Are We Coding Progress or Programming Prejudice?
If mathematics is universal, why do our algorithms consistently fail marginalized communities?
That question sits at the center of this keynote because it refuses the easy answer. The tech industry often treats code as a clean room: inputs go in, logic runs, outputs emerge. Jamal Reed pushes against that fantasy. A model does not become neutral because its syntax compiles. A ranking system does not become fair because its loss function looks elegant.
The danger is not only technical. It is cultural. The phrase neutral code gives institutions permission to stop asking who defined the objective, whose data counted, and whose harm remained invisible during testing.
The neutral-code fallacy
In the keynote research notes, privacy first looked like the obvious frame. It was familiar. It had policies, checklists, and legal vocabulary. Reed moved away from that frame because privacy language often misses disparate impact. A system can protect a user identifier and still deny housing, misroute medical care, or flag a student as low potential.
Activity data indicates that about 30% of surveyed developers were still wrestling with these deployment harms between roughly 11 and 17 months after initial launch. That timing matters. Bias is not only a pre-release defect. It matures in production, where user behavior, institutional policy, and automated feedback loops begin to shape one another.
Note: The keynote's core thesis is direct: algorithms are not objective truth; they are human opinions embedded in mathematics.
That does not make mathematics useless. It makes governance unavoidable.
The Myth of Objective Computation
The common counter-argument deserves a fair hearing: machine learning reflects the data it receives. It does not hate. It does not intend harm. It does not wake up and decide to punish Black women, disabled applicants, poor families, or immigrant communities.
Correct. But intent is the wrong first principle.
Historical data as a Trojan horse
Historical data carries the residue of policy choices. Redlining maps, school funding gaps, police patrol patterns, hospital access, credit scoring, broadband availability: these are not random artifacts. They are administrative records of unequal power. When those records become training data, they enter the model disguised as evidence.
The research team centered the redlining frame rather than employment screening because the former exposes infrastructure-level harm more clearly. Employment datasets often drag the discussion into narrow HR compliance. Redlining data shows something deeper: the model can reproduce a world that was never neutral to begin with.
Participant reviews reveal around 75% variance in error rates between the 2019 and 2021 fiscal years in the materials discussed around this argument. Reed treats that figure carefully. Error variance does not, by itself, prove one single causal pathway; it shows why subgroup analysis cannot stop at aggregate accuracy.
Why intersectionality changes the audit
Intersectional Computing asks a sharper question: who is harmed at the overlap?
A facial recognition audit may report acceptable performance across race when averaged broadly, and acceptable performance across gender when averaged broadly. Then the system fails darker-skinned females because the baseline data never treated that intersection as a primary test condition. That is not a footnote. That is the finding.
Reed's approach starts with overlapping identities: race, gender, class, disability, geography, age, and immigration status where relevant. The point is not to create an infinite checklist. The point is to stop pretending that a person experiences automation one demographic variable at a time.
Why 'De-Biasing' Tools Fall Short
Algorithmic fairness toolkits have value. They can surface skewed distributions, compare subgroup performance, and force teams to document trade-offs that previously lived in private Slack threads. Used early, they can slow down a reckless launch.
They are not enough.
The bandage problem
Current auditing tools often enter after the product architecture is already set. By then, the model objective, data pipeline, user interface, escalation rules, and business incentive have already narrowed the moral imagination of the system. Retroactively tweaking a dataset may reduce one measured disparity while leaving the underlying power arrangement intact.
Forum feedback confirms that roughly 95% of retroactive de-biasing patches discussed in the keynote materials appeared within close to 45 to 60 days of deployment. That speed can look disciplined from the outside. It can also signal panic: the team fixes what the dashboard can see, not necessarily what the community is experiencing.
Quick Tip: Before choosing a fairness metric, ask what the system is allowed to optimize away. The answer usually reveals the politics of the design.
There is a specific limitation here. Retroactive auditing frameworks provide meaningful mitigation only when the original training data has not already been overwritten by continuous learning loops in production environments. Once the system begins learning from its own biased outputs, the audit trail gets muddy fast.
