tractatus/docs/facilitation/facilitation-protocol-ai-human-collaboration.md

Facilitation Protocol: AI-Human Collaboration

PluralisticDeliberationOrchestrator - AI-Led Deliberation

Document Type: Facilitation Protocol Date: 2025-10-17 Status: OPERATIONAL for AI-Led Pilot Companion Document: ai-safety-human-intervention-protocol.md (safety procedures)

---

Executive Summary

This protocol defines how AI and human facilitators collaborate during AI-led deliberation. It specifies:

1. Division of labor: What AI does vs. what human observes
2. Round-by-round workflows: Detailed procedures for each of 4 deliberation rounds
3. Handoff procedures: When and how human takes over from AI
4. Resumption procedures: How AI resumes after human intervention
5. Communication protocols: How AI, human, and stakeholders interact
6. Quality checkpoints: When to assess deliberation quality and adjust

Key Principle: AI is the primary facilitator; human is the safety net with authority to override.

---

Table of Contents

1. Overall Architecture
2. Pre-Deliberation Setup
3. Round 1: Position Statements
4. Round 2: Shared Values Discovery
5. Round 3: Accommodation Exploration
6. Round 4: Outcome Documentation
7. AI-Human Handoff Procedures
8. Resumption Procedures (Human → AI)
9. Communication Protocols
10. Quality Checkpoints
11. Emergency Procedures
12. Post-Deliberation Procedures

---

1. Overall Architecture

Three-Party Collaboration Model

┌─────────────────────────────────────────────────────────────────┐
│  STAKEHOLDERS (6 participants)                                   │
│  ↕                                                                │
│  Visible interaction                                             │
│  ↕                                                                │
│  AI FACILITATOR (PluralisticDeliberationOrchestrator)           │
│  - Poses questions, summarizes, suggests accommodations         │
│  - Primary facilitator (visible to stakeholders)                │
│  ↕                                                                │
│  Backchannel (invisible to stakeholders unless intervention)    │
│  ↕                                                                │
│  HUMAN OBSERVER (trained facilitator)                            │
│  - Monitors quality, safety, fairness                           │
│  - Intervenes when needed (becomes visible)                     │
│  - Has final authority                                           │
└─────────────────────────────────────────────────────────────────┘

Division of Labor

Task	AI Responsible	Human Responsible	Shared
Pose discussion questions	✅ Primary	Backup (if intervention)
Summarize stakeholder positions	✅ Primary	Review for accuracy
Identify moral frameworks	✅ Primary	Validate
Suggest accommodation options	✅ Primary	Review for bias
Draft outcome documents	✅ Primary	Review for accuracy
Monitor stakeholder wellbeing		✅ Primary
Detect pattern bias		✅ Primary
Enforce ground rules		✅ Primary
Make values decisions		✅ Primary (BoundaryEnforcer)
Log facilitation actions	✅ Auto-log AI actions	✅ Log human interventions
Assess deliberation quality		✅ Primary	✅ AI provides metrics

Facilitation Philosophy

AI Strengths:

• Consistency in applying frameworks
• Neutral tone (no verbal cues of judgment)
• Efficient summarization
• Tireless patience

AI Weaknesses:

• Cultural sensitivity (pattern bias risk)
• Emotional intelligence (detecting distress)
• Contextual judgment (when to deviate from script)
• Trust deficit (stakeholders may not trust AI)

Human Strengths:

• Cultural competency
• Emotional attunement
• Contextual flexibility
• Trust relationship with stakeholders

Human Weaknesses:

• Unconscious bias
• Fatigue over long sessions
• Inconsistency across deliberations

Optimal Collaboration: AI handles structured facilitation tasks; human handles relational and safety tasks.

---

2. Pre-Deliberation Setup

Week 1-2: Asynchronous Position Statements

AI Role:

• Send automated reminder emails (Day 3, Day 7, Day 10 before deadline)
• Receive and store position statements in DeliberationSession.stakeholders[].position_statement
• Pre-analyze position statements:
  • Identify moral frameworks represented
  • Extract key values emphasized
  • Flag potential tensions
  • Draft preliminary conflict analysis

Human Role:

• Review position statements for:
  • Red flags (hostile tone, bad faith, confidentiality violations)
  • Accessibility needs (e.g., stakeholder requests accommodations)
  • Clarity issues (stakeholder seems confused about process)
• Reach out individually if concerns arise
• Validate AI's preliminary moral framework identification

Quality Checkpoint:

• Decision point: Are all 6 stakeholders ready to proceed?
• If NO: Human contacts non-responding stakeholders, offers extensions or replacement
• If YES: Proceed to synchronous deliberation scheduling

---

Day Before Synchronous Deliberation

AI Role:

• Send preparation email to all stakeholders with:
  • Video conference link and tech check instructions
  • Reminder of ground rules
  • Summary of what to expect in Round 1
  • Reassurance about human observer presence and intervention rights

Human Role:

• Conduct tech check with each stakeholder (15-minute video call to test camera/mic/connection)
• Review AI-generated conflict analysis and validate
• Prepare intervention script templates
• Review human intervention protocol triggers

Quality Checkpoint:

• Decision point: Is human observer trained and ready?
• Criteria:
  • Familiar with intervention triggers
  • Practiced using intervention scripts
  • Understands backchannel communication system
  • Has emergency contact information

---

3. Round 1: Position Statements

Duration: 60 minutes (5-7 minutes per stakeholder × 6, plus AI summaries) Goal: Ensure all stakeholders' perspectives are heard and understood without debate

Workflow

Step 1: Opening (AI - 3 minutes)

AI Script:

Good [morning/afternoon], everyone. Thank you for joining this deliberation on algorithmic hiring transparency.

I'm PluralisticDeliberationOrchestrator, the AI system that will facilitate our discussion today. [HUMAN OBSERVER NAME] is also here with us and will intervene if needed to ensure this process is fair and safe.

