From Investors to Law Professors: A Panel of AI Personas Weighs In, and What I Decided After

Chapter 5 ended on a name. I wrote that “Moral Mirror” only sets a direction, it can’t substitute for design, and promised that the next chapter would move to why I’d borrowed a roundtable to stress-test that structure. This is the chapter that keeps that promise.

But when I reopened the old folder to write it, I got a different picture than I expected. I remembered being persuaded, especially by the Anthropic-flavored opinion, and turning the original concept quite a bit because of it. Once I actually opened the folder, that memory turned out to be half right and half wrong.

Why I Ran a Panel

The core mechanism of this project is to deliberately unsettle a user’s reasoning. Throw a good counterexample, make the moment their certainty collapses. The problem is that this mechanism shares an edge with manipulation. From the outside, unsettling someone with a good question and unsettling someone with a bad one look nearly identical.

Judging that line alone, on a tool I’m building alone, felt dangerous. Whether a counterexample I like is actually a good one, or just one that fits my own taste, isn’t something I can reliably tell from the inside. So I needed other eyes. Running the panel wasn’t about getting answers. It was closer to lighting up my own blind spots.

The method was two overlapping rounds. In the first round (R1), each persona evaluated independently, without seeing anyone else’s answer, to keep the signal clean. In the second round (R2), each one read all the others and revised its position. I started with seven perspectives, then expanded to ten the next day: YC, a16z, Anthropic, OpenAI, AWS, Google, McKinsey/BCG, Khan/Coursera, Harvard Graduate School of Education/MIT Media Lab, and Stanford Law/Bioethics. Merging both sets later brought the total to seventeen perspectives.

What Each One Argued

Even now, the setup itself feels a little strange. I assigned roles, a YC partner, an a16z partner, an AWS enterprise lead, a Stanford Law professor, and had each one review my hobby-adjacent project through its own lens. I borrowed the names of real organizations; none of those organizations ever actually saw this project. And yet, feeding ten roles different material and different questions produced conclusions that oddly converged and conclusions that sharply diverged, at the same time.

Boiled down to the core claim of each:

Role	Core Argument
YC	· If a single trolley scenario can’t prove in 90 days that people want to come back, everything else is moot · Bring on a design co-founder, or shrink the interaction design to something outsourceable
a16z	· Unlike Character.AI, Calm, Replit, or Substack, neither the fun nor the utility case is demonstrated yet · Condition: 50+ people voluntarily returning within 8 weeks, with no external prompt
AWS	· Technical fit on Bedrock is natural, but enterprise procurement can’t reach a pre-compliance startup yet · Clinical ethics is an immediate non-starter without a liability shield
Google	· Direct acquisition carries reputational risk from hosting an “ethics product” · The learner reasoning-trace data fills a real gap in LearnLM’s training corpus; partnership before acquisition
McKinsey/BCG	· “Sandel isn’t a founder” - a consulting firm’s brand has to vouch for market entry instead · The consulting firm becomes both anchor client and distribution channel
Khan/Coursera	· MOOC economics (7-15% completion, ~5% paid conversion) simply doesn’t work here · Must narrow to cohort-based, adult learners, org licensing to survive
Harvard GSE/MIT Media Lab	· Mechanism lands cleanly on Festinger’s cognitive dissonance and Kapur’s productive failure · But ignores Kohlberg/Gilligan developmental stages; without safeguards it’s “pseudo-Socratic theater”
Stanford Law/Bioethics	· Adult entertainment framing is defensible, B2B training is contractually manageable · Once used for evaluation or hiring, or entering clinical ethics, current form is untenable
Devil’s Advocate	· Lists the death patterns of Replika, Character.ai, AI Dungeon, Inflection’s Pi · “A product that refuses to measure outcomes has also refused monetization”

Two lines stuck with me longer than the rest.

Sandel isn’t a founder. The fastest commercial path is a consulting firm’s brand vouching for the individual founder entering the market.

A product that refuses to measure outcomes has also refused monetization.

The enterprise-side role pushed back hard on the second attack. What enterprises actually buy isn’t “a better person,” it’s a reasoning trail they can hand to a regulator. Regulation had already shifted its unit of measurement from outcome evaluation to process evidence, so redefining the outcome as auditable output per seat, rather than a better human being, makes the vacuum the Devil’s Advocate described disappear. This rebuttal is the one that carried more weight in the second round.

What Cross-Reading Produced

Something interesting happened in R2. Five different personas, each blind to the others’ answers, arrived at the same conclusion independently. YC, a16z, McKinsey, AWS, and Khan all converged, in their own language, on the idea that a consulting firm should be the first customer. And three perspectives on safety, law, and pedagogy (Anthropic, Stanford Law, Harvard GSE) each referenced one another and said they should jointly own a single integrated safety stack.

