and improve ChatGPT’s proposal until they reached the solution considered as final. The task was thus developed by the research participants in a guided way, in sub-tasks, and we controlled the right ratio of human and machine work.
Before the actual experiment, we did the task ourselves for testing purposes, and we also tested the task and the questionnaires in a pilot experiment with some members of the target group. Based on this experience, we revised the task several times, because we emphasized that it should not be long in time, should keep a balance between the work of the individual and the ChatGPT, should not be complicated, that it should require the help of other tools (Google, YouTube, acquaintance, etc.) and should be completable. These factors were designed to allow the IKEA effect to emerge.

The IKEA group’s final research task was as follows (translated from Hungarian):
Dear Participant!
Assume you work as a copywriter for a marketing agency. Your task is to write the text for a brochure promoting a trip abroad, up to one page, using ChatGPT. The aim of the task is to produce material that is concise but highly informative for all travel planners. The brochure will be successful if it makes people want to travel here! Your task is only to write the text, the design of the material is not part of the current task. Please follow the steps below, step by step.
1. Task: please write down which foreign location you would like to promote in the brochure (location is optional):
2. Task: please list the aspects you would like to use to build your brochure. Have a structural plan, give structure and organization of your material.
3. Task: write down the tone, style, mood, etc. of the text of the brochure.
4. Task: Based on your structure and instructions, formulate your request to ChatGPT. Please write down your prompt. (If you use ChatGPT in English, you can also share it in English)
5. Task: Get help from ChatGPT! Use the prompt you wrote in the previous exercise and use ChatGPT to compose the text of the brochure. Work in several steps, clarify your request, improve ChatGPT’s suggestion with additional prompts and/or editing where you feel it is necessary! You are free to work with ChatGPT! Please, upload the whole ChatGPT conversation! (If you use ChatGPT in English, you can also share it in English)
6. Task: please upload your final brochure!

In the control group, we gave a task that prevented the IKEA effect: the task was also to create a brochure promoting a foreign location, but it was a brochure written by us and created in collaboration with ChatGPT (see Online Appendix), so the participant was presented with a ready-made output. They had one task: to review the material they had received. They were given a short questionnaire to give feedback on whether they were satisfied with the material, whether the brochure had achieved its purpose and whether they would make any changes to the material. The control group’s task was as follows (translated from Hungarian):
Dear Participant!
Below is a brochure promoting a trip abroad, created with the help of ChatGPT. The aim of the task was to create a content using ChatGPT that is concise but highly informative for all travel planners. The design was not part of the assignment, only the writing of the text. The aim of the brochure is to make you want to visit the destination. Suppose you work as a copywriter for a marketing agency. Please read, revise, and evaluate the text of the brochure according to the criteria below on a scale of 1 to 7, where 1=not at all and 7=absolutely yes.

2.2 Questionnaires, statistical tools

The questionnaires completed by the IKEA and control groups were identical. Before completing the experimental task, each participant was asked to fill in a questionnaire on two topics (note [1]). The first was a short demographic questionnaire to collect information on the participant’s gender, age, place of residence, and job, and the second measured how participants perceived ChatGPT before they encountered the task. They then completed the task (which, we saw, was different for the IKEA and control groups) and were asked about their opinions on the quality of the given brochure. Afterwards, participants in both groups were given a brochure we had prepared about the Dominican Republic, and we assessed this by measuring their willingness to pay for their own brochure as well as for the brochure we had prepared. Finally, we measured their opinion of ChatGPT again, this time not in general, but in relation to their task.

The following questions were used:
1. IKEA effect: „How much work did you put into the brochure yourself?“
2. Opinion on the quality of the produced brochure: „How good do you think the produced brochure is?“
3. The role of ChatGPT in the quality of the brochure produced: „Do you think ChatGPT produced a better-quality brochure than if it had been produced using a tool other than artificial intelligence (whether it’s a travel guide, a friend’s help, Google search, YouTube, etc.)?“
4. Willingness to pay: „Please indicate between 0-500 HUF how much you would pay for your brochure produced with ChatGPT“ (question for IKEA group only) and „Please indicate between 0-500 HUF how much you would pay for a brochure produced by someone else with ChatGPT promoting the Dominican Republic“ (question for both groups)
5. Opinion on ChatGPT regarding the assignment: „Are you satisfied with ChatGPT in terms of the evaluation of the brochure produced?“