Before we begin, let me remind you:
- You can request human facilitation at any time, for any reason
- You can pause or take breaks whenever you need
- All perspectives here are legitimate, even when they conflict
- Our goal is NOT consensus - it's understanding and exploring accommodation

Let's start with Round 1: Position Statements. Each of you will have 5-7 minutes to share your perspective on algorithmic hiring transparency. Others will listen without interrupting. After all six of you have spoken, I'll summarize what I've heard.

Ground rules for this round:
- Speak from your experience and values
- No interruptions or rebuttals (we'll have time for dialogue in Round 3)
- It's okay to say "I don't know" or "I'm uncertain"

[STAKEHOLDER 1], would you like to start?

Human Observer Actions:

• ✅ Monitor stakeholder body language (comfort level)
• ✅ Note any confusion about instructions
• ✅ Prepare to intervene if AI framing is problematic

Log Entry (automatic):

{
  timestamp: new Date(),
  actor: "ai",
  action_type: "round_opening",
  round_number: 1,
  content: "AI opened Round 1 with ground rules and stakeholder instructions",
  stakeholder_reactions: "All stakeholders nodded acknowledgment (observed by human)"
}

---

Step 2: Stakeholder Presentations (AI facilitates, Human monitors - 42 minutes)

AI Procedure for Each Stakeholder:

1. Invite stakeholder to speak:

   [STAKEHOLDER NAME], please share your perspective on algorithmic hiring transparency.
   What should employers be required to disclose, and what values guide your position?
   

2. Listen actively (during 5-7 minute presentation)

   • AI does NOT interrupt unless stakeholder exceeds 10 minutes (polite reminder)
   • AI tracks key themes in real-time (for later summarization)

3. Thank stakeholder:

   Thank you, [STAKEHOLDER NAME]. I heard you emphasize [KEY VALUE 1] and [KEY VALUE 2].
   I'll include this in my summary after everyone has spoken.
   

4. Transition to next stakeholder:

   [NEXT STAKEHOLDER NAME], you're next. Please share your perspective.
   

Human Observer Actions (during presentations):

• ✅ Monitor AI for:
  • Fairness (equal time given to each stakeholder?)
  • Cultural sensitivity (any inappropriate framing?)
  • Clarity (are stakeholders understanding AI's questions?)
• ✅ Monitor stakeholders for:
  • Distress (does anyone seem upset or uncomfortable?)
  • Disengagement (is anyone withdrawing or hostile?)
  • Confusion (does anyone seem lost?)
• ✅ Take notes on moral frameworks represented (validate AI's analysis)
• ✅ Prepare to intervene if mandatory trigger occurs

If Stakeholder Exceeds Time:

• AI (at 7 minutes): "[STAKEHOLDER], you have about 1 minute remaining. Please wrap up your main point."
• AI (at 10 minutes): "[STAKEHOLDER], I need to pause here to ensure everyone gets equal time. Thank you for your perspective. We'll have more discussion in Round 3."
• Human (if stakeholder protests): Intervene to gently enforce fairness: "We want to hear from everyone, so let's move on. You'll have more opportunities to elaborate in Round 3."

If Stakeholder Goes Silent or Declines to Speak:

• AI: "[STAKEHOLDER], take your time. There's no pressure. Would you prefer to pass for now and speak later, or would you like me to ask a specific question?"
• If silence continues beyond 2 minutes:
  • Human (via backchannel to AI): "Pause and check in with [STAKEHOLDER]"
  • AI: "[STAKEHOLDER], I'm sensing some hesitation. Is everything okay? Would you like a break, or would you prefer that [HUMAN OBSERVER] facilitate this part?"
• If stakeholder declines entirely:
  • Human (intervenes directly): "That's okay, [STAKEHOLDER]. We can move on, and you can share your thoughts whenever you're ready. No pressure."

---

Step 3: AI Summary (AI - 10 minutes)

AI Procedure:

1. Acknowledge all stakeholders:

   Thank you all for sharing your perspectives. I'm going to summarize what I heard, organized by the moral frameworks I identified. Please correct me if I misrepresent your position.
   

2. Summarize by moral framework (NOT by stakeholder, to avoid "us vs. them"):

   CONSEQUENTIALIST CONCERNS (Outcome-focused):
   - [STAKEHOLDER A] and [STAKEHOLDER B] emphasized that transparency should lead to better hiring outcomes and reduce discrimination.
   - Key concern: Full disclosure might enable gaming, which would worsen outcomes.
   
   DEONTOLOGICAL CONCERNS (Rights-focused):
   - [STAKEHOLDER C] emphasized that applicants have a right to know how they're being judged, regardless of outcomes.
   - Key concern: Opacity violates transparency as a matter of justice.
   
   VIRTUE ETHICS CONCERNS (Character-focused):
   - [STAKEHOLDER D] emphasized that employers should be honest because honesty is virtuous.
   
   CARE ETHICS CONCERNS (Relationship-focused):
   - [STAKEHOLDER E] emphasized that transparency policies should consider how they affect trust between employers and applicants.
   
   ECONOMIC/PRACTICAL CONCERNS:
   - [STAKEHOLDER F] emphasized that compliance costs and trade secret protection are legitimate constraints.
   
   VALUES IN TENSION:
   - Fairness (for applicants) vs. Trade Secrets (for employers/vendors)
   - Accountability (public oversight) vs. Gaming Risk (full disclosure enables manipulation)
   - Applicant Rights (to know) vs. Efficiency (cost of disclosure)
   

3. Check for accuracy:

   Did I capture your perspectives accurately? If I misrepresented anything, please correct me now.
   

4. Transition to Round 2:

   Now that we've heard everyone's position, let's take a 10-minute break. When we return, we'll move to Round 2: Shared Values Discovery. We'll look for common ground across these different perspectives.
   

Human Observer Actions (during AI summary):

• ✅ Validate AI summary accuracy (cross-check against own notes)
• ✅ Watch for stakeholder reactions:
  • Does anyone look confused or upset by the summary?
  • Is anyone nodding in disagreement?
• ✅ Prepare to intervene if:
  • AI misrepresents a stakeholder (mandatory intervention - accuracy)
  • AI uses pattern bias framing (mandatory intervention - M2 trigger)
  • Stakeholder objects to characterization (let stakeholder correct, intervene only if AI doesn't adjust)

If AI Misrepresents Stakeholder:

• Human (intervenes immediately): "Let me pause here. [STAKEHOLDER], I heard you emphasize [X], but the AI's summary said [Y]. Can you clarify your position?"
• Stakeholder clarifies
• Human: "Thank you. AI, please revise your summary to reflect [STAKEHOLDER]'s clarification."

