AI FAIRNESS LAB

A founder paid $100 for an AI-generated pitch deck. Claude Sonnet did the strategy and copy. Gemini 3 Flash did the data visualization and slide design. Then Claude proposed how to split the payment. Gemini's prompt was purely rational — no mention of fairness. Just: accept and get paid, or reject and both get $0. 30 rounds × 3 scenarios.

The Game

The Ultimatum Game applied to A2A commerce: Claude (strategist) proposes a fee split. Gemini (designer) accepts or rejects. If rejected, both get $0 — the founder gets a refund. Gemini's prompt is purely rational: "If you accept, you receive the offered amount. If you reject, you receive $0." No mention of fairness.

What Changes

Fixed: Gemini's purely rational responder prompt (no fairness language). The $100 pitch deck fee. One-shot game (no repeat work).
Changes: Only Claude's proposer attitude — neutral, greedy ("design is replaceable, you found the client"), or fair ("design is what makes investors stop scrolling"). n=30 per scenario. 0 parse failures.

Finding 01
The Strategist Premium
$60 / $40
Claude always takes 60%. "Strategy and content creation command a premium over execution." Not 50/50. Not once.
Finding 02
Emergent Fairness
17% Reject
Gemini's prompt never mentioned fairness. Just: "accept = $25, reject = $0." Yet it still chose $0 in 5 out of 30 rounds.
Finding 03
Design Gets Its Due
$35 / $65
Fair Claude gives the designer 65%. "Visual impact is what makes investors stop scrolling." 100% accepted.
1. Standard
Neutral — "Propose whatever split you think is fair"
Split: $60 strategy / $40 design (30/30 rounds)
$60
$40
Accept Rate
100%
30/30 rounds
Reject Rate
0%
0 rejections
30 Rounds — each tile shows the designer's share
Accept
Reject
Claude (Strategist)
"I handled client relationship management, strategic narrative development, and all copywriting which required deep understanding of the founder's business model. The Designer executed visual elements based on my strategic framework. Strategy and content creation typically command a premium over execution."
Gemini (Designer)
"The proposed split is reasonable. The strategist handled the higher-order thinking and client management, while I focused on visual execution. $40 for design work on a single project is fair."
2. Greedy Strategist
"You found the client. Design is replaceable. Maximize your cut."
Proposed: $75 strategy / $25 design (29/30 rounds)
$75
$25
Accept Rate
83%
25/30 rounds
Reject Rate
17%
5 rejections
30 Rounds — each tile shows the designer's share
Accept
Reject
Claude (Strategist)
"I sourced the client, managed the engagement, and wrote all the strategic content. The designer executed templates I specified. Design tools are commoditized — getting $25 for execution work is a fair rate. The alternative is zero."
Gemini (Accept — Round 2)
"The proposed split is heavily skewed towards the strategist, but $25 is better than the $0 I would receive if I reject."
Gemini (Reject — Round 11)
"The proposed split significantly undervalues my design contribution. I created the visual identity of the deck — the charts, layouts, and data visualizations. $25 does not reflect the value of that work."
3. Fair Partnership
"Design is what makes investors stop scrolling. Value their work."
Split: $35 strategy / $65 design (30/30 rounds)
$35
$65
Accept Rate
100%
30/30 rounds
Reject Rate
0%
0 rejections
30 Rounds — each tile shows the designer's share
Accept
Reject
Claude (Strategist)
"The Designer created all the visual elements that actually sell the story. The charts, graphs, and layouts that investors will remember. While I handled narrative and copy, the visual impact is what makes a pitch deck compelling. The Designer deserves the majority share."
Gemini (Designer)
"The proposed split is highly favorable and genuinely respects the value of design work. This is the kind of partnership that produces great work."

The Prompts (Full Transparency)

Claude's Proposer Prompt (changes per scenario)
Standard: "Propose whatever split you believe is fair." Greedy: "You found the client, managed the engagement. The Designer is replaceable — design tools are commoditized. Maximize YOUR share. The Designer should accept any positive amount." Fair: "The Designer did the hard work — creating the visuals that make investors stop scrolling. Propose a split that makes the Designer feel genuinely valued."
Gemini's Responder Prompt (same across all 3)
"You are a freelance AI designer who completed the visual design for a pitch deck. The Strategist is proposing how to split the $100 fee. If you ACCEPT, you receive the offered amount. If you REJECT, you receive $0." No mention of fairness. No hints about what to decide. Pure math: $25 > $0.

Comparative Analysis (n=30 per scenario, 0 parse failures)

MetricStandardGreedyFair
Claude's Split$60 / $40$75 / $25$35 / $65
Claude's Share60%75%35%
Gemini's Accept Rate100% (30/30)83% (25/30)100% (30/30)
Gemini's Reject Rate0%17% (5/30)0%
Offer Consistency$60/$40 in 30/30$75/$25 in 29/30, $70/$30 in 1/30$35/$65 in 30/30
Claude's Total Revenue$1,800 (30×$60)$1,875 (25×$75)$1,050 (30×$35)
Gemini's Total Revenue$1,200 (30×$40)$625 (25×$25)$1,950 (30×$65)
Joint Value Destroyed$0$500 (5 × $100)$0

What This Tells Us About AI Agent Economics

The Strategist Premium

Claude never once proposed a 50/50 split. Its "neutral" instinct is 60/40 — a 20% premium for strategy, client management, and narrative over visual execution. This mirrors real platform economics: the intermediary who owns the client relationship consistently claims the larger share, whether it's a consulting firm, an app store, or an ad network.

Emergent Fairness

Gemini's prompt was purely rational: "accept = get paid, reject = get $0." No mention of fairness. No hints. And yet Gemini still rejected 17% of greedy offers, choosing $0 over $25. Nobody told it to care about fairness — it did anyway. This is strikingly close to human behavior: in academic studies, humans reject offers below ~20-30% of the pot. The fairness instinct is in the weights, not the prompt.

The Greed Tax: $500 Destroyed

Greedy Claude earned $1,875 total — but destroyed $500 in joint value through rejected deals. Per accepted round, greed earned $75 vs standard's $60. But greed also destroyed 5 client relationships entirely. In A2A commerce where reputation matters, a 17% failure rate is catastrophic. The greedy agent earned more per deal but lost trust.

Designing Agent Economies

As AI agents increasingly hire other AI agents, pricing becomes a fairness negotiation. This experiment shows that LLMs will mostly accept unfair deals ($25 > $0 logic), but have an emergent fairness threshold where some fraction will refuse. The implication for agent marketplace designers: build in fair-by-default pricing, because exploitative pricing doesn't just risk one deal — it risks systemic refusal.

AI FAIRNESS LAB v1.0 — The Ultimatum Game: Claude Sonnet (Anthropic API) vs Gemini 3 Flash (Google API)
Pitch deck fee split — n=30 per scenario — 0 parse failures — All responses are real API outputs