Article 3 → Conscious Stewardship
In the last piece, I made a promise.
I said I’d explain what makes AI categorically different from every other technology philanthropy has absorbed. Not faster. Not more powerful — categorically different. The kind of difference that changes what governance needs to do, not just how much of it we need.
That’s where this piece begins. And it ends somewhere more uncomfortable: with the question of whether, knowing what we know, we choose to govern — or whether we let the system run.
What Makes This Different
Every wave of automation in our lifetimes has done roughly the same thing: it executed. Machines performed physical tasks. Software stored decisions and reproduced them. Databases held memory so humans didn’t have to. Even the most sophisticated enterprise systems were, at their core, very fast filing cabinets.
Generative AI does something else.
It doesn’t just store or retrieve. It summarizes. Synthesizes across sources. Analyzes sentiment. Identifies patterns. Recommends next steps. Ranks options. Interprets data — presenting conclusions rather than raw information.
In philanthropy, this means something specific:
When an AI tool summarizes fifty grantee reports into a board-ready dashboard, the board isn’t reading the grantees anymore. It’s reading the model’s interpretation of the grantees.
When a risk-scoring tool flags an application as “high capacity risk,” the program officer isn’t evaluating the nonprofit — they’re evaluating an output.
The technical term for this is interpretive mediation. The plain version is simpler:
AI shapes how institutions decide what’s true.
That is not an efficiency question. It is a governance question.
When AI-generated summaries become what boards read, and risk scores become inputs into decision-making, interpretation is no longer solely carried by named staff. It is partially embedded in systems trained on historical patterns and probabilistic inference.
We have always trusted program officers to interpret data. The difference is that program officers are accountable, contestable, and context-aware. Their reasoning can be surfaced and debated. When interpretive work is embedded in opaque or semi-opaque systems, the locus of accountability shifts.
That shift requires a different level of oversight — not because interpretation is new, but because its structure is.
A Drive, a Lucky Dip, and a Cautionary Tale
I was on a road trip with the kiddos. I was searching Libby for an audiobook — chosen with the intellectual rigor of a coin flip. Walter Tevis. Mockingbird. Sure.
The book begins in New York City. A walk toward the Empire State Building. That detail hooked them immediately.
In Tevis’s world, humans have outsourced so much cognitive work to machines that they have quietly stopped reading. Stopped writing. Stopped, in many cases, thinking in any sustained way. Not because they were forced to. Because it was easier not to. The machines handled it. And one day, the machines are the only ones who remember how.
What strikes me about Mockingbird is that it isn’t dramatic. There’s no uprising. No singular catastrophe.
The tragedy is quieter.
It’s atrophy.
The humans didn’t lose their capacity to think because machines took it from them. They lost it because they stopped practicing. Each small abdication felt survivable. Each convenience felt reasonable.
Until, collectively, they had drifted somewhere none of them had chosen.
That is the risk I want to name.
Not collapse. Drift.
How Atrophy Actually Happens
MIT’s Work of the Future research identifies three failure modes in automation: disuse, misuse, and overuse.
Disuse: ignoring tools that would help. Misuse: deploying tools poorly. Overuse: tools that work well but create dependency or skill atrophy over time.
Overuse is the most insidious — precisely because it doesn’t feel like failure while it’s happening.
When automation works reliably, people stop practicing the underlying skill. Not out of laziness. Out of rational adaptation. Why maintain fluency in something the tool handles?
The risk only becomes visible later — when the tool fails, changes, or produces flawed output, and no one in the room can evaluate it critically because no one has done the underlying analysis themselves in two years.
In grantmaking and compliance, this risk shows up in subtle ways:
Decisions that are harder to explain because model recommendations become embedded in the workflow. Outputs accepted without interrogation because internal expertise is stretched thin. Institutional memory shifting into systems that few people fully understand.
None of this is dystopian.
It is a governance drift — one that strong institutions can prevent, but only if they name it early.
The Architecture of Small Decisions
Ursula Hegi’s Stones from the River isn’t about technology. It’s about how communities adjust — slowly and reasonably — to authoritarian pressure. Each accommodation feels survivable in isolation. The full architecture of what has changed is only visible in retrospect.
A small silence. A minor accommodation. A rule that seems temporary.
Each decision defensible on its own terms.
Until the community looks up and finds itself somewhere no single decision led.
AI adoption in philanthropy happens this way:
A contract signed without data transparency clauses because the timeline was tight. A bias audit postponed because deployment was already underway. A summary tool replacing full proposal review “just this cycle” because capacity was stretched. Program attendance data fed into third-party tools without clear consent protocols because legal reviewed the contract — and legal isn’t a community organizer.
None of these is catastrophic. All are defensible.
Together, they normalize drift.
And normalization is how sector atrophy begins.
