ATLAS: Navigating Culture as Infrastructure for Designing and Evaluating AI

Culture as Infrastructure for AI

AI systems that perform well in one setting often fail to carry over to another, because the knowledge, relationships, and practices that make AI output usable are rarely captured by the technical system itself. This talk presents ATLAS, a playbook grounded in ethnographic and participatory research across agriculture, small business, health, and workplace settings, aimed at helping teams surface cultural assumptions across the AI life cycle: design, development, deployment, and evaluation.

Reframing Culture for AI Design

The core argument is that culture functions as infrastructure for AI interactions. A person expresses a need, the AI generates a response, and the person must then interpret that response and act on it in the physical world. At each step, culture shapes the exchange: shared language determines how a problem is expressed (English is the norm, but many people work in Swahili, Portuguese, and other lower-resource languages), knowledge and relationships determine how the response is judged, and local practices and responsibilities determine whether the response can be acted upon.

Drawing on social science theories of culture, including multi-level models, cultural dimensions such as individualism and power distance, and practice-based views, the talk emphasizes that culture is complex, dynamic, and situated. Rather than reducing it to a fixed representation, ATLAS treats culture along three interlocking axes:

• Contextual: the domain, the task, the modality, and the setting in which the AI is used.
• Relational: how people connect with each other and with the system, and how the AI disrupts or becomes part of existing knowledge practices.
• Consequential and dynamic: how information shapes actions and outcomes, and how those consequences shift as relationships and conditions evolve.

From this, the playbook asks which evidence matters, whose knowledge is needed, and which decisions could change over time.

Project Gecko: Farmers and AI in Kenya and India

Project Gecko studied how farmers in Kenya and India, working with partner Digital Green, used an AI advisory system between November 2024 and June 2025. The methods combined interviews, farm visits, onboarding observation, prototype workshops, and co-design across different regional communities, each with its own languages, resources, seasons, and goals.

Three findings connected the work across sites:

1. Knowledge practices and relationships make advice usable. AI entered an existing information ecology of peers, extension officers, and digital tools. Farmers had to translate and reinterpret AI output for their communities before acting on it. Advice that is on topic can still fail if it ignores local conditions, resources, timing, and responsibilities. In one example, a farmer asking about the cost of raising pigs received a response covering housing, breeding stock, and cost factors, but no usable pricing that would support a decision.

2. Context and intent need to be made visible. The system could identify the topic of a question but could not supply the practical detail needed for action. Farmers, working through questions together in group sessions, naturally added specifics such as the animal’s weight and the quantity of feed needed. This pointed toward designing AI that asks for clarification, invites refinement after a question is asked, presents sources and limitations, and hands off appropriately to the human.

3. Evaluation is social and personal. Accuracy alone does not capture quality when truth is defined locally and community wisdom can outweigh institutional general knowledge. Farmers trusted other farmers above nearly any other source, and often took AI responses back to group chats for a sense check. Completeness was subjective: the pig answer was accurate but did not meet what the specific farmer needed. Usefulness was also personal, with some farmers wanting fewer words, others wanting diagrams, and all needing tailoring to their context. This led the team toward human-centered evaluation, where AI quality is judged against user needs, goals, and values.

When AI responses contradicted local knowledge or the experience of trusted people, farmers tended to trust themselves and stop using the system. Where responses were more ambiguous, farmers consulted others before acting. Usage followed a pattern of initial excitement and exploration, a sharp drop-off after the system’s limits became clear, and renewed engagement tied to seasons such as planting and harvest. Newer farmers engaged more than experienced ones, who used the system mainly for knowledge checks. Farmers also approached the tool with an entrepreneurial mindset, hoping it could help increase revenue, which the system as built could not fully support.

Applying ATLAS in Practice

ATLAS organizes its questions around four themes: context (what must be understood about a task or setting before designing for it), relationships (whose knowledge and which relationships shape use, relevant to adoption and market fit), consequences (what decisions and outcomes follow, and how the system should adapt), and revisiting assumptions over time. These questions are mapped to decision points across the life cycle, whether a team is evaluating an existing system, adapting a component, or co-creating from scratch. The playbook includes templates that help teams articulate what a project assumes, what affects a decision, whether a question is a design, development, deployment, or evaluation problem, and how assumptions link explicitly to decisions.

In a worked example, a team redesigning an AI advisory tool that assumed users knew what to ask could run a local inquiry and discover that farmers formulate questions together in groups. That finding could expand the user profile to include family members, allow multiple users to add context within a conversation, and surface credibility signals such as community-endorsed guidance.

Live Evaluation and Feedback Loops

The team’s own partners iterated on the app several times, incorporating insights about missing knowledge into the knowledge base. A persistent open question was how to capture live evaluation. Observed interaction signals, such as users rephrasing a question with one or two changed words or ending a session after an unsatisfying answer, offered a way to detect unmet needs that post hoc qualitative feedback missed. Feedback mechanisms like thumbs up or down are common in current systems but do not reveal intent. The speakers argued that future systems should personalize in the moment and update their knowledge bases from field feedback.

Open Questions

Several research directions remain open. The first is intent alignment: people hold different definitions of terms such as fairness or what counts as a good farming outcome, and common ground emerges through conversation. Aligning system vocabulary with the goals farmers actually use is ongoing work. The second is interaction modality. Beyond text, farmers may want to interact through speech in local languages, which the team explored for Kikuyu and Swahili and then extended to four other Kenyan languages, or through images and video. The third concerns conversational quality: AI often agrees with users even when they are wrong, which does not help people reach their goals. The fourth is community-centered design. Many AI tools assume a single individual user, but farmers often work through cooperatives and women’s groups and want to discuss AI output collectively. The speakers argued that AI adoption across the continent will require redesigning interaction models rather than borrowing the individual-user model from Western contexts.

• Tags:
• Low-Resource Languages
• Culture-Aware AI
• Human-Centered Evaluation
• Participatory Design
• Ethnographic Research
• Agricultural AI
• Speech Interfaces
• Community-Centered AI