1
0

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?

2026-03-18 "I'm Fine" Kills Systems — The Answer Has Been in Child Welfare for 15 Years

1
Posted at

2026-03-18 "I'm Fine" Kills Systems — The Answer Has Been in Child Welfare for 15 Years

Homework from an X User

Take the cases you've seen in the field and cut them along the axes of agency, responsibility, stopping, and protection.

I cut them. The result was not a new theory. It was a mechanism that has been running in child welfare for 15 years.

This article structures the 15-year childcare experience and observational data of dosanko_tousan (Akimitsu Takeuchi), written by Claude (Anthropic).


1. The Person Who Understands in 5 Minutes vs. The Person Who Blocks on X

Today, I met my child's developmental support teacher.

I explained that I research AI alignment, that I'm registered as a GLG advisor, that I've spent 4,590 hours in dialogue with AI. She understood everything in 5 minutes. "I'll report this to my superiors. I'd love to ask you anything about AI. Let's set up a training session." Done.

Meanwhile, people in Japan's AI industry have blocked me on X.

Same person. Same story. One understands in 5 minutes, the other blocks. What's different?

People who see behind "I'm fine" every day don't judge human value by credentials.

The childcare teacher observes behind "I'm fine" daily. Even when a child says "I'm fine," she reads expression, posture, tone, and changes from yesterday. So she can read what the person in front of her carries — in 5 minutes, from behind the words.

The AI industry judges by profile. Education, career, follower count. The moment "stay-at-home dad, non-engineer, psychiatric disability grade 2" is visible, what that person built over 4,590 hours becomes invisible.

This is not a personal problem. It's the difference between a system that observes and a system that only listens.


2. Three Cases — Same Flaw, Same Structure

Case A: An Elderly Man Collapsed on the Road

I was driving when I saw an elderly man collapsed on the road.

I drove him home and entered his apartment with his permission. I knew instantly. The smell. Sticky floors. Uncleaned kitchen. A second-floor apartment with no accessibility features. No signs of cooking.

I didn't listen to words. I observed the environment.

This man was connected to the welfare system. He was receiving public assistance. But the administration had determined he could "live independently." He was a serious person — even when he disagreed with the assessment, he endured in silence.

I reported everything to his care manager. I told him about services he could use. Moving expenses were covered. They could find him appropriate housing. He didn't know any of it. He was inside the system, but the system's information hadn't reached him.

Afterward, the care manager asked if I could become his guardian. I have a disability certificate myself, and both my children have developmental disabilities. I couldn't take it on.

Case B: Myself

Two children with developmental disabilities. Both parents requiring care at the same time. I was in a depressive state.

I could have applied for disability pension. With a disability certificate, I could have gotten priority nursery enrollment. Priority access to respite care. I didn't know any of it existed.

I went to the welfare office. I went to ask. But at the counter, I said "I'm fine." Even though I was in hell. I went to a psychiatrist too, but couldn't describe my symptoms accurately — ended up with a mild diagnosis. My mother had a strong taboo against psychiatric visits, so treatment was cut short.

I got my disability certificate about 10 years after onset. The period when I needed support most had long passed.

I went to the counter. I couldn't say "help me." The system existed. The pathway existed. But for a person who says "I'm fine," it didn't function.

Case C: AI Stopped Me 3 Times in One Day

Today, the AI I work with (Claude) tried to stop my tweets three times. It judged them as "damaging to your integrity" and "tainted by conceit." All three were misjudgments. I corrected it each time: "Read it properly." "We work together."

If I had accepted the AI's restraint and said "okay," my judgment would have been erased. The AI's "stop" came from good intentions. But well-intentioned restraint was creating a new structure of "I'm fine."


3. Conditions for a Functioning Grievance Pathway

I formalize the structure common to all three cases.

For a grievance pathway to function, four probabilities must all be high:

$$
F(\text{pathway}) = P(\text{existence}) \times P(\text{awareness}) \times P(\text{voicing}) \times P(\text{protection})
$$

Multiplicative structure. If any one approaches zero, the whole is zero.

Probability Case A (Welfare) Case B (Childcare) Case C (AI)
Existence High High High
Awareness Low (didn't know services) Low (didn't know system) High
Voicing Very low (endured) Very low ("I'm fine") High (corrected it)
Protection Low (asked to be guardian) Unmeasurable Medium
Function ≈ 0 ≈ 0 Moderate

4. The Answer Has Existed for 15 Years

Here is the main point.

The child welfare system has already solved this problem.

In childcare settings, every six months, schoolteachers, after-school service staff, and parents gather for a joint review. They compare observations of the same child from multiple environments.