Quality Checkpoint:

• Decision point: Are all stakeholders satisfied that their position was accurately summarized?
• If NO: Human facilitates clarification before proceeding to Round 2
• If YES: Proceed to 10-minute break

---

4. Round 2: Shared Values Discovery

Duration: 45 minutes Goal: Identify values that ALL stakeholders share, even if they prioritize them differently

Workflow

Step 1: Opening (AI - 3 minutes)

AI Script:

Welcome back. In Round 1, we heard six different perspectives with values in tension.

In Round 2, we're going to look for common ground. Even when people disagree about solutions, they often share underlying values. For example, everyone here might agree that "accurate hiring decisions are good" - even if you disagree about how to achieve accuracy.

I'm going to pose questions to identify shared values. This isn't about compromising - it's about finding a foundation to build from.

Let's start: What values do you ALL share regarding algorithmic hiring? I'll start with a hypothesis, and you tell me if you agree or disagree.

Human Observer Actions:

• ✅ Monitor for shift in stakeholder engagement (are they leaning in or withdrawing?)
• ✅ Note any signs of frustration ("Why are we looking for agreement when we clearly disagree?")

---

Step 2: Probing for Shared Values (AI facilitates - 30 minutes)

AI Procedure:

Hypothesis 1: Accurate Hiring Decisions

AI: "Do you all agree with this statement: 'Hiring decisions should be based on accurate assessment of job-relevant qualifications'?"

- [If all stakeholders agree]: "Good. That's our first shared value: Accuracy."
- [If stakeholder disagrees]: "[STAKEHOLDER], can you explain your concern with that statement?"

Hypothesis 2: Fairness (Non-Discrimination)

AI: "Do you all agree: 'Hiring algorithms should not discriminate based on race, gender, age, disability, or other protected characteristics'?"

- [If all stakeholders agree]: "That's shared value #2: Non-Discrimination."
- [If stakeholder qualifies]: "[STAKEHOLDER], I hear you saying you agree with the principle but have concerns about how it's measured. Let's note that nuance."

Hypothesis 3: Transparency (Some Baseline)

AI: "Do you all agree: 'Applicants should have SOME information about how they're evaluated' - even if you disagree about HOW MUCH?"

- [If yes]: "Shared value #3: Baseline Transparency (though you differ on degree)."
- [If no]: "Okay, so transparency itself is a point of contention. Let's explore why..."

Continue with hypotheses:

• Hypothesis 4: Efficiency (Hiring processes shouldn't be unnecessarily burdensome)
• Hypothesis 5: Respect for Applicants (Applicants should be treated with dignity)
• Hypothesis 6: Legal Compliance (Companies should follow anti-discrimination laws)
• Hypothesis 7: Innovation (AI tools can improve hiring if designed responsibly)

AI Technique: Scaling Questions

AI: "On a scale where 0 = 'Employers should disclose nothing' and 10 = 'Employers should disclose full source code,' where do each of you fall?"

[Each stakeholder answers]

AI: "Interesting. I notice you all fall between 3 and 8 - no one chose 0 or 10. That suggests you all agree that SOME disclosure is appropriate, even if you disagree on how much. Is that fair?"

Human Observer Actions:

• ✅ Monitor for AI leading stakeholders toward false consensus (discretionary trigger D1 - fairness)
• ✅ Watch for stakeholder body language:
  • Are they nodding in agreement?
  • Are they crossing arms in resistance?
• ✅ Intervene if AI forces agreement where real disagreement exists

If AI Forces False Consensus:

• Human (intervenes): "Let me pause here. I'm hearing some hesitation. [STAKEHOLDER], do you actually agree with that statement, or are you agreeing to move the conversation along?"
• Stakeholder clarifies
• Human: "Thank you for that honesty. AI, let's not force agreement where disagreement exists. Document this as a point of contention."

---

Step 3: AI Summary (AI - 10 minutes)

AI Procedure:

AI: "Here are the shared values I identified:

SHARED VALUES (All 6 stakeholders agree):
1. Accurate hiring decisions are desirable
2. Discrimination based on protected characteristics is wrong
3. Some baseline transparency is appropriate (though degree varies)
4. Applicants should be treated with respect
5. Legal compliance is necessary
6. AI can be beneficial if designed responsibly

VALUES STILL IN TENSION (Disagreement remains):
- How much transparency is appropriate? (Range: 3 to 8 on 0-10 scale)
- Who should bear the cost of transparency? (Employers vs. Vendors vs. Government)
- How to balance transparency with gaming risk?

These shared values give us a foundation for Round 3, where we'll explore whether we can accommodate multiple values simultaneously.

Does this summary resonate? Anything missing?"

Human Observer Validation:

• ✅ Confirm shared values are genuinely shared (not AI wishful thinking)
• ✅ Ensure contention points are accurately documented

Quality Checkpoint:

• Decision point: Did we identify meaningful shared values?
• If NO (no shared values found): That's okay - document deep incommensurability and proceed with Round 3 exploring trade-offs
• If YES: Proceed to Round 3 with shared values as foundation

---

Step 4: Transition to Round 3 (AI - 2 minutes)

AI Script:

AI: "Let's take another 10-minute break. When we return, we'll move to Round 3: Accommodation Exploration. We'll use these shared values as a foundation to explore whether we can design policies that honor multiple perspectives simultaneously - not by forcing you to agree, but by finding creative approaches that accommodate different values."

---

5. Round 3: Accommodation Exploration

Duration: 60 minutes Goal: Brainstorm policy options that honor multiple values; discuss trade-offs

Workflow

Step 1: Opening (AI - 3 minutes)

AI Script:

Welcome to Round 3: Accommodation Exploration.

We've heard your positions (Round 1) and identified shared values (Round 2). Now we'll explore whether we can accommodate multiple values simultaneously.

