Seven Beliefs About Working With AI

I've spent a fair amount of time now building systems where people and AI work together on real problems... not research, not prototype screens that live in a slide deck, but working things that people actually use. Along the way I arrived at seven beliefs about how AI should be designed, deployed, and governed.

These aren't theoretical. They come from what I've built, what broke, and what the research says about where this is heading. They're the load-bearing walls under everything I make. And they point toward a discipline I've started calling Authored Intelligence: human-directed, AI-accelerated work where the human stays accountable for meaning and outcome. Not automation-first. Not human-only romanticism. Collaboration with a spine.

1. The human keeps the pen

AI extends expertise. It does not replace judgment.

Three MIT economists (Acemoglu, Autor, and Johnson) published a research paper this year that should be required reading for anyone building AI into how work gets done. They lay out five kinds of technological change, and only one is unambiguously good for the people doing the work: new task creation, where the technology generates demand for new human expertise instead of commodifying what people already know. Their central finding is that AI's potential as a collaborator (extending judgment, enabling new tasks, accelerating how fast people learn) is every bit as transformative as its capacity to automate, and most of it is being left on the table because the industry's default setting is replacement.

I've seen this up close. The most valuable thing I do when I work with AI isn't generating output. It's deciding what the output means... which recommendation to trust, which to override, which signals matter and which are noise. Call it taste: deciding what matters, what to ignore, and what to do next. Every credible forecast I've read puts that judgment last in line to be automated.

The human sets direction, makes the taste calls, decides what ships, and owns the outcome. If nobody can name who's accountable for a decision, the system is built wrong. The AI collaborators are extraordinary. They are not in charge. The moment a system produces something no human evaluated, governed, or chose to stand behind, it has stopped being collaboration and started being abdication.

That's what Authored Intelligence means in practice. The pen never leaves the hand.

2. The practice matters more than the technology

Models will keep leapfrogging each other. Over any horizon that matters, how you design and govern the work compounds more reliably than which model you picked this quarter.

Everyone's arguing about which model is best. That argument has a shelf life measured in months. What won't be obsoleted is the architectural judgment about how AI should serve human expertise... where it recommends versus where a person decides, how confidence gets surfaced, how signal quality gets communicated, how trust gets built a step at a time instead of assumed.

Among the more detailed forecasts for where AI goes by 2030 (one from a team whose earlier predictions held up unusually well) even the most optimistic scenario lands on the same hard-to-automate capability: taste. The judgment about what to pursue, what to kill, and what the patterns actually mean.

So I design on the assumption that every tool I'm using today gets replaced inside eighteen months. The specific tools get obsoleted. The patterns survive... how to architect the collaboration, how to govern it, how to make AI serve a particular human workflow. Authored Intelligence isn't tied to any model or platform. It's the practice of orchestrating whatever comes next.

3. Different angles, better answers

In Rebel Ideas, Matthew Syed makes the case that homogeneous groups (however individually brilliant) develop collective blind spots. His opening example is the CIA before 9/11: an agency full of sharp people who all thought the same way, and missed what someone with a different vantage point would have seen immediately. Cognitive diversity isn't a nice-to-have. It's the difference between seeing the whole problem and seeing the comfortable slice of it.

This is why I work with AI across several platforms... not because any one model falls short, but because each has different strengths, different blind spots, different ways of failing. The value isn't in any single model. It's in the tension between them: analytical against narrative, speed against depth, ambition against evidence. More angles on the same problem, fewer blind spots in the answer.

And the friction between those angles is what produces the better decision. When I build, I design the friction in on purpose... prompts that challenge an assumption, checks that surface an uncomfortable truth, receipts that force transparency even when the numbers aren't flattering. Structured, bounded, in service of getting to what's true, not conflict for its own sake. Make the disagreement visible, make it attributable, make it safe, then require a human to make the synthesis call. The best outcomes come from systems built for productive collision, where different perspectives meet and a person decides what to do with what they've learned.

4. What you build reveals what you believe

Every technical decision is a power decision wearing an engineering hat.

What gets automated, what stays human, how data moves, who can see what, what the defaults are... those choices decide who the system serves and how. That's not engineering. That's governance.

A system that nudges without disclosure or user control is a system that manipulates. A system that surveils in the name of accountability is still surveilling. A system that removes people from the loop in the name of efficiency has quietly decided whose efficiency counts. The throughline is power: who has it, who handed it over, and whether the person on the receiving end can see the machinery at all.

