🤖 AI

Whose Good, Whose Place? The Moral Geography of Agentic AI for Social Good

This paper analyzes 112 studies on agentic AI for social good to reveal a critical "moral-geographic asymmetry" where researchers frequently omit geographic context—particularly for institutional goals like SDG 16—and rarely report real-world deployments, thereby highlighting significant accountability gaps and proposing new reporting standards for more context-specific and participatory AI systems.

Original authors: Poli Nemkova, Haeshitha Indukuri, Jaedon Charles

Published 2026-05-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Poli Nemkova, Haeshitha Indukuri, Jaedon Charles

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a group of architects designing "smart robots" to help humanity solve its biggest problems, like hunger, climate change, or unfair laws. They call these robots "Agentic AI" because the robots can plan, make decisions, and act on their own, not just follow simple commands.

This paper is like a massive inspection of 112 blueprints (research papers) for these robots. The authors, who are researchers themselves, looked at these blueprints to answer a simple but crucial question: Who are these robots actually supposed to help, and where?

Here is the story of what they found, told in everyday terms.

1. The "Global" vs. The "Local" Problem

The researchers noticed a strange pattern in how these blueprints describe the robots' jobs. They call this a "Moral-Geographic Asymmetry." That's a fancy way of saying: The robots' jobs are described very differently depending on what kind of problem they are solving.

The "Concrete" Jobs: When the robots are designed to fix physical or health problems (like delivering medicine, cleaning water, or tracking climate change), the blueprints usually say exactly where they will work. They might say, "This robot will help farmers in Kenya" or "This system will monitor floods in Florida." It's like a contractor saying, "I'm building a house on 4th Street."
The "Abstract" Jobs: When the robots are designed to fix social or political problems (like improving peace, fixing legal systems, or reducing inequality), the blueprints almost never say where. They often just say, "This robot will help the world" or "This will fix global governance." It's like a contractor saying, "I'm going to build a house... somewhere... maybe everywhere."

The Metaphor: Imagine you are hiring a chef.

If the chef is making soup (a physical task), they tell you exactly which kitchen they are using and what local ingredients they have.
If the chef is making laws (a social task), they just say, "I will feed everyone in the universe," without naming a single city, country, or community.

The paper argues this is dangerous. You can't build a fair legal system or a peace treaty without knowing the specific culture, laws, and people involved. By saying "everywhere," the researchers are actually saying "nowhere specific," which makes it impossible to know if the robot will actually work or if it might cause harm.

2. The "Paper Tiger" Problem

The second big finding is about how many of these robots actually exist in the real world.

The Reality Check: Out of 112 blueprints, only 25% (about 1 in 4) showed any evidence that the robot was ever tested with real people or used in a real situation.
The Simulation Trap: The other 75% were just concepts (ideas on paper) or simulations (robots playing in a video game world).

The Metaphor: Imagine a car company that releases 100 new car designs.

75 of them are just beautiful drawings and computer animations. They look great on a screen.
Only 25 have ever been driven on a real road.
The company keeps saying, "Our cars will save the world!" but they haven't actually driven any of them yet.

The paper points out that this isn't just a problem for one type of robot or one year; it's a problem for the entire field. Whether the robot is new or old, simple or complex, most of them are still just ideas.

3. The Five Missing Pieces (The Accountability Gaps)

Because of the two problems above, the authors say the field is missing five key ingredients to be truly responsible:

Missing Targets: We don't know who the robots are for. If you don't name the community, you can't ask them, "Is this actually good for you?"
Missing Maps: Without a specific location, we don't know if the robot fits the local laws or culture.
Missing Proof: We have too many "what ifs" and not enough "it works."
Missing Rules: We don't know who is in charge. If the robot makes a mistake, who is responsible? The paper notes that even the researchers can't agree on how "autonomous" (in charge of themselves) these robots really are.
Missing Mistakes: The papers mostly talk about successes. They rarely talk about when the robots failed or when they shouldn't have been used. It's like a pilot only writing about perfect landings and never mentioning the near-crashes.

The Bottom Line

The paper isn't saying these robots are bad ideas. It's saying that the way we are talking about them is too vague.

When we talk about fixing the weather or growing food, we are specific. But when we talk about fixing society, justice, or peace, we get fuzzy and say "everywhere." The authors argue that you cannot fix a specific place's problems without naming that place.

They propose a simple new rule for anyone building these robots: Stop saying "for the world." Start saying "for [Specific City/Community] with [Specific Goal]." If you can't name the place and the people, you can't claim you are doing "social good." You are just writing a story.

Technical Summary: "Whose Good, Whose Place? The Moral Geography of Agentic AI for Social Good"

Problem Statement
The paper addresses a critical gap in the emerging literature on agentic AI for social good. While researchers increasingly propose autonomous systems (planning, coordinating, allocating resources, persuading) to advance the United Nations Sustainable Development Goals (SDGs), the invocation of "social good" does not inherently establish accountability to the specific communities these systems claim to serve. The authors argue that the field suffers from contextual underspecification, particularly regarding geographic focus. This lack of specificity is not uniform; it is concentrated in domains where local political, legal, and cultural contexts are most critical for feasibility and legitimacy. Furthermore, there is a significant disconnect between conceptual proposals and real-world implementation, with the majority of literature remaining at the level of simulation or theory without empirical validation in deployed settings.

