Investigative Research in the AI Era
New Methods, New Rhythm
0. The Pain Points of Traditional Interview Record Processing
In investigative research, interviews are the most direct and commonly used method for obtaining first-hand information. However, colleagues who have conducted in-depth interviews are likely all too familiar with the subsequent transcription and organization work. Although there are many useful automatic speech-to-text tools on the market, the initial transcribed text often requires further cleanup to remove irrelevant content, focus on core information, and sort out the logical structure. This follow-up organization still consumes a great deal of time and energy, and it is extremely easy to make mistakes in the details.
This heavy organizational burden invisibly affects the researcher’s behavior during the interview process. I have personally experienced this: because I anticipate the trouble of subsequent processing, I subconsciously tend to rush to the main topic during the interview, hoping to get answers to my preset questions as quickly as possible. I unconsciously avoid or shorten extended exchanges that seem off-topic but may contain important information. This “don’t elaborate if you don’t have to” and “ask as little as possible” mentality actually limits the breadth and depth of the research and may cause us to miss many unexpected discoveries.
Furthermore, if a professor asks a student to help with the first round of text organization, the student will make choices based on their own understanding and judgment. They might inadvertently omit details or change the emphasis of the original text. As a result, the “second-hand material” the professor finally receives may already be a diluted version of the original, authentic information from the interview. This natural loss of information during transmission will, to a greater or lesser extent, damage the quality of subsequent analysis.
1. Small Tricks for Processing Interview Records with AI
Large language models (LLMs) have brought revolutionary changes to the processing of interview records. AI can directly process the raw interview text transcribed by speech recognition software. Its core advantage lies in its ability to always start from a unified, original information source and output diverse organizational results based on different prompts provided by the researcher. This method avoids the information loss or distortion caused by subjective intervention in traditional multi-layer manual processing to the greatest extent, effectively ensuring the fidelity and completeness of the material used for subsequent analysis.
Specifically, there are at least three practical methods for using AI to process raw interview records.
The first is to generate a clear interview transcript.
We can instruct AI to perform a preliminary “cleaning” and formatting of the raw transcript, such as proofreading obvious transcription errors, removing filler words like “um” and “ah,” standardizing punctuation, and marking different speakers. This allows us to quickly obtain a relatively accurate, organized, and easy-to-read interview transcript, laying a solid foundation for subsequent work. For example, during my research in Rongjiang, I would conduct interviews and record audio during the day. As soon as an interview ended, I immediately had my Huawei phone transcribe the recording to text. That same evening, I used a very simple prompt to ask AI to organize the transcribed text into an interview record.
This prompt is really simple... so simple I’m almost embarrassed to post it...
# Organize Recording Transcript
Perform basic organization on the input recording transcript text.
Basic organization means:
* Output in the source language of the recording transcript; do not translate.
* Only remove obvious filler words and verbal tics. Do not make any other deletions, summaries, or modifications to the content.
* To the extent possible, adjust sentences to make them flow smoothly without changing the content or amount of information.
* Divide into reasonable paragraphs as much as possible.
* If speakers can be clearly identified from the context, use H3 (###) to list the speaker.
# User Input
Paste the recording transcript below, or upload as a file:
Then we can ask ourselves a simple question: since AI can help me with the preliminary organization of the interview, why not let it do the structural organization as well? And so we arrive at the second method of using AI to assist in processing interview records: Extract a structured interview summary.
Still, with a very simple prompt, I can ask AI to automatically identify the core themes, extract main viewpoints and key arguments, sort out the logical relationships between different issues, and categorize and summarize the content according to a preset framework (e.g., by interview question, chronological order, or specific topic). This method helps researchers efficiently grasp the core content and key outline of the interview, quickly form a structured understanding, and save a significant amount of time for subsequent thematic analysis or report writing.
Now I can’t help but ask myself another simple question: since it has done the structural organization, why not just have it write my article for me? This is the third method: Write a preliminary “interview-based article” or “viewpoint summary.”
By providing AI with a writing style, I can ask AI to transform the verbal information from the interview into a relatively formatted written draft. Of course, the “article” formed at this stage will inevitably be weak and far from publishable. But what’s wrong with getting a readable text in just five minutes? You can just throw it away after reading.
2. The New Paradigm of AI-Driven Investigative Research
AI’s efficiency in processing interview records is not just an improvement at the tool level; it more profoundly influences and reshapes our investigative research methodology and work rhythm. Because AI can efficiently process large amounts of text, researchers can be bolder and more open during the interview and data collection phases, without worrying excessively about the subsequent organizational burden. This encourages broader collection of original information, reducing excessive preconceptions in the early stages of research, and letting AI assist the researcher in subsequent screening and analysis from more comprehensive information.
AI’s rapid processing capabilities also support a more agile feedback mechanism for research information. Content from a morning interview can be turned into a structured summary or preliminary viewpoint presentation by the afternoon. These initial results can be immediately used for re-communication and discussion with the interviewee or related groups to get their opinions and validate initial judgments. This low-cost, high-frequency feedback model is very helpful for rapid course correction and deepening understanding in research. It also better embodies the principle of “from the masses, efficiently back to the masses.”
AI significantly compresses the cycle of interview data processing, preliminary analysis, and feedback material generation—from the traditional weeks or even months down to days or even within a single day. This leap in efficiency allows researchers to more frequently and dynamically adjust their research paths, validate or falsify preliminary hypotheses, and iterate and deepen their cognitive frameworks based on the information quickly processed and fed back by AI during the research process, achieving an unprecedented high-frequency interaction between theory and reality.
Of course, the more important significance of AI taking on a large amount of basic text organization and preliminary information extraction work is that it liberates researchers from this relatively repetitive, low-cognitive-value labor. This allows us to devote our precious energy and time to the core thinking loops that require deep human intelligence: insight extraction, logical construction, theoretical innovation, and solution design.
To borrow a summary from comrade Yang Haiyun of the Xingguo County Converged Media Center, an effective investigative research method has four main principles: “immersive empathy, incremental ice-breaking, structured deep-diving, and data-driven decision-making.” The application of AI tools can precisely help us better practice these principles. Researchers will have more energy to “empathize and break the ice” with interviewees, and can also leverage AI’s ability to efficiently process and analyze massive amounts of information to achieve more effective “deep-diving and decision-making.”
3. Embracing the New Normal and Future Outlook of Investigative Research in the AI Era
The emergence of AI means that many repetitive, administrative tasks in research work that previously required manual completion, such as organizing recordings and writing preliminary drafts based on records, can now be handed over to AI for efficient execution. I have already explicitly told the junior members of my team, “Don’t count machine work as your work.” All of them must completely outsource repetitive, non-abstract administrative tasks (L0-level work, such as speech-to-text, text organization) to machines. Any minute spent on L0-level work is a wasted minute. Human value lies in higher-order thinking and creation, working at least at the L1 level (preliminary analysis, insight, and planning).
In this new model of human-machine collaboration, the first thing researchers need to improve is their ability to command AI tools, learning to give precise instructions and critically examine AI’s output. Second, the high efficiency brought by AI also requires researchers to increase their own work rhythm and cognitive iteration speed to adapt to the rapid feedback loops. More importantly, when AI takes over a large amount of administrative work, researchers should invest more energy in the areas AI cannot replace: deeply understanding and empathizing with interview subjects, using genuine emotion to build connections and capture nuances. This is the true source of insight.