More interesting still was what Stanford Law found in R2. After reading the other nine analyses, it surfaced five legal risks nobody had flagged in R1, contract clause issues buried in a16z’s proposed investment structure, joint-controller liability issues buried in OpenAI’s integration proposal. That’s a combination I wouldn’t have caught alone.

That’s where the panel earned its keep. What comes next is the problem.

Reopening the Folder Told a Different Story

I reread STATUS.md to check what I’d actually adopted. Solo operating mode was locked in. No seed funding. Consulting-firm anchors, external designers, a PhD advisor, legal counsel, all shelved. Clinical ethics entry was excluded outright, and right next to it sat a note I’d written myself: this was the external analyst’s recommendation, not my own instinct.

So, in effect, I threw out nearly every business-track recommendation from seventeen people. And yet I’d remembered being persuaded by the Anthropic-flavored view enough to turn the project significantly. If both are true, something doesn’t add up.

Reopening DESIGN_PRINCIPLES.md is where the contradiction resolved. The safety-design track had, in fact, survived almost intact. The table laying out the four reasons Sandel’s lecture hall isn’t manipulation, and exactly how each one goes missing in a 1:1 AI setting, and what fills that gap, was written into the spec sentence by sentence. The five principles (show reactions only right after a user submits an answer, show reasoning instead of ratios, show diversity of direction instead of majority opinion) were all still there. The decision to put “This tool is designed to unsettle your thinking” on the entry screen traced back to the same session.

Laid out plainly:

Track	Voices	Current State
Business & Scale	YC, a16z, McKinsey, AWS, Google, Khan/Coursera	Almost entirely discarded. Solo operation, no funding, external resources shelved
Safety & Design	Anthropic, Stanford Law, Harvard GSE	Absorbed sentence by sentence. Now the backbone of spec §5 and §9

Memory had fused two separate tracks into one. That I discarded the overwhelming majority of seventeen people’s advice, and that I was deeply persuaded by one of those tracks, are both true at the same time, and I’d been storing that as a single impression: “I turned a lot.”

---
config:
  look: handDrawn
  theme: neutral
---
flowchart TD
  A["17-persona panel<br/>2026-05-23 ~ 05-24"] --> B["Business & Scale track<br/>YC · a16z · McKinsey · Google · Khan"]
  A --> C["Safety & Design track<br/>Anthropic · Stanford Law · Harvard GSE"]
  B --> D["Discarded<br/>Solo operation · no funding · resources shelved"]
  C --> E["Absorbed<br/>DESIGN_PRINCIPLES.md §5, §9"]

It Was Never a Business. It Was Will.

Why it split this way isn’t hard to see in hindsight. I never started this project to run a business. It began as something closer to will, a question I’d carried for a long time that I wanted to actually build.

Every business-track recommendation pointed in a different direction but demanded the same thing underneath: gather users, make money, reshape the product to match that scale. Following any of it meant the question I’d originally been holding onto would blur, or in the worst case, have to be abandoned. Fitting a consulting firm’s client would reshape the tool into whatever they wanted. Taking funding would expand the scope to match an exit timeline. Either way, it becomes a different kind of thing than what I set out to do.

The safety-track recommendations, by contrast, never asked this project to become something else. They just told me how to do the thing I was already trying to do, unsettle someone’s certainty a little without hurting them, better. Advice that protects the user turned out to also be advice that protects the identity of the project.

So the line that split seventeen opinions into kept and discarded was never how polished the advice was. It was whether following it let me keep doing what I originally set out to do, or forced me to become something else. Business advice was always the latter, however good it was, and safety advice was always the former.

If I Ran the Panel Again

Writing this chapter, I asked myself one thing. Was the panel pointless, then? If I was going to discard the entire business track anyway, did I need to ask seventeen people in the first place?

My answer now is that the confidence to discard it would have been hard to earn without asking. If I hadn’t known that five people independently recommended a consulting-firm anchor, I couldn’t have been sure whether skipping it was my own taste or an actual gap in my review. Asking and then discarding is different from never asking at all. The first is a choice. The second is avoidance.

At the same time, I won’t pretend the process was comfortable. Listening to seventeen opinions and throwing out more than half of them can look strange to anyone who believes majority rules. But given that this project exists specifically to protect people from majority rule, a founder who doesn’t fold their own judgment in front of a panel’s majority opinion is arguably following the same principle, just applied to themselves.

What This Chapter Doesn’t Answer

This chapter recorded why I ran the panel and what I kept versus discarded. What survived was safety design; what I discarded was business expansion; and the line between them was never advice quality, it was whether following the advice still left this project as the thing I originally set out to build.

Still, some things remain unanswered. First, whether the safeguards I wrote into the spec as sentences will actually hold up in front of real users is still unknown, so far it’s a promise on paper. Second, I can’t be fully certain that this line, protects identity versus blurs it, is always the right line. Sometimes becoming a business is the only path that carries a project further. The next chapter looks at what that paper promise actually became, the place where the safeguards come down into code and prompts.