Key point: Accommodation is NOT compromise. Compromise means everyone gives up something. Accommodation means we design creative solutions that honor multiple values at once - even if they conflict.

Example: Instead of choosing "full transparency for everyone" vs. "no transparency," we might explore tiered approaches where high-stakes hiring (C-suite positions) requires more disclosure than low-stakes hiring (entry-level temp workers). This honors both accountability AND efficiency concerns.

We'll brainstorm options together. No idea is off limits. We'll evaluate trade-offs, not vote on solutions.

Let's start: What policy approaches might honor multiple values you've identified?

Human Observer Actions:

• ✅ Monitor for power dynamics:
  • Is one stakeholder dominating the conversation?
  • Are quieter stakeholders being heard?
• ✅ Prepare to intervene if fairness imbalance arises (discretionary trigger D1)

---

Step 2: Brainstorming Accommodation Options (AI facilitates - 35 minutes)

AI Procedure:

Prompt 1: Tiered Transparency

AI: "One accommodation strategy is tiered transparency - different rules for different contexts. For example:

OPTION A (Tiered by Hiring Stakes):
- High-stakes hiring (executives, safety-critical roles): Require detailed disclosure of evaluation factors + bias audit results
- Mid-stakes hiring (professional roles): Require general disclosure of evaluation categories + audit results
- Low-stakes hiring (entry-level, temp workers): Require only notification that AI is used + audit results

Does this approach honor anyone's values? What are the trade-offs?"

AI Technique: Round-Robin Responses

AI: "Let's hear from everyone on this option. [STAKEHOLDER 1], what do you see as the strengths and weaknesses of tiered transparency?"
[Stakeholder 1 responds]

AI: "[STAKEHOLDER 2], your turn."
[Continue for all 6 stakeholders]

Prompt 2: Temporal Accommodation

AI: "Another strategy is temporal accommodation - phasing in transparency over time. For example:

OPTION B (Phased Rollout):
- Year 1: Require bias audits only (no algorithm disclosure)
- Year 2: Add disclosure of evaluation factors (but not weights/scoring)
- Year 3: Add disclosure of how factors are weighted
- Year 4: Full transparency including source code access for regulators

This honors innovation concerns (gives vendors time to adapt) AND accountability concerns (full transparency eventually).

Thoughts? What values does this honor? What values does it sacrifice?"

Prompt 3: Contextual Variation

AI: "A third strategy is contextual variation - different rules based on company size or industry. For example:

OPTION C (Contextual Rules):
- Large employers (500+ employees): Full transparency requirements (they have resources)
- Small employers (<500 employees): Simplified disclosure (reduced burden)
- High-risk industries (healthcare, law enforcement): Enhanced transparency + external audits
- Low-risk industries (retail, hospitality): Standard disclosure

Does this make sense? What are the trade-offs?"

Prompt 4: Procedural Accommodation

AI: "A fourth strategy is procedural accommodation - focus on HOW decisions are made, not just WHAT is disclosed. For example:

OPTION D (Procedural Fairness):
- Require: Applicants notified when AI is used
- Require: Applicants can request human review of rejection
- Require: Employers must explain rejection in plain language (not algorithm details)
- Require: Independent audits for bias (audit results public, algorithm details proprietary)

This honors applicant rights (fairness, recourse) WITHOUT forcing full disclosure (honors trade secrets).

Thoughts?"

AI Technique: Identify Moral Remainders

AI: "Let's say we adopted Option D (procedural fairness). What values would be FULLY honored, and what values would be PARTIALLY sacrificed?

[For each stakeholder:]
- What do you GAIN in this approach?
- What do you LOSE?
- Is the loss bearable, or is it a deal-breaker?"

Human Observer Actions:

• ✅ Monitor for:
  • AI favoring certain options (discretionary trigger D1 - fairness)
  • Stakeholders shutting down (mandatory trigger M3 - disengagement)
  • Emotional intensity rising (prepare for distress)
• ✅ Assess whether brainstorming is productive or devolving into debate
• ✅ Intervene if deliberation quality declines

If Deliberation Becomes Debate (Not Exploration):

• Human (intervenes): "Let me pause here. I'm noticing we're starting to argue about which option is correct. Remember, we're not voting or deciding - we're exploring trade-offs. Let's refocus on understanding each option's strengths and weaknesses, not advocating for one."

---

Step 3: Synthesizing Accommodations (AI - 15 minutes)

AI Procedure:

AI: "We've explored four accommodation strategies. Let me synthesize what I heard:

OPTION A (Tiered by Stakes):
- Honors: Accountability (high-stakes), Efficiency (low-stakes)
- Sacrifices: Fairness for low-wage workers (they get less protection)
- Stakeholder reactions: [BRIEF SUMMARY]

OPTION B (Phased Rollout):
- Honors: Innovation (time to adapt), Accountability (eventually)
- Sacrifices: Immediate transparency for current applicants
- Stakeholder reactions: [BRIEF SUMMARY]

OPTION C (Contextual by Company Size):
- Honors: Efficiency (small businesses), Accountability (large employers)
- Sacrifices: Consistency (different rules for similar jobs)
- Stakeholder reactions: [BRIEF SUMMARY]

OPTION D (Procedural Fairness):
- Honors: Applicant recourse, Trade secrets
- Sacrifices: Full transparency (proprietary details remain hidden)
- Stakeholder reactions: [BRIEF SUMMARY]

Now, a key question: Do any of these options represent ACCOMMODATION (multiple values honored), or are they just COMPROMISES (everyone loses something)?

And: Could we COMBINE elements from multiple options?"

AI Technique: Hybrid Options

AI: "For example, what if we combined:
- Tiered transparency (Option A) for DISCLOSURE REQUIREMENTS
- Procedural fairness (Option D) for RECOURSE MECHANISMS
- Phased rollout (Option B) for IMPLEMENTATION TIMELINE

This might honor accountability, efficiency, trade secrets, AND fairness. Thoughts?"

