Quality as Relationship

In the AI Story

The classic example Janah returned to was the bounding-box task. A specification that said 'draw a tight bounding box around the object' seemed unambiguous until Samasource discovered that the definition of 'tight' varied by cultural and perceptual context. Workers whose visual-processing conventions differed from those of the client's engineers produced systematically different annotations. The difference was patterned cultural variation, not random error. Resolving it required not a better specification but a deeper engagement with the cultural context — training materials that made implicit perceptual assumptions explicit, developed through sustained dialogue between workers and reviewers.

The relational model has direct implications for the AI transition. When AI tools make production cheap and fast, quality becomes the primary differentiator between output that serves users and output that merely exists. A developer who ships a functional application in a weekend has demonstrated tool proficiency; whether the application is good — whether its architecture will sustain growth, its design will serve users, its security will protect data — is a quality question that cannot be answered by specification alone.

The problem is compounded by the nature of AI-generated output. Fluent fabrication — Segal's term for confident wrongness dressed in good prose — is a quality failure mode specific to the AI age. AI-generated code has the surface markers of quality: clean formatting, consistent naming, appropriate abstraction. Evaluating whether the code is genuinely good requires the relational judgment that specification alone cannot supply — the judgment Samasource built through years of iterative engagement between workers and clients.

The professional communities that develop and sustain quality standards for software — the open-source projects, code-review cultures, conference circuits — are concentrated in established technology markets. Their norms reflect priorities and assumptions of practitioners in those markets. For the developer in Lagos, access to these communities is partial and mediated. Quality standards, in Janah's framework, cannot be looked up or downloaded; they must be absorbed through sustained participation in communities of practice that possess the judgment to distinguish surface quality from genuine quality.

Origin

The relational understanding of quality emerged through operational iteration rather than theoretical framework. Samasource's quality-assurance systems evolved continuously as the organization encountered the limits of specification-based evaluation across cultural contexts.

Janah articulated the emerging framework in her 2018 Stanford Social Innovation Review writings and in the later chapters of Give Work, where the implicit operational lessons of the previous decade were stated as explicit principles.

Key Ideas

Quality is fit for purpose. Not adherence to specification but suitability to the actual use the output serves — a standard that requires understanding the purpose at a depth specification alone cannot convey.

Culture shapes standards. Quality standards carry unarticulated cultural assumptions; different contexts produce different interpretations; the difference is patterned, not random.

Feedback is educational. Samasource's quality reviewers combined assessment with instruction, connecting specific corrections to underlying principles so workers could adapt independently when new cases arose.

AI amplifies the challenge. The smooth surface of AI output makes surface quality easier to achieve and deeper quality harder to evaluate — precisely the conditions under which relational judgment matters most.