AI is entering institutional life the way cars entered public life: first as convenience, then as infrastructure, and only later as a governance problem too urgent to ignore.
I have been thinking about AI through the lens of car culture.
Not because the technologies are identical. They are not. But because the social pattern is familiar. A powerful tool arrives. Its benefits are real. Adoption accelerates. Convenience becomes expectation. Commercial pressure outpaces public understanding. Harm appears unevenly at first, then more widely. And eventually society has to decide whether it is willing to build the norms, rules, infrastructure, and accountability systems required to live with that tool at scale.
We did not respond to the automobile by pretending it was neutral. We also did not ban driving because risk existed. We built roads, lanes, licenses, insurance, inspections, traffic laws, crash reporting, and standards of care. We created public expectations about who could drive, under what conditions, with what responsibilities, and what happened when something went wrong. In other words, we built governance around a technology that had become part of ordinary life.
That is where we are with AI now. And many of our institutions are not ready to admit it.
This spring, I kept hearing versions of the same warning in very different rooms. At PEAK Grantmaking, the language was community, service, and proximity. At NTEN, the formal debate asked whether nonprofits that fail to adopt generative AI will be unable to compete for talent and funding within five years — but beneath that motion was a more important question: what kind of adoption are we defending, and what kind of institutional unpreparedness are we normalizing? And at the American Bar Association, the conversation turned fully to delegated authority, liability, audit trails, and the hard question underneath all the rest: who is responsible when systems act on our behalf?
Different rooms. Different vocabulary. Same fault line.
Distance Changes Judgment
At PEAK Grantmaking, Jackie Joyner-Kersee opened with a reminder that where you come from matters. What your community gives you matters. The foundation beneath you matters. She spoke about East St. Louis, aspiration, discipline, excellence, and returning to serve with the same seriousness that once fueled athletic greatness.
It was moving. But it was also instructive in a way I did not expect.
Her story was not just about athletics or service. It was about the cost of distance. About what gets lost when institutions make decisions about people whose lives they have never had to live. About what becomes possible when the person closest to real conditions is the one trusted to lead.
That question of proximity has stayed with me for years. It matters in philanthropy. It matters in governance. And it matters in technology.
Because distance changes judgment.
The farther a system gets from the people living with the consequences, the easier it becomes to confuse procedure with wisdom, efficiency with care, and abstraction with stewardship. That is true of bureaucracies. It is true of algorithms. And it is true of any institution that mistakes speed for understanding.
The Governance Question Underneath the Hype
Then I moved into another room entirely — at the American Bar Association, where the vocabulary shifted to chains of accountability, delegated authority, AI liability, autonomous systems, warrants, audit trails, medical records, standards, and human review.
Different language. Same fault line.
Who is responsible when a system acts on our behalf?
Not in theory. In practice.
Who approved the tool? What data did it touch? What authority was delegated? What human was in the loop? Were they competent to intervene? What evidence exists after the fact? Can the decision be traced, replayed, audited, or challenged?
The more I listened, the clearer it became that the central challenge of AI is no longer simply adoption. The deeper challenge is whether our institutions still know how to govern themselves while AI is inside the workflow.
That is a harder question than most organizations want to ask.
Adoption is legible. Procurement is legible. Policy drafts are legible. Tool demos are legible. Governance is slower, less glamorous, and much more revealing. Governance forces institutions to confront questions they have often avoided for years: How are decisions actually made here? Who owns the risk? What counts as meaningful oversight? What gets documented? What gets disclosed? What gets escalated? What are staff permitted to do? What are they prohibited from doing? Which standards are real, and which are merely ceremonial?
This is why the car culture analogy matters.
Cars did not merely give people mobility. They reorganized public life. They changed law, labor, commerce, insurance, urban design, policing, and public expectations about speed, freedom, and risk. Once the car became normal, governance was no longer optional. It became part of what adoption meant.
AI is doing something similar inside our institutions. It does not merely help people work faster. It reorganizes the workflow. It changes expectations. It alters rhythms of review. It creates new dependencies. It reshapes what counts as acceptable speed, sufficient review, and ordinary judgment. It enters as assistance and quickly becomes infrastructure.
That is why I keep returning to a sentence that feels truer every week:
AI can support decisions. It cannot replace responsibility.
I mean that legally, ethically, and operationally.
A system can draft a memo. Summarize a report. Surface themes. Suggest language. Classify content. Flag anomalies. Generate notes. Translate text. Support research. Reduce burden. In some cases, it can do those things well. Good use is possible.
But the responsibility for judgment does not move simply because the work moved faster.
Boards have duties. Executives have duties. Managers have duties.
Care. Loyalty. Obedience. Stewardship. Trust.
None of that disappears because a model produced a recommendation in three seconds. And yet many organizations are behaving as though the presence of a tool changes the location of responsibility. It does not. What it changes is the speed at which poor governance gets exposed.
Speed Without Discipline Is Exposure
This is especially visible in regulated sectors. In health care, for example, the questions are already sharpening: how to assess AI-related risk, govern access, revise agreements, preserve evidence, and define meaningful oversight when automated systems shape records, diagnostics, or recommendations. Liability rises wherever authority becomes blurry and systems influence outcomes without clear lines of review. Those debates are not confined to medicine. They are simply easier to see there because the stakes are so legible.