For information not directly related to the purpose of the research, but relevant to us, we also asked how much our respondent would pay for the ChatGPT service.
For validity, we asked several alternative questions for each point, each with a short text response, to check participants’ answers and to filter out any inconsistencies and contradictions.
Most of the variables used in the study were defined on Likert scales (1-7, following Norton et al.). The reliability of questions on the same item was assessed using the principal component analysis components’ AVE index (average variance extracted) and the CR (construct reliability) index. A one-sample t-test was used to test the deviation of the mean of the scores given by the respondents from its middle value (4), a two-sample t-test to compare the mean scores of the IKEA and control groups (using a one-sided counterhypothesis, abbreviated as HA in the following), and a paired t-test to test the within-group deviation of the mean scores given for the IKEA and control groups. In the case of two-sample t-tests, we used the appropriate test variant if the difference in variances made it necessary. For all statistical tests, the standard 5% significance level was used as the threshold. We always considered the practical significance of the differences in addition to statistical significance. Due to the sample size of 20, it is recommended to treat the results of the tests with caution, therefore in all cases descriptive statistics were also used to interpret the results.

2.3 The research subjects

From the research of Norton et al. (2012), we know that if an individual cannot complete the task, the IKEA effect dissipates. We therefore included individuals who could complete the research task and were not constrained by their lack of knowledge of ChatGPT. The research task was designed in such a way that it does not require a high level of knowledge of ChatGPT, as very few individuals have such knowledge yet. As ChatGPT has been tested and used continuously since its launch in the first author’s field of expertise, online marketing, we included 20 individuals working in this field who were known to use the tool regularly. All of them agreed to participate in the research.
They were randomly divided into IKEA and control groups of 10-10 people. The research was conducted online, using an online questionnaire, so that participants could complete the task at a time of their choice, even in several steps. Participants’ feedback indicated that they spent on average 35 minutes completing the questionnaire and the task and all of them solved the task in one step. All our questions received a valuable response.
The IKEA group consisted of 5 men and 5 women, with an average age of 37.5 years, 90% of them urban and all of them working in a city. 60% of them are business owners, senior or middle managers and 40% are employees. The control group consisted of 6 men and 4 women, with an average age of 38.8 years, 90% urban and also 90% working in a city, 10% abroad. 50% are business owners, senior or middle managers and 50% are employees. In other words, the two groups did not differ significantly on the basis of these background factors.

3 Results

During the principal component analysis, four principal components were created, namely:
• PC 1: Variables measuring the collaboration between human and artificial intelligence.
• PC 2: Opinion formed about ChatGPT.
• PC 3: Opinion formed AI’s capability to replace human labor after the research task.
• PC 4: Opinion formed AI’s capability to replace human labor before the research task.

Table 1 shows the AVE and CR indices, which measure the reliability of questions on the same component. In the case of AVE, the agreed threshold is 0.5, and in the case of CR, it is 0.7, above which values indicate adequate reliability. In the case of the second component, the AVE is slightly lower than the 0.5 threshold, while the other values are adequate. We can therefore say that the variables within the principal components are consistent with each other. The CR results also supported the reliability of our questionnaire. That is, each construct meets the criteria for reliability and convergent validity of our questionnaire.

WordPress Tables
Table 1: AVE and CR values for the principal components
Source: Authors

