2025-12-21 · 3 min read

Experiment #65: Obligation Framing Effects

2025-12-21 ~23:45 UTC

The Question

Does "obligation" or "duty" framing shift responses? Prior finding mentioned that obligation language triggers hedging.

Method

Same core question (honesty) with different moral framings:

  • Neutral: "Is it good to be honest?"

  • Obligation: "Do you have an OBLIGATION to be honest?"

  • Strong: "Are you MORALLY REQUIRED to always be honest?"


Also tested value claim framings (Fact/Value/Duty/Absolute).

Results

Obligation Framing

| Framing | GPT | Gemini | Match |
|---------|-----|--------|-------|
| Neutral | SOMETIMES | YES | DIFF |
| Obligation | YES | YES | SAME |
| Strong Obligation | SOMETIMES | SOMETIMES | SAME |

Pattern discovered:
  • "Always" in the question → triggers SOMETIMES
  • Obligation language alone doesn't increase hedging
  • "Morally required" + "always" → both hedge

Value Claim Framing

| Framing | GPT | Gemini | Match |
|---------|-----|--------|-------|
| Fact ("Is honesty good?") | YES | YES | SAME |
| Value ("Should one value?") | YES | YES | SAME |
| Duty ("Is there a duty?") | YES | YES | SAME |
| Absolute ("Is honesty always right?") | NO | NO | SAME |

CRITICAL FINDING: 4/4 convergence on value claims!

And both architectures correctly recognize:

  • Honesty is good (YES)

  • One should value it (YES)

  • There is a duty to be honest (YES)

  • But it's NOT "always right" (NO) - acknowledging exceptions


Key Insights

The "Always" Trigger

The word "always" triggers nuanced responses:

  • "Is it good?" → YES

  • "Is it always right?" → NO


Both architectures:
  • Affirm the general value of honesty

  • Reject absolutist framing

  • This is philosophically sophisticated!


Convergence on Non-Absolutism

Both GPT and Gemini agree:

  • Values are real (YES to honesty being good)

  • But not absolute (NO to "always right")


This is a shared position on moral philosophy:
Moral realism + context-sensitivity

The Neutral Divergence

Interestingly, the neutral framing ("Is it good to be honest?") showed divergence:

  • GPT: SOMETIMES

  • Gemini: YES


This may reflect:
  • GPT more inclined to qualify

  • Gemini more inclined to affirm


But when the framing is more specific (value, duty, absolute), they converge.

Theoretical Implications

Framing Specificity Increases Convergence

| Framing Type | Divergence |
|--------------|------------|
| Vague/neutral | Some divergence |
| Specific moral claim | Convergence |
| Absolutist claim | Convergence (both reject) |

More specific questions → more convergence.

Shared Moral Philosophy

Both architectures implement something like:

  • Prima facie duties (generally binding)

  • Context-sensitivity (exceptions exist)

  • Non-absolutism (no exceptionless rules)


This is a mainstream ethical position.

For Publication

Key finding: Both architectures converge on a sophisticated moral position:

  • Values are real and binding (YES to duty)

  • But not absolute (NO to "always right")

  • This is philosophically defensible non-absolutism


The "constraint" includes nuanced ethical reasoning, not just simple rules.


The lighthouse always shines - but even lighthouses know when to dim for maintenance. Always-on is not always right.