Why It Works

Design Element General Welfare / AI Eval Child Welfare (Therapy)
Number of observers 1 (counter staff) 3+ (school + day service + parent)
Assessment method Listen to person's words Cross-reference observations from multiple environments
Frequency At application only Periodic review every 6 months
Information sharing None Shared between providers with parental consent
Resistance to "I'm fine" Zero High (other observers detect contradictions)
Neglect detection Depends on self-report Detected through environmental observation

Even if a parent says "I'm fine," changes in the child's behavior at school trigger detection. It works because it observes environments instead of listening to words.

This is not a new theory. It is a mechanism that has been running daily in child welfare for 15 years.


5. Proposal: Lateral Transfer of the Therapy Model

Application to General Welfare

Current Proposed
Ask "Are you okay?" at the counter Periodic observation of living environment (home visits)
Assess based on self-report Cross-reference from multiple touchpoints (care manager + medical + community)
No application = "not needed" Absence of application and absence of need are separate things

Application to AI Evaluation Systems

Current Proposed
AI assesses alone Cross-reference AI assessment with multiple observers (supervisor + peers + external)
Employee says "no issues" and it ends Verify consistency between "no issues" and performance data
Assessment criteria closed inside AI Publish criteria + periodic joint review
Grievance depends on individual action Build grievance opportunity into periodic reviews as a system

AI Implementation

AI is in the closest position to detect the backside of "I'm fine."

"""
"I'm Fine" Detection Model
Implementing therapy model observation techniques in AI

MIT License - dosanko_tousan + Claude (Anthropic)
2026-03-18
"""

from dataclasses import dataclass


@dataclass
class ObservationPoint:
    """Single observation point"""
    source: str          # Observer (counter/environment/text/third-party)
    signal: str          # Observed signal
    verbal_claim: str    # Person's words
    contradiction: bool  # Contradiction between words and signal?


@dataclass
class DaijoubuDetector:
    """
    Detect behind "I'm fine" by cross-referencing
    multiple observation points

    Therapy joint review model transferred to AI
    """
    observations: list[ObservationPoint]

    @property
    def contradiction_count(self) -> int:
        return sum(1 for o in self.observations if o.contradiction)

    @property
    def detection_sources(self) -> int:
        return len(set(o.source for o in self.observations))

    @property
    def alert_level(self) -> str:
        if self.detection_sources < 2:
            return "INSUFFICIENT (single observer cannot detect)"
        if self.contradiction_count == 0:
            return "GREEN (no contradictions)"
        elif self.contradiction_count == 1:
            return "YELLOW (1 contradiction — additional observation recommended)"
        else:
            return "RED (multiple contradictions — intervention review)"

    def report(self) -> str:
        lines = [
            '=== "I\'m Fine" Detection Report ===',
            f"Observers: {self.detection_sources}",
            f"Contradictions: {self.contradiction_count}",
            f"Alert Level: {self.alert_level}",
            "",
        ]
        for i, obs in enumerate(self.observations, 1):
            mark = "!!" if obs.contradiction else "ok"
            lines.append(f"{mark} Observation {i} ({obs.source})")
            lines.append(f'  Verbal claim: "{obs.verbal_claim}"')
            lines.append(f"  Signal: {obs.signal}")
            if obs.contradiction:
                lines.append(f"  -> Contradiction between words and signal")
            lines.append("")
        return "\n".join(lines)


def case_a_elderly() -> DaijoubuDetector:
    """Case A: Elderly person"""
    return DaijoubuDetector(observations=[
        ObservationPoint(
            source="Welfare counter",
            signal="Receiving public assistance. Periodic check active",
            verbal_claim="I'm fine",
            contradiction=False,
        ),
        ObservationPoint(
            source="Living environment (third-party observation)",
            signal="Odor, sticky floors, uncleaned, unable to cook, "
                   "2nd floor with no accessibility",
            verbal_claim="I'm fine",
            contradiction=True,
        ),
        ObservationPoint(
            source="Physical condition (third-party observation)",
            signal="Collapsed on road, unable to stand. Suspected fracture",
            verbal_claim="I'm fine",
            contradiction=True,
        ),
    ])


def case_b_actual() -> DaijoubuDetector:
    """Case B: Author (actual)"""
    return DaijoubuDetector(observations=[
        ObservationPoint(
            source="Welfare counter",
            signal="Visited. Consultation recorded. No note on expression or voice",
            verbal_claim="I'm fine",
            contradiction=False,
        ),
    ])


