Working note 001 · April 2026

Can AI align with humanity if humanity isn't aligned with itself?

I used a tool for structured reasoning to examine its own reason for existing. Here's what happened.

The problem

Most AI alignment research starts from an assumption: there's a set of "human values" out there, and our job is to find them and encode them into AI systems. But that assumption has a problem. Humanity disagrees — profoundly, persistently — about what "good" means. Across cultures, generations, political lines, and individual lives, there is no consensus to encode.

So "align AI with human values" quietly becomes "align AI with the values of whoever builds it." That's not alignment. That's power dressed up in technical language.

I think there's a different approach. What if alignment isn't a destination — a correct set of values to find and lock in — but an ongoing process? A process where people surface their disagreements, reason through them transparently, and arrive at decisions whose logic anyone can inspect and challenge. Under this framing, the role of AI isn't to embody the right values. It's to support the infrastructure through which humans work out their values together.

That's what I'm exploring at epistemology.info. And to test whether the idea holds up, I did something recursive: I used a tool I've been developing to examine the thesis underlying its own existence.

What the tool does

The Epistemic Workbench is a reasoning tool. You give it a claim — any claim — and it breaks that claim apart into its supporting arguments, the evidence behind those arguments, the objections that challenge them, and the hidden assumptions the whole thing rests on. Then it tracks whether the claim holds up as you add evidence, raise objections, and respond to challenges.

How it works, briefly: The Workbench builds what researchers call an "argument graph" — a map of how ideas support and attack each other. Each argument gets a confidence score based on the strength of its evidence and the force of the objections against it. When you add a strong objection, the confidence of the claim it attacks goes down. When you respond to that objection convincingly, it goes back up. The structure is persistent — it doesn't vanish like a conversation. Anyone can come back later, find a weak point, and challenge it.

Two things make this different from an ordinary debate. First, the structure of the reasoning is visible. You can see exactly which evidence supports which claim, which objections challenge which argument, and whether those objections have been addressed. Nothing is hidden. Second, the math is separate from the AI. An AI helps break arguments down into their parts, but a formal reasoning engine — not the AI — computes whether the claim holds up. The AI can't hallucinate its way to a conclusion.

The thesis I tested

I started with a foundational claim in democratic theory: "Democracy requires an informed electorate." This isn't just an academic question — it's the premise that justifies the entire project. If democracy doesn't actually need informed citizens, then building infrastructure for structured public reasoning is a solution to a problem that doesn't exist.

The Workbench broke this claim down automatically and assigned it 72% confidence — accepted, but with significant vulnerabilities. Here's what the initial argument map looked like:

The argument, decomposed

Purple = the main claim. Green = arguments that support it. Red = objections that challenge it. Yellow = assumptions that haven't been examined yet.

72% Democracy requires an informed electorate
SUPPORTS Four pillars all point the same way: democratic legitimacy, good policy outcomes, meaningful deliberation, and government accountability all depend on an informed public
OBJECTION "The miracle of aggregation" — Maybe voters don't need to be individually informed. Maybe they use shortcuts like party affiliation, endorsements, and social signals that add up to good-enough collective decisions. (Lupia, Popkin)
SUPPORTS Historical pattern: across time and geography, when people have better access to information, democratic governance tends to be higher quality
OBJECTION The epistocratic critique — Democracies actually function tolerably well despite widespread voter ignorance. If the evidence shows voters are ignorant but democracies still outperform alternatives, maybe the thesis is wrong.
UNEXAMINED Is an informed electorate even achievable at the scale of a modern nation?
UNEXAMINED Can collective wisdom replace individual knowledge?
UNEXAMINED Does democracy actually aim at legitimate governance? (Or is it just a power-transfer mechanism?)

This is a fair summary of where democratic theory actually sits. The informed electorate thesis has serious support — from Jefferson through Dewey to Fishkin's modern deliberative polling experiments — but it faces real objections that haven't been settled.

What happened when I pushed back

Here's where it gets interesting. I entered evidence from the specific infrastructure I'm developing — and from recent research in computational argumentation — into the Workbench. The goal wasn't to "win" the argument. It was to see whether the project's architecture actually addresses the objections, or just talks past them.