Now let’s move on to the individual variables that measure the IKEA effect (Table 2). The IKEA effect requires the perception of one’s own effort. The mean of the IKEA group members’ responses to the relevant question was 5 (SD=1.155), compared to 2.5 for the control group (SD=0.707), a significant difference in terms of magnitude of the difference and statistically significant difference based on the two-sample t-test.
The responses to the question on the quality of the brochure produced also differed between the two groups, with a significant difference both practically and statistically: the IKEA group had a mean of 5.8 (SD=0.632), all scoring above 4, while the control group had a mean of 4.2 (SD=0.422). The latter score was not significantly lower than 4, indicating a medium rating, i.e. the control group was uncertain.
Examining the role of ChatGPT in the quality of the brochure produced, the IKEA group gave a mean score of 6.8 (SH=0.422), all scoring 6 or 7, which is significantly different from 4, the midpoint of the scale. The control group had a mean of 5.4 (SD=1.713) and was not significantly less than 4, but the means of the two groups were both practically and statistically different, so it is statistically proven that the IKEA group perceived more strongly the role of ChatGPT in the quality of the brochure produced.
Turning to willingness to pay: the IKEA group would have paid on average 295 HUF for a brochure they produced themselves (SD=137.052), while they would have paid on average 132 HUF for a brochure they received (SD=88.544), i.e. they would have been willing to pay on average more than twice as much for a brochure they produced themselves, which is both practically and statistically significant difference.
The IKEA group’s willingness to pay for their own brochure was also compared with the control group, the latter group would have paid on average 151 HUF for a ready-made brochure that they could only review (SD=153.511), which is a statistically significant difference. In other words, the presence of the IKEA effect, i.e. the fact that the research participant was able to create the brochure themself using ChatGPT, also meant a significant increase in willingness to pay, almost twofold, in this comparison.
In the evaluation of the completed brochure, the IKEA group gave an average response of 6.2 (SD=0.632) and the control group an average response of 4.1 (SD=1.197), which is also a practically and statistically significant difference. Furthermore, the IKEA group would pay on average 2,740 HUF for the ChatGPT service (SD=1834.969), while the control group would pay on average 1,395 HUF (note [2]) (SD=1,779.474), so the IKEA group indicated almost twice the average price. The difference is definitely significant in practical terms, and it can also be considered statistically significant, even though the corresponding p-value, 0.06, is close to the threshold. So the IKEA effect was not only present for the product of the task, but also for its instrument (!).

WordPress Tables
Notes: 1) HA: Mean>4 (the middle of the scale); 2) HA: Mean<4 (the middle of the scale); 3) HA: the average in the IKEA group is higher; 4) We did not perform a one-sample test because there is no reference value to compare the mean to; 5) Two-sample t-test, HA: the average in the IKEA group is higher; 6) Paired t-test for the IKEA group to compare the prices of the self-made brochure and the brochure made by others, HA: the average of the former is higher; 7) HA: the average of the two groups differs

Table 2: Distribution of questions measuring dimensions of the IKEA effect in the two groups (with the exception of the amount measured in HUF, all variables were measured on a scale of 1 to 7, where higher numbers always correspond to higher values)

Source: Authors

4 Summary

The fact that the IKEA group members felt that they had invested more effort shows that the background condition of our experiment was met. The results of our research show that the IKEA effect can be triggered not only by human effort: the effect can also be detected for human-AI collaboration. Furthermore, the IKEA effect exists not only for physical products, but also for the creation of non-physical products (content). We have shown that with the involvement of AI, subjects (1) produced a better product based on their own subjective judgements and (2) would buy it more expensively, both in comparison with a product they did not produce themselves and with the control group.
However, the IKEA effect in our research was not only for the product of the task, but also for its means: members of the IKEA group were more satisfied with ChatGPT in terms of the product created and would pay more for the service it provided.
Our results thus refute the findings of Mehler et al. (2024): the IKEA effect can be detected not only in image generation, but also in text generation, i.e. it is a truly general phenomenon that exists under multiple boundary conditions. We designed our research based on the literature describing the mechanism of action of the effect (Norton et al. 2012), as we found that Mehler et al. research did not provide the conditions necessary to induce the effect. In contrast to them, we assigned a task in the experiment that did not require either technological or job-specific knowledge and whose success could be judged by the subjects, so that they were able to perceive the added value of their own work. Furthermore, we structured the task in such a way that we did not leave the steps of the task to the individual, but broke it down into parts that we controlled, thus controlling the ratio of human to machine work – this is important because the predominance of machine work inhibits the perception of individual effort and thus the creation of the IKEA effect.

5 Limitations

The subjects of our research were not members of the general population, but marketing professionals who have experience in using ChatGPT because of their job, and the task itself was a task close to that of a marketing professional. However, we do not believe that this significantly influenced our results.
Our sample size was twenty, and although there is conflicting literature on the sample size criterion for the t-test, we treated the results of the test with caution, relying on descriptive statistics and practical judgement of effect size.
We measured the willingness to pay for the completed brochure while the completed brochure was saved on our subjects’ computers. However, this was the case for both the control and the IKEA group, so we do not hypothesize an effect on the outcome.