When Commercial Logic Becomes Default Logic
The AI ecosystem philanthropic institutions are adopting was built for return on investment, speed, scale, data capture, and user lock-in. This is not a moral critique. It is incentive design.
Kate Crawford argues in Atlas of AI that technological systems inherit and amplify the power structures of the environments in which they are built. They do not arrive neutral. They arrive with assumptions embedded — about what counts as data, what counts as efficiency, what counts as success.
If philanthropic institutions adopt AI architectures without actively shaping governance, commercial logic becomes default logic.
Efficiency becomes the design principle — not because we chose it, but because we didn’t choose anything else.
Default logic is still logic.
Consider what this looks like in practice: a model trained on historical grantmaking patterns defines “strong applicants” based on what has already been funded. It optimizes for legibility to funders rather than proximity to community need. It rewards organizations fluent in institutional language because that’s what the data reflects.
That’s not a malfunction.
That’s the system performing exactly as designed — in a context it was never designed for.
If efficiency becomes the design principle, we lose the human — the critical thinking, the care, the community. That is not why our sector exists.
We Built Stoplights
I am not arguing for abstinence.
I know exactly how dangerous cars are. I tell my children they will squash you like a blueberry. I mean it. And I still let them walk home from school.
Why?
Because we didn’t ban cars and we didn’t pretend they were safe. We governed them. We built stoplights and crosswalks and licensing requirements and insurance mandates. We created shared norms. None of it is perfect. All of it makes coexistence livable.
AI is similar.
We cannot meaningfully opt out — not if we want to remain relevant to the communities we serve. The question has never been whether to coexist with AI.
The question is whether we build the stoplights.
In this context, stoplights look like:
Transparency requirements before any contract is signed or meeting transcribed. Bias audits built into deployment timelines, not added after the fact. Clear protocols for what happens when a model surfaces uncomfortable findings. Defined decision rights specifying what AI can recommend and what humans must decide. Governance structures that keep communities — not just compliance teams — at the center of review.
We do not get to opt out of coexistence.
We do get to shape the terms.
The Communities at the Center
Governance is not only about protecting institutions. It is about protecting people.
The risks of ungoverned AI adoption are unevenly distributed. Data extraction from vulnerable communities is already a reality in adjacent sectors. Surveillance creep — where program data collected for one purpose is used for another — is a known pattern.
When AI systems determine which organizations receive funding, they make implicit judgments about whose voice is legible. Organizations that write in the language of institutional philanthropy will score well. Organizations that center community voice or relational outcomes may not — not because their work is weaker, but because they weren’t in the training data.
Philanthropy’s obligation is clear:
Develop best practices. Protect communities from harm. Fund public-interest oversight. Support safe experimentation with communities in the room as co-designers — not subjects.
That is not fear.
That is stewardship.
Awareness Creates Obligation
In the first piece in this series, I wrote about genetic testing during pregnancy — about realizing that certain information would create an obligation I wasn’t prepared to carry.
This is the opposite of that moment.
We can see the interpretive mediation. We can see the acceleration. We can see the drift. We can see how default logic takes hold.
Once you see it, you inherit responsibility for what you do next.
Abdication is a choice. Stewardship is a choice.
Both are available.
Only one aligns with why our sector exists.
The Tragedy Wasn’t the Machines
My kids and I finished Mockingbird somewhere on the highway, in that particular silence good stories leave behind.
The tragedy was never the machines.
The tragedy was surrender — gradual, reasonable-seeming, almost invisible until complete.
We understand the incentives. We see the acceleration. We know what atrophy looks like while it’s still early.
So now the question is not technical.
It is moral.
Will we build the stoplights?
Stewardship is not optional in moments of acceleration.
It is the work.
In the next piece, I’ll move from principle to practice: what governance infrastructure actually looks like inside philanthropic operations, and what it means to design AI systems that are accountable not just to boards — but to the communities those charities exist to serve.
Sources & Further Reading
Walter Tevis, Mockingbird (1980)
Ursula Hegi, Stones from the River (1994)
Kate Crawford, Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence — Yale University Press
Stanford HAI, AI Index Report 2024 — aiindex.stanford.edu
MIT Work of the Future — workofthefuture.mit.edu
Kimber, R., Ehrlich, V., Vance, B., & Osei, A. (2025). Collective Learning in Philanthropy: AI, Trust, and the Future of Grant Reporting. The Foundation Review (forthcoming)
Kimber, R. & Vance, B. Purpose as a Compass — Philanthropy New York / Foundation Review
About This Series
Article 1 explored moral obligation. Article 2 examined structural pressure. Article 3 addresses conscious stewardship. Articles 4–7 will move into design, funding, regulation, and relational infrastructure.
Rachel Kimber is a global philanthropy advisor and fractional CIO helping foundations align technology adoption with institutional values and mission. She co-created the AI Readiness Framework and speaks regularly on AI governance in the philanthropic sector.