The "miracle of aggregation" objection says voters can use shortcuts instead of deep knowledge. The response I entered: that objection assumes unstructured aggregation — people just voting based on gut feelings and heuristics. But what if aggregation happens through a system where every argument's supporting evidence and opposing objections are laid out explicitly? Where participants are verified real people (not bots or sock puppets)? Where the reasoning is saved so anyone can challenge it later? Under those conditions, you're not choosing between individual knowledge and collective wisdom — the infrastructure makes them the same thing. Heuristics become structured reasoning rather than cognitive biases.

The Workbench accepted this response. The objection was marked as answered.

The epistocratic critique is cleverer. It says: if voter ignorance research is itself contested, maybe we can't even determine who's "informed enough." The response: exactly — and that's precisely why you can't solve the problem by having some authority decide who's qualified to participate (that's epistocracy, rule by the knowledgeable). Instead, you make everyone's reasoning visible. The tool doesn't judge whether you're informed. It shows the structure of your argument, and weak reasoning is exposed by the structure itself — not by anyone's assessment of your qualifications. The contested evidence base is actually an argument for transparent reasoning infrastructure, not against it.

This response was also accepted.

For the three unexamined assumptions, I brought in specific evidence. Is an informed electorate achievable at scale? EU-funded deliberative democracy pilots (the ORBIS project) and decades of deliberative polling experiments suggest yes — structured deliberation produces more informed views even in large populations. Can collective wisdom replace individual knowledge? Recent research on combining multiple reasoning agents' arguments into a single robust framework suggests the question is a false choice — the combination process preserves individual reasoning while producing collective judgment. Does democracy aim at legitimacy? Democratic theorists from across the spectrum — Rawls, Habermas, even the minimalist Przeworski — converge on this, differing only in how thick or thin that legitimacy requirement is.

The result

After the intervention — all nodes resolved

Struck-through text = objections that have been addressed. Every assumption has been examined and accepted.

72% ✓ Democracy requires an informed electorate — all objections addressed, all assumptions examined
HOLDS Four pillars: legitimacy, outcomes, deliberation, accountability
ANSWERED "Miracle of aggregation" — Structured argumentation makes aggregation and individual knowledge complementary, not opposed
ANSWERED Epistocratic critique — If we can't determine who's informed, the solution is transparent reasoning, not gatekeeping
EXAMINED Achievable at scale — Yes: deliberative democracy pilots and platforms demonstrate it works
EXAMINED Collective vs. individual knowledge — False choice: combining frameworks preserves both
EXAMINED Democracy aims at legitimacy — Broad convergence across democratic theory traditions

What this is and what it isn't

This session was not a proof that "democracy requires an informed electorate." Someone could read the responses to the objections and find them unconvincing. They could raise new objections I didn't consider. They could challenge the evidence I cited.

And that's the point.

Everything is visible. Every objection is recorded. Every response is linked to the specific challenge it addresses. If you think the "miracle of aggregation" objection was answered too glibly, you can reopen it. If you think there's a hidden assumption I missed, you can add it. The argument graph doesn't disappear — it persists, and it's designed to be contested.

This is what I mean by "alignment as process." Not a static set of encoded values, but an ongoing, structured, transparent conversation where reasoning is visible and anyone can challenge it. The AI helps break arguments down into their parts. The formal engine tracks whether claims hold up. But the humans — all of them, not just the ones who build the system — do the judging.

What I'm building

The project behind this working note sits at the intersection of three things that rarely get built together: tools for structured reasoning (like the Workbench used here), verified digital identity (through Osmio, a digital municipality chartered at the ITU in Geneva that can verify you're a real, unique person without exposing who you are), and persistent, tamper-evident records (so no one can quietly rewrite the argument after the fact).

I think these are the preconditions for genuine collective sensemaking at scale — the kind that could, eventually, inform how we govern AI, how we make policy, and how we coordinate across the deep disagreements that make "aligning AI with human values" such a hard problem in the first place.

This is early-stage work. I'm sharing it as I go, including the rough edges, because transparency about the process is the whole thesis.