We all know the story of the blind men and the elephant. Fun parable that you’ve probably been taking the wrong lesson from. I’ve used this one in facilitation for years, because the wrong lesson is so familiar and the right one is so useful.
The standard read: six blind men each touch a different part; each describes the elephant differently, and the point is humility. Nobody gets the whole. Your own view is partial. Be modest, everyone has their own point of view. Fine, I guess.
But look at what each of them actually said:
- The one holding the trunk said it was like a snake. He was right.
- The one at the ear said a fan. Right again.
- The leg, a tree. Sure.
- The side, a wall. Yes.
- The tusk, a spear. Okay.
Not one of them was wrong about the part they had.
The parable was never about being blind or opinionated. It is about what a group does with several accurate partial truths. And there, most groups fail in one of two ways.
They argue until one man wins. The loudest, or the most senior, or the most confident. Now the group has one man’s elephant, and it threw away five true observations to get there.
Or they compromise. They average the reports into “it’s some sort of a snake-fan-tree-wall-spear” and produce a description that matches nothing. The blended answer is worse than any single honest one, because at least the man at the trunk could tell you something real about the trunk.
Pick a winner and you lose five truths. Average them and you lose all six. The elephant is in neither move.
The third option is the one almost no room reaches by default, but it works. It goes like this:
- Keep the six reports intact and distinct
- Hold them next to each other (without collapsing them)
- Build the elephant from the structured combination
- Put just one person on the hook for the synthesis
The disagreement between the men was never the problem to resolve; it was all valid observational data. Six people who agreed would have told you about one body part. Six who disagreed, kept in relationship, can hand you the whole elephant.
I’ve been thinking about this a lot over the last year or two, because it’s exactly what happens the moment you put a real question to more than one AI model.
Ask three or five of them the same hard thing and they will not agree. And the reflex, the one built into almost every tool, is to make the disagreement go away. Pick the one true model that you trust. Or blend the answers into one smooth paragraph that reads like consensus. Both approaches delete the most valuable thing you were just handed, the data. The models disagreed because they were touching different parts of the thing. The disagreement IS the map.
The blind men (and the AI models) were never the problem.
Most of the work I do now is a long argument with that reflex. Open protocols, plain files, that keep the independent reports on the record instead of averaging them into a confident mush, and that keep a human accountable for the elephant at the end. Not because more voices are automatically wiser. Six people shouting is not intelligence either, no matter how well they may see. But six kept distinct, kept honest, and combined on purpose is the only way anyone has ever seen the whole animal.
That is the principle of Aggregated Intelligence, without the AI part. It holds in boardrooms and cockpits and courthouses, at the scale of institutions. It holds just as hard in the smaller rooms, the ones with fewer people and higher stakes. As you may have noticed, the blind men are in all those rooms too, and the pull to pick a winner or blend them to mush is strongest of all when the elephant is something you love.
And it holds at the speed of machines. AI councils and town councils have more in common than they care to admit. The same principles apply to human-AI pairings at work, and to multiple people plus multiple AI agents. The disagreement is the map.
See this in principle in action through these things I made:
- Harnessie: multi-agentic harness
- PAICE.work: People+AI Collaboration Effectiveness metric
- AIDR: human-owned decision records via hot-swappable AI partners