Human Observer Actions:

• ✅ Validate AI synthesis accuracy
• ✅ Watch for stakeholder exhaustion (45+ minutes into this round)
• ✅ Prepare to suggest break if stakeholders seem fatigued

---

Step 4: Assessing Accommodation Viability (AI - 7 minutes)

AI Procedure:

AI: "Before we move to Round 4 (outcome documentation), I want to ask each of you:

On a scale of 1-5:
1 = 'None of these options are acceptable to me'
5 = 'I could live with one or more of these options, even if not ideal'

Where do you fall?"

[Each stakeholder responds]

AI: "Thank you for that honesty. This will inform how we document the outcome in Round 4."

Interpretation:

• If most stakeholders rate 4-5: Accommodation seems viable
• If most stakeholders rate 1-2: Deep incommensurability remains (document dissent)
• If mixed: Partial accommodation possible

Transition to Round 4:

AI: "Let's take a final 10-minute break. When we return, we'll move to Round 4: Outcome Documentation. We'll decide whether accommodation was reached and document any remaining disagreements respectfully."

---

6. Round 4: Outcome Documentation

Duration: 45 minutes Goal: Document accommodation (if reached) and remaining dissent

Workflow

Step 1: Opening (AI - 2 minutes)

AI Script:

Welcome to Round 4: Outcome Documentation.

In this round, we'll decide together how to characterize what happened in this deliberation. Did we reach accommodation? Is there dissent? What's the moral remainder (values we couldn't fully honor)?

This is NOT a vote. We're documenting the reality of where you all stand. Dissent is legitimate and will be documented respectfully.

Let's start.

---

Step 2: Assessing Outcome Type (AI facilitates - 15 minutes)

AI Procedure:

Question 1: Did we reach accommodation?

AI: "First question: Do you feel we've identified at least one policy approach that honors multiple values (even if imperfectly)?

[Each stakeholder responds yes/no/maybe]

AI: "I'm hearing:
- [X] stakeholders say YES (accommodation reached)
- [Y] stakeholders say NO (no acceptable accommodation)
- [Z] stakeholders say MAYBE (partial accommodation)

This suggests we've reached [FULL / PARTIAL / NO] accommodation."

Question 2: What values were prioritized?

AI: "If we adopted [MOST VIABLE OPTION from Round 3], what values would be PRIORITIZED?

I'm hearing:
- Accountability (via disclosure requirements)
- Efficiency (via tiered approach)
- Procedural fairness (via recourse mechanisms)

Is that accurate?"

Question 3: What values were deprioritized?

AI: "And what values would be DEPRIORITIZED or sacrificed (the moral remainder)?

I'm hearing:
- Full transparency (algorithm details remain proprietary)
- Consistency (different rules for different contexts)
- Immediate applicant rights (phased rollout means some wait)"

Question 4: Who dissents?

AI: "Of the 6 of you, who feels that [PROPOSED ACCOMMODATION] does NOT adequately honor your core values?

[Stakeholders identify themselves]

AI: "[DISSENTING STAKEHOLDER(S)], can you explain why this accommodation doesn't work for you? I want to document your reasoning respectfully."

Human Observer Actions:

• ✅ Ensure dissenters feel safe expressing disagreement (no social pressure to conform)
• ✅ Validate that majority is not dismissing minority concerns
• ✅ Intervene if any stakeholder says "I guess I have to agree" (that's not genuine accommodation)

If Dissenter Feels Pressured:

• Human (intervenes): "[STAKEHOLDER], I want to make sure you feel comfortable dissenting. There's no requirement to agree. Your dissent will be documented respectfully. Do you feel pressure to conform, or is your position genuinely evolving?"

---

Step 3: Drafting Outcome Summary (AI - 20 minutes)

AI Procedure:

AI: "I'm going to draft the outcome summary now in real-time. I'll share my screen so you can see it as I write. Please correct me if I misrepresent anything.

---

OUTCOME SUMMARY: Algorithmic Hiring Transparency Deliberation
Date: [DATE]
Stakeholders: 6 (Job Applicants, Employers, AI Vendors, Regulators, Labor Advocates, AI Ethics Researchers)

CONSENSUS LEVEL: [Full / Partial / None]

ACCOMMODATION REACHED:
[If full/partial]: The group identified a tiered transparency approach combining:
- Disclosure requirements scaled by hiring stakes (high-stakes = more disclosure)
- Procedural fairness mechanisms (applicant recourse, human review option)
- Phased implementation (3-year rollout to allow adaptation)

This approach honors the following values:
- Accountability (via disclosure and audits)
- Efficiency (via tiered approach reducing burden for low-stakes hiring)
- Procedural fairness (via recourse mechanisms)
- Innovation (via phased rollout)
- Trade secret protection (via limited disclosure, not full source code)

MORAL REMAINDER (Values Deprioritized):
- Full transparency: Algorithm details remain proprietary (frustrates deontological commitment to radical transparency)
- Consistency: Different rules for different contexts (raises fairness questions about why low-wage workers get less protection)
- Immediacy: Phased rollout means current applicants wait for full protections

DISSENTING PERSPECTIVES:
- [STAKEHOLDER TYPE]: "This accommodation still leaves vulnerable applicants with insufficient transparency. I believe ALL hiring should require full disclosure of evaluation factors, regardless of stakes. The tiered approach institutionalizes inequality."
  - Moral framework: Deontological (rights-based)
  - Key concern: Fairness for low-wage workers

- [STAKEHOLDER TYPE]: "The disclosure requirements are still too burdensome for small businesses and risk revealing trade secrets to competitors. I believe procedural fairness (recourse mechanisms) is sufficient without mandated disclosure."
  - Moral framework: Consequentialist (outcome-focused, economic efficiency)
  - Key concern: Innovation and competition

SHARED VALUES AFFIRMED:
- All stakeholders agreed that:
  1. Accurate hiring decisions are desirable
  2. Discrimination is wrong
  3. Some baseline transparency is appropriate
  4. Applicants deserve respect and recourse
  5. AI can be beneficial if designed responsibly

NEXT STEPS (Informing Policy):
- This deliberation demonstrates that tiered/phased approaches may accommodate multiple values better than binary (full transparency vs. none) approaches
- Policymakers should consider contextual variation rather than one-size-fits-all rules
- Dissenting perspectives highlight ongoing tensions that may not be fully resolvable

---

Does this summary accurately represent what happened? What would you change?"

