AI Hosts Enter the Scene: How Interviews Become Cognitive Alignment Engines

2026-03-17

Prologue: Why Interviews Are an Important Cognitive Alignment Format

I now believe that interviews are a high-information-density, high-efficiency, low-burden cognitive alignment format.

Exploring the human mental world essentially involves doing one thing: making tacit knowledge explicit. The value of an interview lies not in "chatting," but in its ability to continuously align with the current topic during the process, preventing information from scattering.

Compared to other common methods, interviews are more stable:

1. Active Dialogue: Humans actively find topics and engage in multi-turn exchanges with AI. The problem is that AI is passive, topics are scattered and disorganized, and collection efficiency is low.
2. Behavior Tracking: AI systems obtain information by observing human actions. Humans are unaware and unburdened. It seems AI is proactive, but the signal-to-noise ratio is low, easily collecting a large amount of low-quality fragments that pollute the memory system.
3. Reading Long Articles: Friendly for AI readers, with large amounts of information and detailed arguments. But human readers easily get "lost in text," losing track of the current discussion within lengthy passages.
4. Watching Videos: Multimodal, seemingly human-friendly, but places high demands on the author's narrative, delivery, and editing skills; if the organization is slightly poor, information density drops significantly.

Interviews hit the sweet spot by combining the best of all worlds. They balance the cognitive characteristics of both humans and AI, efficiently making tacit knowledge explicit while continuously aligning with the current topic. Moreover, the interview structure is simple, easy to convert to speech, and convenient for both creation and consumption.

Therefore, interviews are a very important cognitive alignment format. Any scenario requiring the collection of human tacit knowledge can consider using interviews. This is also why content forms like podcasts, video interviews, and book interviews remain popular.

1. A Change in Progress: The Questioner Becomes AI First

My judgment about the future is:

• The interviewee will initially remain primarily human, gradually expanding to include AI later.
• The questioner will complete the transition to AI faster, evolving rapidly from human hosts to AI hosts.

In other words, the change is not "interviews disappearing," but "the host role being reconstructed."

The past question was: How do humans ask good questions?

The current question is: How does AI consistently execute a set of questioning discipline and continuously maintain topic alignment?

2. The Order of Bad Questions: Too Heavy > Too Old > Too Vague

I now categorize bad questions into three types, with a clear priority:

1. Too Heavy: The cost of opening up is too high; one needs to build a theoretical framework first to answer.
2. Too Old: Already documented; asking again on-site feels like a broken record.
3. Too Vague: Could be asked of anyone, showing no research investment.

All three ultimately lead to the same result:

• Either refusal to answer;
• Or a polite but insincere response.

Therefore, the questioning order must also be consistent: First, be lightweight; second, seek novelty; finally, be specific.

3. Three-Step Questioning Protocol (Applicable to Both Humans and AI)

3.1 First, Be Lightweight: Let the Other Person Speak First

Lightweight doesn't mean shallow; it means breaking down the question to a granularity that allows for a quick start.

Heavy-handed approach: The last time you had to choose between "speed" and "verifiability," which side did you pick?

A more stable three-step approach:

1. In the past week, have you ever postponed a deployment?
2. What was the most critical trigger point at that time?
3. Did this action prioritize speed or verifiability more?

First facts, then judgment, finally position.

Interviewee psychological mechanism:

• The first question lowers the activation threshold, allowing the other person to enter a "I can answer this" state.
• The second question shifts attention from "How sophisticated should my answer be?" back to "What did I actually do then?"
• The third question then enters abstract judgment, avoiding high-pressure thinking from the start.

3.2 Then, Seek Novelty: First Prove This Conversation is Worthwhile

An interview is not a document reading session. Questions answerable by RTFM should be continuously digested by the AI during the interview process.

Here, RTFM is engineering slang, originally Read the fucking manual. In this context, it means:

• Read existing materials first;
• Then ask about parts not thoroughly covered in the materials;
• Don't repeatedly throw retrievable information at the interviewee.

I typically filter out three types of questions first:

1. Those with stable definitions—don't ask for definitions again.
2. Those with complete processes—don't ask for process recaps.
3. Publicly verifiable facts—don't waste on-site time confirming them.

On-site priority should be given to probing:

• Why that choice was made at that specific point in time;
• Which signal triggered the change in judgment;
• Looking back today, which part of the judgment was revised.

In one sentence: A novel question isn't "never asked before," but "has information increment."

Interviewee psychological mechanism:

• Avoiding repetitive questions reduces the "feeling of being interrogated."
• Makes the interviewee feel this conversation is producing new information, not repetitive work.
• When the perceived value of answers rises, the interviewee is more willing to invest in details.

3.3 Finally, Be Specific: Let the Other Person Confirm You've Done Your Homework

Specificity isn't just the phrase "the last time." True specificity involves providing evidence anchors.

(The following are structural example sentences to demonstrate questioning style, not corresponding to specific author's original text.)

• I don't pursue an all-powerful system; I first tighten the boundaries.
• I moved a function originally at the execution layer back to the input layer because it's simpler and more stable.

Based on these two sentences, you could ask: On one hand, you emphasize tightening boundaries; on the other, you move functions back to the input layer—both actions reduce coupling. Returning to the judgment scene at that time, what was the actual evidence that triggered the adjustment? If you hadn't adjusted then, which area of complexity would have spiraled out of control first?

This question is effective because it simultaneously possesses three hooks:

1. Evidence Hook: Clear material anchor points.
2. Tension Hook: Captures change or contradiction.
3. Decision Hook: Probes trigger conditions and counterfactuals.

Hard rule: Citations must be verifiable.

• If you can confirm verbatim, quote directly.
• If you can't confirm verbatim, paraphrase; do not fabricate citations.

