CHAPTER 02 / STRUCTURAL ABSENCE AND EPISTEMIC DISTORTION
Herstory as Analytic Lens
“Something shifts in a room when everyone prepares with the same tool. The meetings get smoother. More efficient. The smoothness is the harm.”
Kira Sjöberg Helsinki, Finland
What I Want Readers to Notice
Over eighteen months, I watched more than 2,200 professionals across five nontechnical industries sit down with large language models and compare what the systems produced with what they already knew.
Again and again, something subtle happened.
When an output felt wrong, people did not always evaluate it. They recognized its shape. It sounded plausible, carried the tone of something pre-vetted, and resembled the language they expected to see. The output moved into reports, presentations, communications, and decisions, carrying whatever the system had flattened or displaced along with it.
One moment stayed with me. A logistics coordinator was reviewing an AI-generated route plan that looked entirely reasonable. Then she stopped and said, “This works on paper, but not on a Tuesday in November.”
The system had the routing data. What it did not know how to value was eleven years of situated judgment about a particular place, season, workflow, and set of conditions. That knowledge was not absent. It was present in the room and treated as less authoritative than the fluent output on the screen.
That distinction changed the question for me.
AI governance often assumes that exclusion is primarily a missing-data problem: find what is absent, expand the archive, add more representative material. Sometimes that is exactly what is needed. But representation alone cannot correct a system whose evaluation criteria, alignment objectives, terminology, or institutional workflows have learned to make certain knowledge count less.
I wrote this chapter to make that mechanism visible. Herstory is not used here as a corrective appendix that adds women back into an existing account. It is a detection practice: a way to ask whose standpoint has been treated as neutral, whose knowledge is being discounted, and who has the authority to reopen the standards that produced the distortion.
The resources on this page help teams move from saying, “The AI was biased,” to naming the actual institutional mechanism: This criterion, objective, workflow, or term discounted this knowledge, and these people are responsible for repairing it.
Kira Sjöberg | Helsinki, Finland
Kira is a speaker and advisor specializing in the human and organizational dimensions of artificial intelligence adoption. Her advisory work examines how automation bias, over-trust, consensus culture, and responsibility diffusion shape decisions in AI-enabled environments.
FEATURED RESOURCE
The Mechanism Review Toolkit
A relational impact review for youth AI
The Mechanism Review Toolkit helps teams diagnose how an AI-mediated decision may be reproducing exclusion before they choose the wrong repair. It begins with a distinction that conventional bias reviews often overlook: Absence and Distortion. The response requires reopening evaluation criteria, alignment objectives, terminology, workflow design, or decision authority. Many consequential failures involve both of these concepts. The purpose of the review is not to select the more sophisticated diagnosis. It is to ensure that the repair matches the mechanism rather than merely treating the visible symptom.
FOR
AI-governance teams, public agencies, data and analytics leaders, product groups, organizational-development practitioners, policy teams, educators, compliance functions, community representatives, and cross-functional decision makers
TIME
A 30-minute initial audit or a complete 30-day institutional review
FORMAT
Downloadable audit, Use-Case Charter, review guide, terminology prompts, and Deployment Decision Log
Designed to be tested and adapted for your setting—not followed as a fixed prescription. The toolkit does not replace legal review, regulatory obligations, model documentation, or technical evaluation. It adds a governance layer those processes often lack: mechanism diagnosis, situated evidence, named authority, and institutional accountability.
Why This Matters
NOTICE
Recognize when polished outputs and smooth agreement are concealing an unresolved knowledge problem.
Notice whose judgment is treated as anecdotal, when “sounding right” substitutes for evaluation, when everyone arrives with the same framing, and when present knowledge is repeatedly flattened or made less credible.
DECIDE
Distinguish between what is missing and what has been distorted before prescribing a repair.
Determine whether the problem requires data action, changes to evaluation or alignment, a terminology correction, a workflow redesign, or some combination—and decide whether the system should proceed, pause, be redesigned, or be refused.
SUSTAIN
Make the diagnosis answerable over time.
Name the people responsible for repair, establish authority to halt or revise the system, preserve affected-stakeholder participation, document the decision, and return after 90 days to assess whether the same knowledge is still being discounted.
This work moves beyond abstract AI ethics to confront the deeper questions of power by drawing attention to whose knowledge counts and whose voices shape our technological futures.”
PAYAL ARORA
Author of From Pessimism to Promise: Lessons from the Global South on Designing Inclusive Tech
“
ABOUT THIS CHAPTER
AI bias is commonly described as a problem of incomplete data. The archive excluded certain people, histories, languages, or perspectives, and the answer is to make the dataset more representative.
That account is important, but incomplete.
Herstory as Analytic Lens shows how AI systems can reproduce inequality even when marginalized knowledge is technically present. Evaluation regimes may reward dominant framings. Alignment processes may smooth contested knowledge into agreeable generalities. Organizational workflows may treat fluent output as pre-vetted. Institutional terminology may shift accountability from human design decisions to the system's supposed inner life.
The chapter calls this epistemic distortion.
Distortion explains why adding more data does not necessarily change what an institution treats as correct, credible, or actionable. An organization can expand representation while preserving the standards that made the knowledge peripheral in the first place.
Drawing on feminist standpoint epistemology and epistemic injustice theory, the chapter reframes situated knowledge as a capacity for detection. People navigating systems not built for them are often required to notice contradictions, hidden rules, and local conditions that structurally advantaged actors can afford to overlook.
The governance task is not to romanticize every dissenting observation. It is to create a process through which those observations can be tested without being dismissed before the evaluation begins.
The Structural Absence Audit and Mechanism Review make that process operational by linking diagnosis to evidence, ownership, decision authority, repair, and review.
SUGGESTED USES
AI-governance, model-evaluation, and pre-launch reviews.
Public-sector procurement, vendor evaluation, and contract review.
Reviews of automated decisions in employment, public services, health, education, and risk assessment.
Leadership and cross-functional organizational decision-making.
Executive education and professional learning.
Courses in data ethics, information policy, feminist technology studies, public policy, and applied AI ethics.