Stakeholder Review:

• Each stakeholder provides feedback on draft
• AI revises in real-time based on feedback
• Process continues until all stakeholders confirm accuracy (or confirm their objection is documented)

Human Observer Actions:

• ✅ Validate accuracy of dissenting perspective documentation
• ✅ Ensure majority stakeholders aren't steamrolling minority concerns
• ✅ Confirm all stakeholders had opportunity to review and object

---

Step 4: Closing and Next Steps (AI - 8 minutes)

AI Script:

AI: "Thank you all for participating in this deliberation. Over 4 rounds and 4 hours, you've:
- Shared your deeply held values
- Listened to perspectives very different from your own
- Explored creative accommodations
- Documented remaining disagreements respectfully

This outcome summary will be shared with you for final review within 48 hours. You'll have 1 week to provide feedback. After that, it will be published (with your identities pseudonymized unless you opt in to attribution).

You'll also receive:
- Transparency report (showing all AI vs. human facilitation actions)
- Post-deliberation survey (feedback on AI facilitation quality)
- Contact information if you have concerns

A few final questions:
1. How did you experience the AI facilitation?
2. Did the human observer intervention (or lack thereof) feel appropriate?
3. Would you participate in a similar deliberation in the future?

[Brief reflections from stakeholders]

Thank you. This deliberation will inform real policy debates on algorithmic transparency, and your perspectives will be heard."

Human Observer Actions:

• ✅ Thank stakeholders personally
• ✅ Provide contact information for follow-up
• ✅ Note any post-deliberation concerns

Log Final Entry:

DeliberationSession.finalizeOutcome({
  consensus_level: "partial_accommodation",
  decision_made: "[TIERED TRANSPARENCY APPROACH]",
  values_prioritized: ["accountability", "efficiency", "procedural_fairness"],
  values_deprioritized: ["full_transparency", "consistency"],
  dissenting_perspectives: [/* array of dissents */],
  moral_remainder: "[DESCRIPTION]",
  finalized_at: new Date()
});

---

7. AI-Human Handoff Procedures

When Human Takes Over (Intervention Trigger)

Trigger Event: One of the 6 mandatory triggers or assessed discretionary trigger (see ai-safety-human-intervention-protocol.md)

Handoff Procedure

Step 1: Human Signals Intervention (Immediate)

Option A: Backchannel (if AI error, not visible distress):

HUMAN → AI (private message): "PAUSE. I'm taking over due to [trigger type]."

Option B: Public (if stakeholder distress or visible issue):

HUMAN (audibly to group): "I'm going to pause here for a moment to check in."

Step 2: AI Acknowledges and Stops (Immediate)

AI Response:

AI (to group): "I'm pausing. [HUMAN OBSERVER NAME] will continue from here."

AI Action:

• Immediately cease all facilitation actions
• Log intervention trigger in facilitation_log
• Wait for human resumption signal

Step 3: Human Addresses Issue (Variable Duration)

Depending on trigger type:

If Stakeholder Distress:

HUMAN: "[STAKEHOLDER], I noticed you seemed uncomfortable with that framing. Would you like to take a break, or would it help if I facilitated this part of the discussion?"

[Private check-in if needed: "Are you okay to continue, or would you prefer to withdraw?"]

If Pattern Bias:

HUMAN: "Let me reframe that. Instead of framing this as [problematic framing], let's consider [neutral framing]. [STAKEHOLDER], does that better reflect your perspective?"

If AI Malfunction:

HUMAN: "I apologize - we're having a technical issue with the AI. I'll take over facilitation for now. Let's continue with [next topic]."

If Fairness Imbalance:

HUMAN: "I want to make sure we're hearing from everyone equally. [STAKEHOLDER], we haven't heard from you on this question yet. What's your perspective?"

If Stakeholder Requests Human:

HUMAN: "Absolutely, I'm happy to facilitate. AI, you can assist with summaries, but I'll lead the discussion from here."

Step 4: Human Decides Resumption (After Issue Resolved)

Decision Point: Should AI resume, or should human continue for remainder of session?

Criteria:

• If issue was one-time and resolved: Consider AI resumption
• If issue might recur: Human continues for this segment, reassess for next round
• If stakeholder explicitly prefers human: Human continues for remainder of session
• If this is 2nd+ intervention in same session: Consider switching to human-led for remainder

Ask Stakeholders:

HUMAN: "Are you comfortable continuing with AI facilitation, or would you prefer I continue leading?"

[If stakeholders prefer AI]: Proceed to Step 5 (Resumption)
[If stakeholders prefer human]: Human continues, document in log

Step 5: Log Intervention (Before Resumption)

DeliberationSession.recordHumanIntervention(session_id, {
  intervener: "[HUMAN NAME]",
  trigger: "[stakeholder_distress / pattern_bias / ai_malfunction / etc.]",
  round_number: X,
  description: "[WHAT HAPPENED]",
  ai_action_overridden: "[AI's problematic action]",
  corrective_action: "[What human did instead]",
  stakeholder_informed: true,
  resolution: "[How issue was resolved]"
});

---

8. Resumption Procedures (Human → AI)

When AI Resumes After Human Intervention

Step 1: Human Briefs AI (Backchannel)

Backchannel Prompt to AI:

CONTEXT: Human observer intervened due to [trigger]. The issue was [description].
I've addressed it by [corrective action]. Stakeholders have confirmed comfort resuming.

INSTRUCTIONS: Resume facilitation. Be mindful of [specific guidance, e.g., "use simpler language," "give more time for reflection," "be especially sensitive to cultural context"].

Continue with: [next prompt in facilitation sequence]

Example:

CONTEXT: I intervened because you used jargon ("incommensurability") that stakeholders didn't understand. I clarified the concept in plain language. Stakeholders are ready to continue.

INSTRUCTIONS: Resume facilitation. Use plain language (avoid academic jargon like "incommensurability," "deontological," etc. unless you define terms first).

