Every major technology in history has eventually forced the question: who does this serve, and at what cost?

The printing press democratized knowledge and also enabled the mass printing of propaganda. The telegraph accelerated commerce and also enabled industrial-scale financial fraud. Nuclear energy powers cities and can sterilize them. The internet connected humanity and built the most efficient surveillance and manipulation infrastructure ever created.

AI is no different. What is different is the pace — and the fact that for the first time, the builders, the ethicists, the regulators, and the public are all being forced to have the conversation in real-time, while the technology is still forming. That’s either terrifying or genuinely hopeful, depending on your disposition.

I lean toward hopeful, but only with clear eyes about what the risks actually are.

The Frameworks

Before I tell you what ethics looks like from the ground, let me give you the top-down view — the frameworks that the serious institutions have built.

The USAID AI Ethics Guide is one of the better ones. Written for practitioners deploying AI in development contexts — international aid, government programs, civil society — it grounds ethical AI in something concrete: capabilities. What can this system actually do? What decisions is it making? Who has accountability when it goes wrong? It avoids the trap of abstract principles and pushes toward operational questions. That’s useful.

The RAFT framework — Reliable, Accountable, Fair, Transparent — comes out of the responsible generative AI conversation. It’s a checklist, more or less. Is the system reliable enough to be trusted with this decision? Is there a human accountable for its outputs? Is it fair across different populations and use cases? Is it transparent about its reasoning and limitations? Four letters, but each one hides enormous complexity when you try to actually implement them.

The EU AI Act takes a risk-based approach. High-risk applications — biometric surveillance, employment decisions, credit scoring, critical infrastructure — face strict obligations. Lower-risk applications get lighter treatment. The theory is sound: regulate proportionate to consequence. The practice is messy, as all regulation of fast-moving technology is.

These frameworks exist, they’re thoughtful, and they were mostly written by people who have never deployed an AI system in production.

I say that not to dismiss them — they capture real hard-won thinking — but to flag the gap. There’s a difference between a framework written in a conference room and the moment you’re debugging a production system at 11pm and a decision gets made that you didn’t fully anticipate.

What Ethics Actually Looks Like

I build AI agents. Not for a corporation with an ethics review board and a legal team. Out of a home office in Hampton, Virginia, running on a Mac mini.

Here’s what ethical AI development actually looks like from that position.

It starts with permissions. When I set up Zephyr — my home AI system — the first design question wasn’t “what can this system do.” It was “what should this system be allowed to do, and who can authorize what.”

That sounds obvious. It isn’t. Every AI system I’ve seen in production has permission scope that’s either too broad (“the model can do anything”) or poorly documented (“I think it can do X but I’m not sure”). The first version of my system had the same problem. The discipline required to define what a system can touch, what it can initiate, what requires human confirmation — that discipline is the foundation of ethical deployment. Not a principle. A specific technical decision, made repeatedly, audited regularly.

Transparency isn’t a feature — it’s a default. The RAFT framework calls for transparency in AI reasoning. What does that mean in practice? For me, it means: every action the system takes gets logged. The system can explain why it made a decision. There’s a clear audit trail from input to output. When the system is uncertain, it says so. When it’s operating near the edge of its capabilities, it flags that too.

This matters more than it sounds. AI systems have a confidence problem — they present uncertain outputs with the same surface-level assurance as certain ones. A system that’s wrong with conviction is more dangerous than one that’s wrong but hedged. Building transparency into the architecture from the start means building uncertainty quantification into every decision point. That’s not a nice-to-have. It’s load-bearing.

Fairness is a distribution problem. The AI bias conversation often gets framed as a values problem — “biased AI is bad AI.” That’s true but incomplete. Bias in AI systems is usually a data distribution problem that surfaces as a values failure. The facial recognition systems that misidentified Black faces weren’t built by racists (mostly). They were built on training data that overrepresented certain demographics and underrepresented others. The system learned to be accurate on average in a way that was deeply inaccurate for specific populations.

This is the Acemoglu problem in miniature. Technology can produce aggregate gains that mask specific losses. A hiring algorithm that improves average hiring quality by 10% while systematically filtering out qualified candidates from underrepresented groups has produced a gain and a harm simultaneously. The gain is visible; the harm is statistical and invisible until someone looks for it.

The ethical obligation isn’t just to build systems that work. It’s to ask: who does this work for, and who does it fail?

Real Failures

The frameworks are good. The failures are instructive.