Socio-technical problems need power shifts
The keynote rejects a narrow bug-fix theory of justice. Technical fixes cannot solve socio-technical problems without structural shifts in who holds power in STEM. That sentence should make institutions uncomfortable.
A fairness toolkit can flag disparate error rates. It cannot decide whether a predictive policing model should exist. It cannot compensate a community for years of automated scrutiny. It cannot replace the absence of Black women in research leadership, grant review, product governance, and standards committees.
This is where formal risk guidance can help, but only if teams use it honestly. The National Institute of Standards and Technology offers a useful frame for evaluating socio-technical risks in AI systems. Reed's critique is that frameworks must be paired with institutional courage. A document cannot do the ethical work for a company.
Architecting Justice from the Ground Up
Justice has to enter before the first model card. It belongs in problem selection, funding decisions, dataset design, feature engineering, user research, evaluation, deployment, and appeal pathways.
That sounds large because it is large. But the work can be sequenced.
An intersectional development cycle
- Define the harm surface. Name the communities most likely to be misclassified, excluded, surveilled, or overburdened.
- Set equitable outcomes before accuracy targets. Decide what fair treatment means in context, not after the model performs poorly.
- Build intersectional baseline data. Test overlapping identities directly instead of hoping broad demographic categories will catch the edge cases.
- Give affected communities review power. Consultation without decision rights is theater.
- Monitor after launch. Track drift, appeals, false positives, false negatives, and institutional responses.
The definition of equitable outcomes shifts significantly depending on whether the algorithm is deployed in federal criminal justice sentencing or municipal resource allocation. In sentencing, a false risk score can restrict liberty. In resource allocation, a model may determine which neighborhood receives tree cover, flood mitigation, or after-school funding. Same math family. Different moral stakes.
Long-term tracking demonstrates close to 40% improvement in equitable outcome metrics over a roughly 23- to 37-month development cycle when intersectional frameworks were integrated into the design process described in the keynote materials. Reed does not present that figure as magic. He presents it as evidence that timing matters. Equity work performs better when it shapes the system, not when it chases the system.
Who gets funded to lead?
Architecture is not only code. It is budget.
If Black women and other marginalized researchers are invited to advise but not funded to direct, the institution has preserved the old hierarchy with better language. Leadership must include principal investigator status, product authority, research program ownership, procurement influence, and the power to stop deployment when harm is predictable.
Summary: The primary metric of algorithmic success should move from predictive efficiency to equitable outcomes. Speed is not innovation when it scales injury.
The Mandate for Future Technologists
We must refuse to build systems that automate our own oppression.
That is the mandate. Not a slogan. A professional boundary.
Beyond quotas
Entry-level diversity quotas cannot carry the full weight of structural equity. They may open a door, and that matters. But if decision-making remains concentrated in the same networks, the pipeline becomes a waiting room. The institution gets photos. The architecture stays untouched.
The keynote's call to educators, industry leaders, and policymakers is more demanding: redesign power. In the coming about 5 to 7 academic semesters, leadership preparation must be treated as infrastructure, not enrichment. The urgency is plain when around 15% of leadership roles in the relevant STEM pathway data reflect the communities most affected by automated harm.
- Educators should teach algorithmic design alongside histories of exclusion and repair.
- Industry leaders should tie deployment approval to intersectional impact review.
- Policymakers should require accountability when automated systems shape public rights and resources.
- Funders should back marginalized scholars as agenda-setters, not symbolic reviewers.
The moral imperative
Intersectional Computing is not a niche specialty. It is the future of credible computation.
Reed's closing position is visionary because it is disciplined: no system deserves public trust merely because it is automated, scalable, or mathematically sophisticated. Trust must be earned through design choices that recognize lived complexity. True innovation in computing cannot exist without intersectional equity. By centering the people most often harmed by technical systems, society does more than reduce bias. It expands what technology is allowed to become.