Methodology
The authors conducted a structured survey of 112 papers published between 2015 and 2026 (91% from 2020 onward). The corpus was constructed using a PRISMA-style systematic review across four academic platforms (Google Scholar, ACL Anthology, Semantic Scholar, Elicit), utilizing queries organized around all 17 SDGs to minimize retrieval bias.

Annotation Procedure: To manage the scale of annotation across eight dimensions, the authors employed a dual-LLM annotation procedure using GPT-4o and Llama-3.3-70B with identical controlled-vocabulary prompts.
Human Validation: A stratified subsample of 50 papers was manually coded by a human annotator to validate the LLM outputs. Consensus labels were constructed by prioritizing human labels for the subsample and selecting the LLM with higher human-LLM agreement (Cohen's $\kappa$ ) for the remainder.
Coding Dimensions: Each paper was coded for:
1. Domain
2. SDG Alignment (1–17)
3. Agent Type (e.g., LLM agents, multi-agent systems)
4. Autonomy Level
5. Real-World Deployment Status
6. Evaluation Type
7. Agent Goal
8. Geographic Focus (coded as "Not specified," "Global/Broad," or "Concrete")
Statistical Analysis: The study utilized descriptive statistics and inferential tests (Chi-square tests of independence, Fisher's exact tests) to examine relationships between SDG clusters and geographic specification. Sensitivity analyses were performed using alternative SDG cluster assignments and strict vs. broad geographic coding definitions.

Key Contributions

Structured Mapping: The paper provides the first systematic taxonomy of agentic AI for social good, mapping the field across domains, architectures, autonomy levels, and evaluation types.
Identification of Moral-Geographic Asymmetry: The authors identify a non-uniform pattern of contextual abstraction. While health and ecological domains often specify a geographic context, institutional and social-policy domains (specifically those aligned with SDG 16: Peace, Justice, and Strong Institutions) overwhelmingly fail to do so.
Accountability Agenda: The paper proposes a minimal reporting standard and identifies five specific accountability gaps that limit the maturity of the field, arguing for a shift from "possibility" to "accountable sociotechnical intervention."

Results

Geographic Underspecification: 73% (82/112) of the papers specify no geographic context. This absence is sharply uneven.
- Papers aligned with health or physical/ecological SDGs specify a geography 37–40% of the time.
- Papers aligned with institutional and social-policy SDGs specify a geography only 13% of the time (broad coding) and 0% under strict coding (which excludes "global" claims).
- SDG 16 is the most-covered goal in the corpus (39 papers) but has the lowest geographic specification rate (13% broad; 2.6% strict).
- Statistical analysis confirms this asymmetry is significant ( $\chi^2(2) = 8.78, p = .012$ ) and robust across seven of eight sensitivity configurations.
Proposal-to-Practice Gap: Only 25% (28/112) of the papers report any real-world deployment or small-scale test. The remaining 75% are conceptual or simulation-only.
- This gap is field-wide, not localized to specific architectures or eras. Papers naming a geography do not deploy at higher rates than those that do not. Post-2023 LLM-agent papers do not deploy at lower rates than earlier multi-agent work.
Evaluation Deficits: Evaluation practices mirror the deployment gap. 36 papers are conceptual studies, and 30 are simulations. Only 5 report field studies.
Autonomy Ambiguity: Cross-LLM agreement on "Autonomy Level" was low ( $\kappa = 0.56$ ), indicating the field lacks precise definitions for the boundary between system-initiated and human-authorized action.

Significance and Claims
The paper claims that the field's current trajectory treats institutional good as universal in a way it does not treat health or ecological good. This "moral abstraction" is ethically problematic because institutional interventions (governance, justice, policy) are highly dependent on local political and legal contexts. By failing to name the communities or places these systems serve, the literature obscures power dynamics and accountability.

The authors argue that the field must move beyond asking "what agents can do" to asking "what agentic AI for social good owes to the people and places in whose name it is built." They propose that future work must:

Specify target contexts (communities, institutions, settings).
Match evidence to claims (e.g., field studies for claims of community benefit).
Engage affected stakeholders early in the design process.
Justify architectural complexity and autonomy levels explicitly.
Document failures and non-deployments to enable collective learning.

The paper concludes that expansion in the number of proposals does not equate to maturity; true maturity requires context-specific, participatory, and accountable systems.

1. The "Global" vs. The "Local" Problem

2. The "Paper Tiger" Problem

3. The Five Missing Pieces (The Accountability Gaps)

The Bottom Line

Technical Summary: "Whose Good, Whose Place? The Moral Geography of Agentic AI for Social Good"

More like this