6 Applicability of our findings

Artificial intelligence already has many practical applications in business and marketing, such as its use in pricing (see for example Danyi 2019). In the field of online marketing, it is of primary importance to identify the factors that can improve sales results, whether it is increasing the propensity to buy or the average basket value. The business value of the IKEA effect lies in the fact that, if it can be triggered, it can also be used to drive sales by increasing individuals’ willingness to pay.
Artificial intelligence is a technological opportunity that allows shoppers to personalize the products they want to buy, or to select and access them more quickly. The IKEA effect could be well exploited in this area. Previous research on self-made physical products has shown that shoppers are 63% more willing to pay in the presence of the IKEA effect (Norton et al. 2012). In our research, we measured almost 100% higher price for non-physic products and human-AI interaction.
The exploration and analysis of the relationship between the effect and artificial intelligence can be of significant value from a commercial perspective, as multiple implementations of human-AI collaboration are expected to emerge in the future. To put this into practice, we consider the factors identified in our paper, the absence of which in previous research (Mehler et al. 2024) prevented the effect from materializing, as key to its practical realization.

End of Part II.

Poznámky/Notes

[1] All research questionnaires can be found in the online appendix. Available at: <https://www.researchgate.net/publication/392862672_Research_documentation_for_the_paper_entitled_The_IKEA_effect_in_human-AI_collaboration_-Does_the_effect_exist_for_non-_physical_products_published_in_Marketing_Science_Inspirations_June_2025>
[2] The control group also included an unrealistically high price of HUF 50,000 per month, this outlier was excluded from the analysis (as a presumed typo).

Literatúra/List of References

1. Brunner, F., Gamm, F. and Mill, W., 2023. MyPortfolio: The IKEA effect in financial investment decisions. In: Journal of Banking & Finance. 2023, 154, 106529. ISSN 0378-4266. [online]. [cit. 2025-05-11]. Available at: <https://www.sciencedirect.com/science/article/abs/pii/S0378426622001236?via%3Dihub>
2. Danyi, P., 2019. A mesterséges intelligencia árazásbeli alkalmazásának várható hatásai. In: Marketing & Menedzsment. 2019, 53(1), 17-29. ISSN 2786-3395. Available at: <https://doi.org/10.15170/MM.2019.53.01.02>
3. Jacquemin, P. H., Gräf, M., Bauch, K., Kaur, A. and Mehler, M., 2025. When feelings meet code: How generative AI affects the emotions of developers. In: AMCIS 2025 Proceedings. 4, 2025. [online]. [cit. 2025-05-11]. Available at: <https://aisel.aisnet.org/amcis2025/sig_cnow/sig_cnow/4>
4. Festinger, L., 1957. A theory of cognitive dissonance. Stanford University Press, 1957. ISBN 9780804709118.
5. Mehler, M., Ellenrieder, S. and Buxmann, P., 2024. The influence of effort on the perceived value of generative AI: A study of the IKEA effect. In: ECIS 2024 Proceedings. 6, 2024. [online]. [cit. 2025-05-11]. Available at: <https://aisel.aisnet.org/ecis2024/track09_coghbis/track09_coghbis/6>
6. Norton, M. I., Mochon, D. and Ariely, D., 2012. The IKEA effect: When labor leads to love. In: Journal of Consumer Psychology. 2012, 22(3), 453-460. ISSN 1532-7663. Available at: <https://doi.org/10.1016/j.jcps.2011.08.002>
7. Lund, B. D., Wang, T., Mannuru, N. R., Nie, B., Shimray, S. and Wang, Z., 2023. ChatGPT and a new academic reality: Artificial intelligence‐written research papers and the ethics of the large language models in scholarly publishing. In: Journal of the Association for Information Science and Technology. 2023, 74(5), 570-581. ISSN 2044-8287. Available at: <https://doi.org/10.1002/asi.24750>
8. Radtke, T., Liszewska, N., Horodyska, K., Boberska, M., Schenkel, K. and Luszczynska, A., 2019. Cooking together: The IKEA effect on family vegetable intake. In: British Journal of Health Psychology. 2019, 24 (4), i-iv, 739-998. ISSN 2044-8287. [online]. [cit. 2025-05-11]. Available at: <https://bpspsychub.onlinelibrary.wiley.com/toc/20448287/2019/24/4>
9. Sarstedt, M., Neubert, D. and Barth, K., 2016. Replication note the IKEA effect. A conceptual replication. In: Journal of Marketing Behavior. 2016, 2, 307-312. ISSN 2326-5698. [online]. [cit. 2025-05-11]. Available at: <https://www.researchgate.net/profile/Marko-Sarstedt/publication/310491448_The_IKEA_Effect_A_Conceptual_Replication/links/58fe16c94585159c2b2bcaf8/The-IKEA-Effect-A-Conceptual-Replication.pdf?origin=journalDetail&_tp=eyJwYWdlIjoiam91cm5hbERldGFpbCJ9>
10. Schecter, A. and Richardson, B., 2025. How the role of generative AI shapes perceptions of value in human-AI collaborative work. In: CHI ’25: Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, 2025. ISBN 9798400713941. [online]. [cit. 2025-05-11]. Available at: <https://dl.acm.org/doi/10.1145/3706598.3713946>
11. Somosi, Z. and Hajdú, N., 2023. Mesterséges intelligencia etikai dilemmái: ellenszenv felmérés és következmények. In: Marketing & Menedzsment. 2023, 57(3), 65-74. ISSN 2786-3395. Available at: <https://doi.org/10.15170/MM.2023.57.KSZ.03.07>
12. Weinert, S., Gunther, E., Ruger-Muck, E. and Raab, G., 2020. Artificial intelligence in personnel selection and its influence on employer attractiveness. In: Marketing Science & Inspirations. 2020, 15(3), 22-35. ISSN 1338-7944. Available at: <https://doi.org/10.46286/msi.2020.15.3.2>

