WASHINGTON, D.C. — 童颜视频 has a long history of testing new data collection methods as technology, communication norms and respondent expectations evolve. From computer-assisted telephone interviewing (CATI) to web surveys and text messaging, our approach to innovation has been consistent: test carefully, evaluate rigorously, and be transparent about what works, what doesn’t, and what questions remain.
Today, we are in the early stages of testing another emerging technology: artificial intelligence-enabled phone interviewing. As with prior innovation work at 童颜视频, our goal is not simply to determine whether a new approach to interviewing can work, but to understand under what conditions it should be used and where caution is warranted.
AI-driven voice systems have advanced rapidly in recent years. Modern AI interviewing systems can speak naturally, adapt to respondents’ answers and handle conversational turns in ways that previous computerized interactive voice response (IVR) systems could not. These advanced AI systems can probe, ask follow-up questions and transcribe longer qualitative responses. At the same time, the scalability of AI phone interviewing creates the potential for faster and less expensive data collection and may ultimately give more people the opportunity to have their voices heard.
Our research to date has consisted of a series of pilot tests conducted across four continents, encompassing more than half a million call attempts in seven languages and using both random digit dial (RDD) samples and preconsented panel samples. These pilots have been designed to help us learn more about the key regulatory, technological and methodological questions that must be addressed before broader adoption can be considered.
Legal and Regulatory Compliance
Any use of AI in data collection must consider legal and regulatory compliance. For AI phone interviewing, this includes considerations such as:
- consent and disclosure requirements for automated or AI-assisted calls
- differences in regulations across countries and jurisdictions
- data privacy, storage and processing requirements
- regulations for how automated calls (regardless of the use of AI) can be placed (such as the Telephone Consumer Protection Act in the United States)
As a global research organization, 童颜视频 must ensure that the use of AI phone interviewing complies with local laws as well as ethical standards and Institutional Review Board (IRB) requirements. Regulations governing AI, automated calling and data protection continue to evolve, and they can vary widely across regions, countries and states. For example, in some countries, such as the United States, it may not be legal to use an autodialer with the AI infrastructure or to get consent via an AI interviewer. This would necessitate using a preconsented sample or human interviewers to begin the calls and then to transfer to the AI system.
Technological Considerations
After legal and regulatory requirements are addressed, the next question is whether the technology can meet the standards required for high-quality survey research. As with any new data collection technology, the issue is not simply whether it functions, but whether it functions reliably with real respondents.
AI phone interviewing differs from other automated systems in important ways. Traditional IVR systems follow a fixed logic. Regardless of whether a respondent says, "Yes," or presses "1," the system records the answer and moves to the next question. The same input always produces the same result. An AI interviewing system also follows a predetermined survey script, but in addition, it interprets open-ended speech rather than matching responses to fixed options. Because of that interpretive step, identical responses from two different respondents may not be handled in exactly the same way, similar to the way two human interviewers may not handle every scenario identically. That adaptability can create a more natural interaction than an IVR system does, but it also introduces new considerations for consistency and reproducibility.
This is not entirely different from the judgments human interviewers make. For example, if a respondent is asked, "Did you feel happy most of the day yesterday?" and says, "Yesterday was difficult because of work, but the evening was better," both a human interviewer and an AI interviewing system must weigh what the respondent most likely meant and decide on a next action. The AI might determine that the overall sentiment is positive, or it might ask a clarifying follow-up: "Would you say that, overall, you felt happy most of the day?" However, the types of errors and deviations from protocol will likely differ between AI agents and human interviewers. Understanding where those differences emerge requires looking at how the AI system processes a conversation.
Five interconnected components handle a distinct step in the interviewing process, shown in the image below.
Each of these components introduces its own potential for error, whether in capturing a respondent's words, deciding what to say next or coding responses for analysis. Careful testing is required to understand how these systems behave across diverse respondents, languages and calling environments. Our evaluation of the technology focuses on several core questions:
- Does speech recognition perform reliably across different call conditions, such as accents, speaking styles and variable audio quality?