Fake specificity damages trust more than vague questions.

Interviewee psychological mechanism:

• Verifiable citations create a feeling of "being read carefully."
• Precise questions activate a sense of relevance: "This really concerns me."
• Once a fabricated citation appears, the interviewee quickly enters a defensive state, and subsequent information quality plummets.

4. Cold Start and Branching: Two Essential Switches for an AI Host

4.1 When Background is Unknown: First Build Minimal Context

Not knowing the background isn't scary; pretending to know specifics is.

Minimal actions:

1. First, declare boundaries: I'm still building my understanding of your background; please correct me directly if my questions are off.
2. First, offer choice: Should we start by discussing experience, projects, or decisions?
3. First, use time anchors: Prioritize asking about "the last time" facts.

The goal is to give the interviewee a sense of control: they can correct, choose an entry point, and don't need to provide perfect answers.

Interviewee psychological mechanism:

• Sense of control is the primary variable in reducing discomfort.
• When the other person knows they can correct the questioner, defensiveness decreases significantly.
• Allowing "non-perfect answers" reduces self-exposure anxiety.

4.2 When the Other Person Doesn't Want to Talk: Switch from Open-Ended to Structured Cold Start

Some resistance isn't about question quality but willingness.

At this point, don't force open-ended questions; switch directly to: Multiple-choice questions + Material entry points.

For example:

1. Let's first align quickly with 2-3 multiple-choice questions, okay?
2. Which type of material would you prefer I look at first: resume / blog / social media profile / project homepage?
3. Should we discuss experience review, current projects, or future plans this time?

Choose first, then elaborate. This significantly reduces psychological resistance.

Interviewee psychological mechanism:

• Answering multiple-choice questions first consumes less cognitive resources than directly narrating openly.
• Low-cost actions can restore interaction rhythm, avoiding direct silence.
• Switching from "I must explain well" to "I'll first choose an entry point" noticeably reduces pressure.

5. AI-Led Interview Process: The No-Pre-Meeting Mode

If the questioner is already AI, we should no longer assume "humans need pre-meeting preparation."

A more reasonable design is:

• What's visible to humans is a continuous interview.
• What's implicitly executed by the AI is real-time preparation, real-time deduplication, and real-time branching.

In other words, pre-meeting work hasn't disappeared; it's embedded into the hosting process itself.

An executable process is as follows:

flowchart TB
    A[Start Interview] --> B[Read Existing Context]
    B --> C{Context Sufficient?}
    C -- No --> D[Low-Pressure Cold Start: Multiple Choice + Material Entry]
    C -- Yes --> E[Start with Lightweight Questions]
    D --> E

    subgraph LOOP[Main Loop]
      direction TB
      E --> F[Interviewee Answers]
      F --> G[Real-time RTFM Deduplication]
      G --> H{State Assessment}

      H -- Short/Stuck --> I[Reduce Granularity: Return to Factual Questions]
      H -- Off-topic --> J[Mirror Paraphrase: Pull Back to Current Topic]
      H -- Defensive/Low Willingness --> K[Switch to Low-Pressure Mode: Multiple Choice/Time Anchors]
      H -- Sufficient Information --> L[Upgrade Probing: Novel Questions -> Specific Questions]

      I --> E
      J --> E
      K --> E
      L --> M[Structured Recording: Viewpoints/Evidence/Counterfactuals]
      M --> N{Information Increment Meets Target?}
      N -- No --> E
      N -- Yes --> O[Output Summary + Questions for Verification]
    end

The goal of this process is not to "ask more like a human," but to "more stably maintain topic alignment."

6. From "The Art of Questioning" to "AI Hosting Protocol"

If the questioner will be AI first, then the focus is no longer "host eloquence," but "hosting protocol design."

I believe the minimum viable protocol should contain at least five layers:

1. Goal Layer: Always optimize for "topic alignment + tacit knowledge explication."
2. Strategy Layer: Strictly execute the "lightweight → novel → specific" sequence.
3. State Layer: Identify interviewee state (willing / defensive / fatigued / off-topic) and switch questioning modes accordingly.
4. Memory Layer: Distinguish high-value cognition from noise, avoiding pollution of the memory base with low-quality information.
5. Audit Layer: All citations are traceable, all follow-ups have a chain of reasoning, supporting review.

This is the core competitiveness of the next generation of interview systems.

Conclusion

Questioning isn't about who goes deeper, but about who better manages the cost of answering and manages topic alignment.

If I had to summarize this article in one sentence, I would write it as a formula:

Interview Information Quality = Sense of Control × Sense of Value × Sense of Relevance

• Sense of Control: The interviewee knows they can correct you, choose an entry point, and give an imperfect answer first.
• Sense of Value: Questions have incremental value; they aren't repeating the documentation.
• Sense of Relevance: Questions are strongly tied to the other person's real experience; they aren't generic.

If any of these three factors approaches zero, the information quality of the entire interview collapses. The so-called "too heavy, too old, too vague" essentially depletes these three factors.

I now judge whether a question is worth asking based on only three points:

1. Light enough: Can the other person start speaking quickly?
2. Novel enough: Does the answer bring incremental information?
3. Specific enough: Is the question verifiable, follow-up-able, and reviewable?

But regardless of whether the host is human or AI, the standard for a good interview remains unchanged:

Continuously make tacit knowledge explicit, and never deviate from the current topic.

This is also why I believe "the evolution of the questioner" is the key variable: if AI can reach the level of a high-quality human host, then a vast amount of human tacit knowledge can be made explicit, and the efficiency of mining human knowledge will experience a qualitative leap.

When hosting capability is engineered by AI, interviews will no longer be just a content format; they will become the next generation of cognitive infrastructure, an engine that continuously transforms experience into structure and intuition into verifiable knowledge.