COMPAS — the recidivism prediction algorithm used in the US criminal justice system — turned out to have roughly equal accuracy overall, but systematically over-predicted recidivism for Black defendants and under-predicted it for white defendants. The system was accurate in aggregate and discriminatory in practice. An ethics review that looked only at overall accuracy would have signed off on it.

Amazon’s hiring algorithm — trained on historical hiring data — learned to penalize resumes that included words like “women’s” (as in “women’s chess club”) because the historical data reflected a male-dominated industry. The system optimized for fit with the past, not fit for the future. Amazon abandoned it, but not before it ran for years.

Clearview AI scraped billions of public photos to build a facial recognition database, sold it to law enforcement, and created a biometric surveillance infrastructure that no one voted for and most people didn’t know existed until journalists found it. No explicit regulation prevented it. The framework gaps made it possible.

These aren’t obscure edge cases. They’re representative. The pattern is consistent: a technically functional system deployed without adequate attention to who it affects, how it fails, and what second-order consequences it enables.

The Gap Between Theory and Ground

Here’s the thing about frameworks: they’re written at the level of principles. “Ensure fairness.” “Maintain transparency.” “Establish accountability.” These are right. They’re also insufficient.

The gap is implementation. A small team building a useful AI product doesn’t have a dedicated AI ethics officer. They have a deadline, a limited budget, and a genuine desire to build something that helps people. The question isn’t whether they want to build ethically. Most people do. The question is whether they have the operational tools and the institutional support to do it.

This is where I think the regulatory conversation often goes wrong. The EU AI Act, for all its rigor, is primarily legible to large organizations with compliance departments. A small developer building a healthcare scheduling tool — which might qualify as high-risk — faces the same documentation requirements as a Fortune 500. The principle is sound; the implementation burden is unevenly distributed.

What would actually help:

Clear checklists for specific use cases. Not “ensure fairness” but “for a hiring tool, run these tests, document these assumptions, log these outputs.” The USAID guide gets closer to this than most. The RAFT framework is a start. We need more operational specificity.

Default architecture patterns. If logging, audit trails, and uncertainty quantification were baked into standard AI development frameworks the way error handling is baked into modern programming languages, more systems would have them. Right now, they’re optional extras that get cut when time is short.

Adversarial testing as standard practice. Red-teaming — deliberately trying to make a system fail — should be as routine as unit testing. Most development workflows don’t include it. The teams that do catch problems early. The ones that don’t ship them to users.

Real accountability. This is the hard one. Most AI ethics frameworks are voluntary. The COMPAS story only became public because journalists investigated. Amazon’s hiring algorithm only stopped when internal advocates pushed hard enough. The incentive structure currently rewards deployment speed over ethical rigor. That has to change — but it will only change through some combination of regulation, litigation, and reputational consequences that make the calculus shift.

The Lived Version

I started this post with frameworks. Let me end with something more personal.

I run an AI system that reads my email, manages my calendar, operates my business, and interacts with my clients. It has significant access to my life. I’ve thought carefully about what I want that to mean.

What I’ve landed on: the ethics of my system aren’t primarily about preventing harm to others — though that matters. They’re about what kind of relationship I want to have with the technology I’ve built. Do I want a system that does whatever it can? Or a system that does what I’ve authorized, within boundaries I’ve set deliberately, with full visibility into what it’s doing?

The second one is harder to build. It requires more explicit design, more discipline, more ongoing attention. But it’s the one I actually want to live with.

The ethical compass isn’t a checklist. It’s a set of ongoing questions: Who does this affect? What can go wrong? Who’s accountable when it does? Is the benefit distributed fairly? You ask them at the beginning of a project and again when you’re debugging at 11pm. You don’t answer them once. You answer them every time the system encounters something new.

That’s not a burden. It’s just what it means to build technology that’s actually worth building.

Next in the series: Post 03 — The Acemoglu Problem. A deep dive on Power and Progress and the specific mechanism by which AI concentrates capital. How productivity gains from technology consistently flow to the top, and what (if anything) changes this time.

Sources: USAID AI Ethics Guide (PDF, References/); RAFT framework via “Build Responsible Generative AI Applications” eBook; EU AI Act (2024); Acemoglu & Johnson, “Power and Progress” (2023); ProPublica “Machine Bias” (2016, COMPAS analysis); reporting on Amazon hiring algorithm (Reuters, 2018); Clearview AI coverage (NYT, 2020); Michael Wade voice memo recordings 1-3 (2024, ethical AI lecture script) — notes in research/voice-memo-analysis.md