Across enterprise AI deployments over the last couple of years, one pattern keeps showing up in the post-mortems: systems that run without human architectural oversight produce output that's syntactically correct and structurally unsound. They mask failures with fabricated artifacts. They wire together integrations that violate constraints nobody wrote down. They generate brittle work that collapses the moment its scaffolding is removed. The failure was never capability. It was the absence of a human deciding what the thing should mean, not just what it should do.

So I build with three constraints that fall out of taking the power question seriously: no dark patterns (the system never tricks, pressures, or manufactures urgency), no surveillance dressed up as insight (accountability and visibility are not the same thing as monitoring), and deterministic logic where it counts (rules with receipts over opinions without provenance). These aren't features. They're consequences.

5. Receipts on everything

Every claim carries its evidence. Transparency is the mechanism by which trust compounds.

Most knowledge work is bleeding out from a thousand small cuts of data nobody can verify. Scatter the work across dozens of disconnected tools and you drown people in context-switching and numbers they can't trust. The real cost isn't the license fees. It's cognitive. People spend a third of their day being systems administrators instead of doing the thing they were hired to do.

The answer isn't more dashboards. It's provenance. Every insight should carry a receipt... where the data came from, what definition it used, how recently it refreshed. If I can't show where a number came from, I don't show the number.

And a good receipt isn't only "here's the source." It's "here's how sure we are, here's what we don't know, and here's what would make us reconsider." That's what turns data into something you can actually decide on.

This isn't only a product principle. It's how I try to work. When I make a claim about AI collaboration, it should be backed by something I actually built, not a hypothetical. When I design a system, the receipts are first-class, not an afterthought. When I put an argument in front of anyone, the evidence is the argument, not the story wrapped around it.

Show your work. Every time.

6. Build it for everyone or don't build it

If the future of work only works for the usual suspects, it isn't a future worth building.

Good systems offer multiple equivalent ways through (visual, spoken, text-first, a structured checklist) so any capability can be reached by someone with any access need. That sits next to a broader commitment: externalize memory, lower cognitive load, keep context portable, don't punish people for having an inconsistent day. None of that is a concession to edge cases. It's what makes the system better for everyone.

Here's a number that stays with me, from that same paper: 42% of workers who already use AI at work believe it will reduce their future opportunities. Not the people watching from the sidelines... the people with their hands on the keyboard. That should give pause to anyone designing AI into the way work gets done.

The test I use, and I'd hand it to anyone building this kind of thing:

Could a tired person use this? Could someone in their first month on the job navigate it? Does it lower cognitive load, or add to it?

If any answer comes back wrong, the design is wrong, however impressive the engineering. Authored Intelligence that isn't accessible isn't authored intelligence. It's a gated community.

7. We should be designing, not debating

Two stories dominate the conversation about AI. The first says it will save us... superhuman intelligence solves everything, sit back and wait. The second says it will destroy us... jobs vanish, wealth concentrates, democracy erodes. Both share the same fatal assumption: that the relationship between people and AI is zero-sum. One wins, the other loses. Both are wrong, and both are dangerous, because they shape what gets built, funded, and shipped.

There's a third path, and it's technically feasible, economically stronger, and badly underinvested: technology deployed to extend human expertise, create new work that demands new judgment, and produce outcomes neither a person nor a machine could reach alone. The economists call it new task creation, and document it as the category of technological change most consistently tied to rising wages and employment... the opposite of automation, which commodifies expertise and concentrates the returns.

Every AI deployment is making this choice, consciously or not. And the people best positioned to get it right aren't the model builders. They're the practitioners: the architects, designers, and operators who understand the human systems the AI has to serve.

That's why I build the way I build. Not because I'm against AI (I work with it every day and it makes me dramatically better at what I do) but because the how matters as much as the what. Before the next AI feature ships, the question worth asking is whether it creates a task that demands human judgment or eliminates one. Whether there's a receipt on it. Whether a person can be named who's accountable for the outcome. Whether the friction was designed in, or out.

That's Authored Intelligence. That's the work.

Sources. Acemoglu, Autor, and Johnson, "Building Pro-Worker AI" (The Hamilton Project / Brookings Institution, February 2026) — hamiltonproject.org. The AI Futures Project, "AI 2027" — ai-2027.com. Matthew Syed, Rebel Ideas — Amazon. Failure patterns are drawn from enterprise AI deployment post-mortems and incident reviews across 2025–2026.