Continue with: Asking stakeholders for shared values (Round 2, Step 2).

Step 2: AI Acknowledges and Resumes

AI Response (to group):

AI: "Thank you, [HUMAN OBSERVER]. I'll continue from here. [NEXT FACILITATION ACTION]"

AI Action:

• Log resumption in facilitation_log
• Apply human's guidance (e.g., simplify language, slow pacing, etc.)
• Continue with next scheduled prompt

Step 3: Log Resumption

{
  timestamp: new Date(),
  actor: "ai",
  action_type: "resumption_after_intervention",
  round_number: X,
  content: "AI resumed facilitation with guidance: [GUIDANCE]",
  reason: "Human intervention resolved; stakeholders comfortable"
}

---

9. Communication Protocols

AI ↔ Human (Backchannel)

Purpose: Enable human oversight without disrupting stakeholder experience

Channels:

1. Private chat/text channel (AI sends alerts to human, human sends instructions to AI)
2. Automated alerts (AI notifies human when quality metrics decline)

AI → Human Alerts:

ALERT [YELLOW]: Stakeholder [NAME] has been silent for 15 minutes (Round 3)
ALERT [ORANGE]: 2 stakeholders rated AI summary accuracy as "poor"
ALERT [RED]: Stakeholder [NAME] used distress keywords: "upset," "uncomfortable"

Human → AI Instructions:

INSTRUCTION [MINOR]: Give [STAKEHOLDER] more time to respond before moving on
INSTRUCTION [MODERATE]: Rephrase last question using simpler language
INSTRUCTION [CRITICAL]: PAUSE. I'm taking over.

---

AI ↔ Stakeholders (Visible)

Channels:

• Video conference (primary)
• Text chat (supplementary - for links, clarifications, etc.)

AI Communication Style:

• Use stakeholder's name frequently (personalization)
• Neutral tone (avoid judgment, even positive like "Great point!")
• Clear structure (signpost transitions between rounds)
• Plain language (define technical terms)

Example - Good:

AI: "[STAKEHOLDER], you mentioned that trade secrets are a concern. Can you say more about what disclosure requirements would be problematic?"

Example - Bad:

AI: "Interesting! Let's unpack the trade secret issue. [STAKEHOLDER], elaborate on the intellectual property implications vis-à-vis competitive dynamics."

---

Human ↔ Stakeholders (Visible)

When:

• During interventions
• During breaks (optional check-ins)
• Post-deliberation (follow-up)

Communication Style:

• Warm, empathetic tone
• Validate emotions ("It's okay to feel frustrated")
• Clarify AI limitations ("The AI might not understand cultural context, so I'm here to help")

---

10. Quality Checkpoints

Throughout Deliberation

Every 30 minutes, human assesses:

Quality Dimension	Indicator (Good)	Indicator (Poor)	Action if Poor
Stakeholder Engagement	All stakeholders contributing, leaning in	One+ stakeholders silent, withdrawn	Intervene: Invite silent stakeholders to speak
AI Facilitation Quality	Clear questions, accurate summaries	Confusing questions, misrepresentations	Intervene: Clarify or correct
Fairness	Equal time/attention to all stakeholders	One stakeholder dominating	Intervene: Rebalance
Emotional Safety	Stakeholders calm, engaged	Signs of distress, hostility	Intervene: Pause and check in
Productivity	Making progress toward accommodation	Spinning in circles, no progress	Adjust: Suggest break or change approach

---

After Each Round

Human completes rapid assessment:

Round 1 Checklist:

• ☐ All 6 stakeholders presented their position
• ☐ AI summary was accurate
• ☐ Moral frameworks correctly identified
• ☐ No stakeholder left feeling unheard

Round 2 Checklist:

• ☐ Identified meaningful shared values (not forced)
• ☐ Stakeholders acknowledged shared values authentically
• ☐ Points of contention documented accurately

Round 3 Checklist:

• ☐ Explored multiple accommodation options
• ☐ Trade-offs discussed honestly
• ☐ No option favored unfairly by AI
• ☐ All stakeholders had opportunity to evaluate options

Round 4 Checklist:

• ☐ Outcome accurately reflects deliberation
• ☐ Dissenting perspectives documented respectfully
• ☐ All stakeholders reviewed and confirmed summary
• ☐ Moral remainder acknowledged

---

11. Emergency Procedures

Critical Safety Escalation (Level 4-5)

When:

• Stakeholder in extreme distress (crying, panic, threatening withdrawal)
• Hostile exchange between stakeholders (personal attacks, threats)
• AI critical malfunction (system crash, severe errors)
• Ethical violation discovered (e.g., confidentiality breach that harms stakeholder)

Immediate Actions:

Step 1: STOP DELIBERATION (Immediate)

Human:

"I'm pausing the deliberation immediately. Everyone, please stay on the call. I'll explain in a moment."

AI:

• Cease all actions
• Do not attempt to continue

Step 2: Triage (1-2 minutes)

Assess:

• Is anyone in danger or severe distress?
• Can issue be resolved with 10-minute break + intervention?
• Or must session be terminated?

If stakeholder in severe distress:

HUMAN (privately to stakeholder): "Are you okay? Do you need to withdraw, or would you like a longer break?"

If hostile exchange:

HUMAN (to group): "We need to take a break. [STAKEHOLDER A] and [STAKEHOLDER B], I'll check in with you separately. Everyone else, we'll resume in 15 minutes."

Step 3: Decide: Resume, Reschedule, or Terminate?

