Donations in the digital age: Effects of human-machine interaction on donation behavior

How to cite

Essl, Andrea; Bieberstein, Frauke von; Hafner, Daniele; Dadic, Hana. 2025. "Donations in the digital age: Effects of human-machine interaction on donation behavior." AEA RCT Registry. AEARCTR-0017292. https://doi.org/10.1257/rct.17292-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Andrea Essl* , Frauke von Bieberstein , Daniele Hafner , Hana Dadic

* Primary investigator

Registry ID
AEARCTR-0017292
DOI
https://doi.org/10.1257/rct.17292-1.0
First registered
Last updated
Status
In development
Trial dates
2025-11-24 – 2027-02-28
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/17292

Abstract

The rise of digital donations and the potential of artificial intelligence in fundraising are reshaping charitable giving. While fundraising has traditionally relied on human interaction or non-interactive channels, digitalization and rising costs are prompting charities to explore automated alternatives such as chatbots. Research suggests that people respond differently to machines than to humans, often showing weaker emotional and social reactions, yet little is known about how such differences affect actual donation behavior. We address this gap with a controlled online experiment using a real-effort task followed by a dictator game in which participants decide how much of their earnings to donate to a charity. Participants are randomly assigned to one of three conditions, differing through which channel they report their donation amount: (i) FORM, where they report donations via a non-interactive input form, (ii) CHATBOT, where they interact with a chatbot (based on Open AI technology), or (iii) CHAT HUMAN, where they interact with a real person via live chat. All conditions follow the same structure for comparability, with participants fully informed about whether they are interacting with a chatbot or a human agent. We test whether interactive chats increase donations compared to forms, and whether human chats elicit higher giving than chatbot chats. We also examine mechanisms such as social image concerns, perceived closeness, and emotional responses. The findings will provide novel insights for behavioral research and practical guidance for nonprofit fundraising in the digital age.

Intervention

The online experiment uses a between-subject design with three conditions:

1. FORM: Participants provide their donation choice via a non-interactive online form. They select a charity from a list and enter the amount they wish to donate in a simple input field. Instead of a chat, they can access an FAQ page to address questions about the charities.
2. CHATBOT: Participants provide their donation choice in a chat with a chatbot. Within the chat, they can ask questions about the charities (FAQ), select their preferred charity, and indicate their donation amount. Participants know they are communicating with a machine.
3. CHAT HUMAN: Participants provide their donation choice in a live chat with a real person. As in the chatbot condition, the chat includes the FAQ function, the selection of a charity, and the input of the donation amount. Participants know they are communicating with a real person.

Experimental design

With our experiment we want to test two hypotheses. First, we expect that people will donate more when they report their donation amount to either a chatbot or a human agent, compared to a non-interactive form. This prediction is based on the idea that both the human-like features of a chatbot and the real human presence can activate social image concerns. Second, we expect that donations will be higher when participants interact with a human rather than a chatbot. Here, the presence of a real person may enhance feelings of social image concerns, social and emotional closeness, and the desire for social approval, all of which can lead to more generous behavior.

To test these hypotheses, we conduct a between-subject online experiment. In the first part, participants complete a real-effort task (counting zeros) to earn money and then in the second part they play a dictator game in which they decide how much of their earnings to donate to charity. We will offer well-known charities. Participants are randomly assigned to one of three treatments. In the FORM treatment, participants read FAQs, select their preferred charity, and indicate their donation amount via a simple, non-interactive input field. To maintain interactivity and comparability with the other treatments, information is revealed step-by-step, mimicking a conversational flow. In the CHATBOT treatment, they enter a chat with a chatbot. In the third treatment, called CHAT HUMAN, they enter a live chat with a real person. In all treatments, the procedure follows the same structure: a brief greeting, a limited-time opportunity to ask questions/in the form treatment they can access a FAQ page, the selection of a charity, the input of a donation amount, and a closing thank-you message including the opportunity to receive confirmation for their donation. In the CHATBOT condition, the chatbot is explicitly prompted to provide these elements, while in the CHAT HUMAN condition the human agent receives the same instructions as the chatbot prompts, ensuring comparability across treatments.

Participants are informed in advance about their reporting channel (form, chatbot, human). Those in the CHATBOT treatment know they are chatting with a chatbot, while those in the CHAT HUMAN condition know they are chatting with a real person. To further enhance the credibility of these framings, we inform participants in advance that a voluntary verification step will follow the experiment. In the CHAT HUMAN condition, participants are given the option to initiate a one-way live call, accompanied by a short chat option, with the assistant they have been chatting with. In the CHATBOT condition, we provide subtle cues for verification, such as requests for the bot to generate a specific poem.

In addition, we will elicit decision makers’ beliefs on injunctive and descriptive norms. Belief elicitation is incentivized and takes place after participants made their donation decision.

Participants will also complete a short survey on self-report scales (social image concerns, Inclusion of the Other in the Self (IOS) Scale, emotions, technical affinity, perception of the donation process, and socio-demographic variables (e.g., age, gender, income, education).

Design details

Randomization method
Computer (online experiment)
Randomization unit
Individual
Clusters
The experiment will be run on Prolific. The number of participants planned for the experiment is 519 people completing the study. We advertise for 580 participants to take into account individuals who might not be matched with a human agent and/or who must be excluded due to the restrictions below. We will exclude subjects who: - complete the survey faster than two standard deviations from the average completion time; - do not complete the study within 45 minutes of starting; - do not complete parts with a time limit within the given time; - do not complete the study for other reasons (e.g., no match in the defined time) - exit and then re-enter the task as a new subject (as these individuals might see multiple treatments); - do not answer one of the control question correctly the third time - fail the attention check question - incorrectly answer the control question on the channel - engage in behavior that suggests deliberate disruption of the task (e.g., offensive content) Based on previous research, we expect to exclude about 5 percent of participants because they do not meet the restrictions above.
Observations
The experiment will be run on Prolific. The number of participants planned for the experiment is 519 people completing the study. We advertise for 580 participants to take into account individuals who might not be matched with a human agent and/or who must be excluded due to the restrictions below. We will exclude subjects who: - complete the survey faster than two standard deviations from the average completion time; - do not complete the study within 45 minutes of starting; - do not complete parts with a time limit within the given time; - do not complete the study for other reasons (e.g., no match in the defined time) - exit and then re-enter the task as a new subject (as these individuals might see multiple treatments); - do not answer one of the control question correctly the third time - fail the attention check question - incorrectly answer the control question on the channel - engage in behavior that suggests deliberate disruption of the task (e.g., offensive content) Based on previous research, we expect to exclude about 5 percent of participants because they do not meet the restrictions above.
Arms
For each of the experimental treatments: About 184 individuals
Primary outcomes
Amount of money donated
Secondary outcomes
Beliefs on social norms, social image concerns, Emotions, Other in the Self (IOS) Scale, use of skip-button, time to skip (right-censored if no skip)

Cited by

No citation data available for this registration yet.

Keywords