But the broader pattern is visible everywhere. It is a philanthropy story. A nonprofit story. A higher education story. A legal story. A civic story. The sectors differ. The governance problem rhymes.
We are living through a period in which speed is treated as virtue. The race to market. The race to scale. The race to monetize. The race to signal relevance. But speed without discipline is not innovation. It is exposure.
One of the strongest ideas I carried out of these conversations was the need to engineer discipline for enthusiasm. That line has stayed with me because so much of our current AI discourse still runs on enthusiasm, fear, or vibes — less often on structure, less often on standards, less often on operating rules sturdy enough for ordinary people doing ordinary work under real constraints.
Many organizations now have an AI policy. Or a committee. Or a values statement. Or a folder of examples borrowed from peers. That may be a beginning. But a policy is not governance, just as a driver’s manual is not a transportation system.
Governance lives in practice.
Which tools are approved? For what uses? With what data? Under what conditions? What is prohibited? What requires disclosure? What requires legal review? What requires human sign-off? Which human? With what competence? What gets logged? What gets retained? What gets audited? Who owns the decision when the output influences the result?
Those are not side questions. That is the work.
Equity Is Not a Side Note
This is also where equity enters — not as a decorative concern, but as a central governance question.
AI bias is too often discussed as though it begins and ends with offensive outputs or obvious misclassification. Those matter. But bias also works quietly. It works through credibility thresholds. Through dialect and tone. Through assumptions about professionalism. Through whose language gets treated as evidence and whose gets treated as noise. Through who is asked for more documentation, more justification, more proof. Through which organizations are expected to absorb administrative burden with the least staff capacity and the smallest margins.
AI does not invent these inequities. But it can formalize them, scale them, and obscure them behind the language of efficiency.
That is why I remain unsatisfied when people reach too quickly for the phrase human in the loop.
A human where? Doing what? With what skill? Under what pressures? With what authority to interrupt or override? Using what standards? Backed by what institution?
A ceremonial sign-off is not accountability. An underskilled human clicking approve is not governance. The question is not whether a human is somewhere nearby. The question is whether the right human is in the right place, with the competence, authority, context, and support to exercise judgment when it matters.
That is a much more serious design problem than most organizations are prepared to face.
When Convenience Becomes Governance
It is also why I am increasingly interested in bounded and federated approaches to AI. In settings where trust, privacy, and professional responsibility are high stakes, it matters whether the data stays home, whether authority is validated at the point of action, whether there is proof of presence, whether decisions can be replayed, and whether the institution preserves evidence rather than merely generating output.
Because the goal is not only insight. It is trusted insight. And trust is not built by aspiration alone. It is built by design.
This is one reason I keep thinking about Sateesh Nori’s framing of the AI notetaker as a kind of Trojan horse. It enters the meeting as convenience. Nobody is quite sure who invited it. People shrug. Later, its transcript becomes the record, its summary becomes the memory, and its action items begin to shape the plan.
It is a small example. But a telling one.
Unaccountable systems do not always arrive dramatically. Often they arrive ambiently, frictionlessly, and by the time an institution recognizes that convenience has become governance, the habit is already in place.
That may be the real danger in this moment.
Not only misuse. Not only dramatic failure. Not only scandal.
The greater danger may be gradual normalization.
A tool joins the workflow. A recommendation becomes the default. A summary becomes the understanding. A draft becomes the final. A ranking becomes the decision-support layer. An AI feature becomes “just part of the platform.” And eventually no one can clearly say where human judgment ended and automated influence began.
That is not readiness. That is drift.
Governance Is a Form of Care
So what would seriousness look like?
Not panic. Not technophobia. Not one more panel where everyone agrees the issue is important. And not a shared drive full of policy PDFs that no one knows how to operationalize.
Seriousness would look like building institutions that can still explain themselves.
Clearer boundaries on approved and prohibited uses. Better training. Stronger oversight. Real role clarity. Disclosure norms with staff, clients, grantees, students, patients, and partners. Better data stewardship. More thoughtful procurement. Tighter contracting. Stronger review pathways. Preservation of evidence. Meaningful auditability. Practice guidelines that can survive contact with actual work.
In other words, it would mean doing for AI what societies eventually had to do for cars: recognizing that once a powerful tool becomes part of ordinary life, responsibility cannot remain abstract.
You need lanes. You need rules. You need training. You need enforcement. You need consequence. You need infrastructure. You need a shared understanding of what safe enough means, who decides, and who is answerable when things go wrong.
If that sounds unromantic, good.
Governance is often unromantic. It is also one of the highest forms of care.
That is the throughline I kept hearing all spring — from PEAK to NTEN to ABA, even as the vocabulary changed. The institutions worth trusting will not be the ones that moved fastest or sounded smartest. They will be the ones willing to build the discipline underneath the enthusiasm. The ones that stayed close enough to real life to govern wisely. The ones that understood that technology may accelerate work, but it does not absolve duty.
The organizations that learn this now will not be the ones falling behind. They will be the ones still able to explain what they are doing, why they are doing it, who it serves, and who is answerable when it matters.
Further reading:
Sateesh Nori, “The AI Notetaker: The Trojan Horse of Unaccountable AI Agents?”
Rachel Kimber, Valerie Ehrlich, Blanch Vance, and Abigail Osei, “Collective Learning in Philanthropy: AI, Trust, and the Future of Grant Reporting,” The Foundation Review