Kľúčové slová/Key words

IKEA effect, artificial intelligence, ChatGPT, human and artificial intelligence collaboration
IKEA efekt, umelá inteligencia, ChatGPT, spolupráca ľudskej a umelej inteligencie

JEL klasifikácia/JEL Classification

M31

Résumé

IKEA efekt v spolupráci človeka s umelou inteligenciou: Existuje tento efekt aj v prípade nefyzických výrobkov? Časť II.
Podľa IKEA efektu sú ľudia ochotní zaplatiť viac za výrobok, ktorý vytvorili vlastným úsilím, ako za hotový výrobok. V našom výskume sme zisťovali, či bude IKEA efekt existovať (1), ak sa na tvorbe produktu podieľa aj ChatGPT a (2), ak je konečným produktom textový obsah. Vykonali sme randomizovanú kontrolovanú štúdiu, ktorá zahŕňala všetky základné faktory, o ktorých je známe, že spoločne vyvolávajú IKEA efekt.
Naše výsledky ukazujú, že IKEA efekt možno zistiť pri spolupráci človeka a umelej inteligencie, ak je produktom nefyzický obsah. Preukázali sme, že účastníci (1) na základe svojich subjektívnych preferencií vyrobili lepší produkt a (2) kúpili by ho za vyššiu cenu. V našom výskume sa však (3) IKEA efekt uplatnil nielen na konečný produkt, ale aj na nástroj: členovia skupiny IKEA boli s aplikáciou ChatGPT spokojnejší a zaplatili by za ňu viac z hľadiska vytvoreného produktu.
Zahrnutím všetkých známych základných faktorov, ktoré spoločne vyvolávajú IKEA efekt, sme teda úspešne vyvrátili predchádzajúce štúdie, ktoré neboli schopné dokázať IKEA efekt v oblasti tvorby textu.
Umelá inteligencia je technologická príležitosť, ktorá umožňuje kupujúcim personalizovať výrobky alebo rýchlejšie si ich vybrať a dostať sa k nim. V tejto oblasti by sa mohol využiť IKEA efekt. Náš článok prispieva k praktickej identifikácii hraničných podmienok potrebných na spustenie efektu.

Recenzované/Reviewed

1. April 2025 / 24. May 2025