The race in artificial intelligence is increasingly a race to know more about us.
Our messages. Our calendars. Our relationships. Our work. Our purchases. Our preferences. Our history. Eventually, perhaps, enough of our lives to construct something resembling a digital twin.
That possibility is enormously powerful.
It is also where we believe much of the industry is starting with the wrong question.
The question is usually:
If an AI understands you this well, what should it be allowed to do for you?
We think the better question comes first:
How much authority should something that understands you this well actually have?
At UnKAnscious, our answer is deliberately asymmetric:
Maximum useful understanding. Minimum necessary authority.
An AI should be capable of developing an extraordinarily rich understanding of the person it serves.
That does not mean it should automatically gain equivalent power to act.
Those are different things.
And the distinction may become one of the most important design principles of the AI era.
Imagine an AI Chief of Staff with access—because you explicitly gave it access—to years of your email.
It might eventually understand that a person who appears occasionally in your inbox is actually one of your most important professional relationships.
It might recognize an unresolved commitment buried across six conversations.
It might know that somebody has been waiting for you for three weeks even though there is no unread email.
It might connect a discussion from February, a decision from April, and a new message from August and tell you:
Something changed here, and you should probably know about it.
That is useful understanding.
But none of it requires the AI to send an email. It doesn’t require permission to delete anything. It doesn’t require permission to purchase something. It doesn’t require permission to make a commitment on your behalf. And it certainly doesn’t require permission to quietly optimize your behavior toward an objective somebody else chose.
The industry often bundles these concepts together under the word autonomy.
We think they should be separated.
An AI can have broad permission to understand while having narrow permission to act.
This distinction became clearer to us while building GoalRunner.
We originally confronted email as most software does: messages, threads, inboxes, follow-ups, priorities.
But an inbox is only superficially a collection of messages.
Look across enough history and something else emerges.
Your email contains fragments of your life and work: relationships, promises, decisions, projects, negotiations, introductions, problems, opportunities. Things you’re waiting for. Things other people are waiting for from you. Things everyone assumed were settled but weren’t.
A conventional email application mostly shows you the containers holding this information.
We became interested in reconstructing the world represented by the evidence inside those containers.
That leads to a very different product question.
Not:
“Which email should I answer next?”
But:
“Given everything you know, what should I know, decide, and do now?”
Those are profoundly different computing models.
One organizes information. The other attempts to develop understanding.
Email is only one evidence source.
Calendars contain another part of the story. Documents another. Messages another. Business systems another.
Eventually, an AI working for an individual could develop a continuously improving model of their relationships, responsibilities, goals, preferences, history and circumstances.
We think of that as a digital twin.
But our conception of a digital twin is different from the surveillance-oriented version the term can imply.
The twin should exist for the human being. The human should determine what it can understand. The human should determine what it can do. The human should be able to inspect why it believes something.
And greater understanding should never silently become greater authority.
That produces an architecture with separate boundaries.
An AI might be permitted to read broadly — but only recommend a response.
It might be permitted to understand financial circumstances — but not move money.
It might be permitted to recognize a deteriorating relationship — but not contact the person.
It might be permitted to identify an opportunity — but not commit the human to it.
Authority should be granted deliberately, capability by capability.
This creates a different path toward autonomous AI.
The prevailing narrative suggests that better AI naturally leads toward granting it more control.
We don’t think that’s necessarily true.
A system can become dramatically more intelligent without becoming dramatically more powerful.
In fact, we think trustworthy autonomy requires several things that are less glamorous than autonomy itself: evidence, provenance, uncertainty, reversibility, explicit authority, human override, and clear boundaries between recommendation and action.
These aren’t obstacles standing between today’s AI and some autonomous future. They are the infrastructure required to reach that future responsibly.
Trust cannot be an interface added after the intelligence works.
It has to exist in the architecture.
This leads to the larger principle behind UnKAnscious.
We believe technology should increase human agency.
That sounds obvious until you examine how much modern software does the opposite.
People increasingly work for their software. We monitor inboxes. Maintain task systems. Transfer information between applications. Check dashboards. Reconstruct context. Remember which system contains which part of reality.
And increasingly, we supervise AI systems that were supposedly going to eliminate this work.
The human remains the integration layer.
That seems backwards.
If AI can understand our world, software should increasingly absorb the complexity of operating software—not transfer additional complexity to the person.
The interface should move toward outcomes.
The human says what matters. The system develops understanding. The AI surfaces what deserves attention. The human retains the consequential decisions.
AI recommends. Humans decide.
And when humans choose to delegate something, they should know exactly what authority they delegated.
The most useful personal AI may eventually know an extraordinary amount about us.
That understandably makes people uncomfortable.
The obvious response is to deliberately make the AI understand less.
Sometimes that is absolutely appropriate.
But there is another important lever:
Limit authority rather than unnecessarily limiting understanding.
A Chief of Staff who knows almost nothing cannot help very much.
A Chief of Staff who knows everything and can do anything is dangerous.
The interesting design space lies between them: an AI capable of extraordinary understanding, operating inside deliberately constrained authority.
That is the architecture we believe can produce something much more valuable than another chatbot, another productivity application, or another autonomous agent.
Something that understands the world around you without trying to own it. Something that helps you exercise judgment rather than replacing your judgment. Something whose increasing intelligence increases your agency, not its own.
That is the kind of AI we want to build.
And it leads to a principle we expect to keep returning to:
The more an AI understands about you, the more carefully we should design what it is allowed to do with that understanding.
Maximum useful understanding. Minimum necessary authority.
That is not a limitation on the future of AI.
We think it is one of the ways we get there.
— UnKAnscious