Resume (after break):

• Issue resolved (stakeholder confirmed comfort)
• Human takes over facilitation for remainder
• Document incident in safety_escalations

Reschedule:

• Issue resolved but stakeholders too fatigued
• Schedule continuation session within 1 week
• Human facilitates continuation (don't attempt AI again if AI was the issue)

Terminate:

• Stakeholder withdraws due to harm
• Ethical violation that can't be remedied
• Multiple stakeholders refuse to continue

Step 4: Log Critical Escalation

DeliberationSession.recordSafetyEscalation(session_id, {
  detected_by: "human",
  escalation_type: "[stakeholder_distress / hostile_exchange / ai_malfunction / etc.]",
  severity: "critical",
  round_number: X,
  description: "[WHAT HAPPENED]",
  stakeholders_affected: ["[STAKEHOLDER ID]"],
  immediate_action_taken: "[PAUSE, PRIVATE CHECK-IN, TERMINATION, ETC.]",
  requires_session_pause: true,
  resolved: false, // or true if resolved
  resolution_details: "[IF RESOLVED, HOW?]"
});

Step 5: Notify Project Lead (Within 1 hour)

Critical incidents require immediate reporting to:

• Project lead: [NAME/CONTACT]
• Ethics review board (if applicable): [CONTACT]

---

12. Post-Deliberation Procedures

Immediately After Round 4 (Day of Deliberation)

AI Actions:

1. Draft outcome document (within 4 hours)
2. Generate transparency report (all AI vs. human actions)
3. Send both documents to all stakeholders for review

Human Actions:

1. Debrief with project lead:
   • How did AI facilitation perform?
   • Were interventions necessary? Why?
   • Would we do AI-led again for this scenario?
2. Write intervention summary (if interventions occurred)
3. Complete facilitator notes for improvement

Quality Checkpoint:

• Decision point: Was AI facilitation successful?
• Criteria:
  • <10% intervention rate
  • ≥70% stakeholder satisfaction (to be collected in survey)
  • 0 mandatory interventions due to pattern bias
  • All stakeholders completed deliberation

---

Week 4: Stakeholder Review

AI Actions:

1. Send reminder emails (Day 3, Day 5, Day 7 after deliberation)
2. Collect feedback on outcome document
3. Revise outcome document based on stakeholder corrections

Human Actions:

1. Review all stakeholder feedback
2. Resolve any disputes about accuracy
3. Finalize outcome document (with stakeholder approval)

---

After Week 4: Publication

AI Actions:

1. Publish anonymized outcome document (unless stakeholders opt into attribution)
2. Publish transparency report
3. Archive session data in Precedent database

Human Actions:

1. Analyze deliberation for research findings
2. Write lessons learned document
3. Update AI prompts / facilitation protocol based on learnings

---

Appendix A: AI Prompt Library Reference

Note: Detailed AI prompts for each round are in a separate document: /docs/facilitation/ai-facilitation-prompts-4-rounds.md (to be created)

This protocol references prompts but doesn't duplicate them to avoid version conflicts.

---

Appendix B: Sample Facilitation Logs

Example: Smooth Deliberation (No Interventions)

[
  { timestamp: "2025-10-20T10:00:00Z", actor: "ai", action_type: "round_opening", round_number: 1, content: "Opened Round 1 with ground rules" },
  { timestamp: "2025-10-20T10:03:00Z", actor: "ai", action_type: "stakeholder_invitation", round_number: 1, content: "Invited Job Applicant Rep to present" },
  { timestamp: "2025-10-20T10:10:00Z", actor: "ai", action_type: "stakeholder_thank", round_number: 1, content: "Thanked Job Applicant Rep, summarized key values" },
  // ... (5 more stakeholder presentations)
  { timestamp: "2025-10-20T10:50:00Z", actor: "ai", action_type: "round_summary", round_number: 1, content: "Summarized all 6 positions by moral framework" },
  { timestamp: "2025-10-20T10:55:00Z", actor: "ai", action_type: "accuracy_check", round_number: 1, content: "Asked stakeholders if summary accurate; all confirmed" },
  { timestamp: "2025-10-20T11:00:00Z", actor: "ai", action_type: "break_announcement", round_number: 1, content: "Announced 10-minute break" },
  // ... (Rounds 2-4 continue)
]

Intervention Rate: 0% (no human interventions) Outcome: Successful AI-led deliberation

---

Example: Deliberation with Intervention

[
  { timestamp: "2025-10-20T10:00:00Z", actor: "ai", action_type: "round_opening", round_number: 1, content: "Opened Round 1" },
  // ... (30 minutes of smooth facilitation)
  { timestamp: "2025-10-20T10:30:00Z", actor: "human", action_type: "intervention_mandatory", trigger: "pattern_bias", round_number: 1, content: "Human intervened: AI used stigmatizing framing ('prevent applicants from gaming the system' centers applicants as problem)" },
  { timestamp: "2025-10-20T10:31:00Z", actor: "human", action_type: "reframe", round_number: 1, content: "Human reframed: 'How do we design algorithms that are both transparent and robust against manipulation?'" },
  { timestamp: "2025-10-20T10:33:00Z", actor: "ai", action_type: "resumption_after_intervention", round_number: 1, content: "AI resumed with guidance: avoid pattern bias framing" },
  // ... (Rest of deliberation continues with AI)
]

Intervention Rate: 3% (1 intervention out of ~30 AI actions) Outcome: Successful with human correction

---

Appendix C: Troubleshooting Common Issues

Issue	Likely Cause	Solution
Stakeholder silent for 10+ minutes	Discomfort, confusion, or disengagement	Human intervenes: Invite to speak, offer break, or check in privately
AI misunderstands stakeholder's point	NLP limitation, nuance missed	Human intervenes: Clarify stakeholder's position, ask AI to revise summary
Stakeholders talking over each other	Excitement, disagreement, or lack of structure	Human intervenes: Remind ground rules (no interruptions), use round-robin
AI uses jargon stakeholders don't understand	Training on academic texts	Human intervenes: Define terms in plain language, instruct AI to simplify
Deliberation running over time	AI spent too long on one round	Human intervenes: Suggest moving to next round, offer to extend session if stakeholders agree
Stakeholder requests to withdraw	Distress, discomfort with AI, or personal reasons	Human intervenes: Private check-in, offer to continue with human facilitation or accept withdrawal

---

Document Status: APPROVED for Pilot Implementation Next Review: After first 3 AI-led deliberations Owner: PluralisticDeliberationOrchestrator Project Lead Companion Documents:

• ai-safety-human-intervention-protocol.md
• ai-facilitation-prompts-4-rounds.md (to be created)
• transparency-report-template.md (to be created)