- Is voice synthesis natural enough to sustain respondent engagement throughout the interview?
- Is the system's response time fast enough to maintain conversational flow with respondents?
- Under what circumstances does the system perform with high quality, and under what circumstances does it make mistakes?
- Can large language models (LLMs), a type of AI that interprets and generates text, accurately code respondent answers?
Our goal is to understand the conditions under which the technology can meet the standards required for high-quality data collection, and those under which each layer in the system may introduce nonresponse or measurement bias.
Methodological Considerations
Beyond legal and technological considerations, the central methodological questions are familiar ones: Who responds, and how does the mode itself affect measurement? With any new data collection technology, it is crucial to understand how it compares with existing methods and how it may introduce or reduce bias.
Who is willing to respond?
童颜视频 is exploring participation in AI phone surveys by looking at response rates, consent rates and completion rates. The central question is whether those who agree to participate in AI interviews, and those who end up completing an AI interview, differ systematically from those who decline to participate or who do not complete the entire interview.
Even after respondents consent to participate, additional barriers to completion that are true with any telephone interview may emerge, such as language barriers or dropped calls. AI may also introduce new challenges that affect the interview and respondents’ willingness to continue. For example, respondents may get frustrated with minor glitches in the technology that are uncommon in human interviews (for example, repeating the same question twice) and break off the interview.
Respondent perceptions of the AI survey experience also matter. Respondents must feel comfortable with how their data are collected and used, particularly with respect to privacy. If respondents have a negative experience with an AI phone interview, they may be reluctant to consent to future AI interviews. The experience must be positive enough that respondents are willing to participate again. Our early findings from the United States suggest that many respondents evaluate the survey experience positively. Yet this positive evaluation and willingness to participate in a future survey will likely vary by country, culture, prior exposure to AI and type of sample.
How does measurement differ?
AI phone interviewing does not fit cleanly into existing theoretical frameworks for data collection modes. In some respects, it closely resembles interviewer-administered CATI surveys. Unlike IVR, AI systems have become so sophisticated that it is possible for a respondent to complete an entire interview without knowing they did not talk to a human, unless explicitly informed.
Like surveys administered by human interviewers, AI telephone interviews are conversational. Respondents hear questions, respond verbally and can ask for questions to be repeated or ask for clarification. As with human-administered interviews or IVR, some respondents will adapt more quickly to the structured nature of the interview, while others will find it more difficult.
At the same time, AI phone interviewing shares important features with self-administered modes (such as web or IVR). Rapid, large-scale call volume is not constrained by the availability of human interviewers. Respondents also have greater flexibility and can complete the survey at their own pace or dial into the system at a time that is convenient for them. If the respondent needs to pause the interview, the AI agent can patiently wait as long as necessary and resume the interview when the respondent is ready again. The AI system can also switch languages or voice characteristics in real time.
These features may reduce certain known biases in human-administered surveys, such as social desirability or acquiescence bias. Alternatively, if the system sounds convincingly human, these biases may persist while introducing others. For example, AI interviewer systems may be less effective than humans at probing, managing the pace of the interview, nonresponse conversion or encouraging a respondent to complete longer interviews.
These features raise important questions about how measurement from AI phone interviews will compare with other methods and whether responses will more closely resemble self-administered or human-administered response distributions. These are all empirical questions that our testing seeks to address.
What Comes Next?
童颜视频’s approach to AI phone interviewing is deliberate. We are testing across contexts, with an emphasis on understanding risks as well as opportunities. Testing new methods, including the use of AI phone interviewing, is part of 童颜视频’s long-standing commitment to measuring public opinion rigorously in a changing world. As with past innovations in survey methodology, transparency remains essential. As new findings emerge, we will share what we learn, as well as the questions that remain unresolved.
Stay up to date with the latest insights by following @童颜视频 and .