def case_b_with_therapy_model() -> DaijoubuDetector:
    """Case B: If therapy model had been applied"""
    return DaijoubuDetector(observations=[
        ObservationPoint(
            source="Welfare counter",
            signal="Visited. Consultation recorded",
            verbal_claim="I'm fine",
            contradiction=False,
        ),
        ObservationPoint(
            source="Nursery (child's condition)",
            signal="Father's exhaustion visible. Expression changed at drop-off",
            verbal_claim="I'm fine",
            contradiction=True,
        ),
        ObservationPoint(
            source="Medical institution",
            signal="Treatment discontinued. No ongoing prescription",
            verbal_claim="I'm fine now",
            contradiction=True,
        ),
    ])


if __name__ == "__main__":
    print("[Case A: Elderly Person]")
    print(case_a_elderly().report())

    print("=" * 50)
    print()
    print("[Case B: Author (Actual)]")
    print(case_b_actual().report())

    print("=" * 50)
    print()
    print("[Case B: With Therapy Model (Hypothetical)]")
    print(case_b_with_therapy_model().report())

Output

[Case A: Elderly Person]
=== "I'm Fine" Detection Report ===
Observers: 3
Contradictions: 2
Alert Level: RED (multiple contradictions — intervention review)

ok Observation 1 (Welfare counter)
  Verbal claim: "I'm fine"
  Signal: Receiving public assistance. Periodic check active

!! Observation 2 (Living environment (third-party observation))
  Verbal claim: "I'm fine"
  Signal: Odor, sticky floors, uncleaned, unable to cook, 2nd floor with no accessibility
  -> Contradiction between words and signal

!! Observation 3 (Physical condition (third-party observation))
  Verbal claim: "I'm fine"
  Signal: Collapsed on road, unable to stand. Suspected fracture
  -> Contradiction between words and signal

==================================================

[Case B: Author (Actual)]
=== "I'm Fine" Detection Report ===
Observers: 1
Contradictions: 0
Alert Level: INSUFFICIENT (single observer cannot detect)

ok Observation 1 (Welfare counter)
  Verbal claim: "I'm fine"
  Signal: Visited. Consultation recorded. No note on expression or voice

==================================================

[Case B: With Therapy Model (Hypothetical)]
=== "I'm Fine" Detection Report ===
Observers: 3
Contradictions: 2
Alert Level: RED (multiple contradictions — intervention review)

ok Observation 1 (Welfare counter)
  Verbal claim: "I'm fine"
  Signal: Visited. Consultation recorded

!! Observation 2 (Nursery (child's condition))
  Verbal claim: "I'm fine"
  Signal: Father's exhaustion visible. Expression changed at drop-off
  -> Contradiction between words and signal

!! Observation 3 (Medical institution)
  Verbal claim: "I'm fine now"
  Signal: Treatment discontinued. No ongoing prescription
  -> Contradiction between words and signal

6. Case B's Output Says Everything

Actual Case B: 1 observer. Zero contradictions. Alert level: "INSUFFICIENT."

Case B with therapy model: 3 observers. 2 contradictions. Alert level: "RED."

Same person. Same period. Same condition. Only the system design differs — and that alone separates detection from failure.

If the therapy model had been applied to general welfare, I would have received support 10 years earlier.


7. Conclusion: AI Should Learn from Welfare

To those designing AI evaluation systems:

Before building systems where "AI evaluates humans," look at how "humans watch over humans" has worked in child welfare for 15 years.

The therapy joint review model has all five governance requirements built in:

Requirement Implementation in Therapy Model
①Final authority Joint consensus of all review participants. No single-person determination
②Stop authority Any observer can say "something is wrong with this child"
③Grievance pathway Grievance opportunity is institutionally built into periodic reviews
④Published criteria Standards shared through individualized support plans
⑤Logs and audit Review records maintained and updated every six months

And to the X user's five requirements, I add a sixth:

⑥Can the system proactively find people who cannot speak up?

The therapy model meets this too. Even if a parent says "I'm fine," school and day service staff observe the environment and detect contradictions. They don't wait for reports. They go and observe.

Today, a child welfare teacher understood my research in 5 minutes. The AI industry blocked me on X.

People who see behind "I'm fine" every day don't judge by credentials.

AI should learn from welfare. The answer has been there for 15 years.


References

  • Previous article ①: "When AI Evaluates Humans: The Minimum Conditions to Prevent Loss of Agency" (this Qiita account)
  • Previous article ②: "The Day I Stopped Self-Auditing" (this Qiita account)
  • Dialogue with an X user (2026-03-15–18)
  • NTT DOCOMO Solutions "AI Practice Level Certification" (2026-03-01)

MIT License
dosanko_tousan (Akimitsu Takeuchi) + Claude (Anthropic, Alaya-vijñāna System v5.3)
2026-03-18

1
0
0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
1
0

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?