OPINION

We Can Build AI Faster Than We Can Explain It

As machines learn to act, the bottleneck is no longer capability. It is comprehension.

A robotic hand and a human hand reaching toward one another.

Opinion

3 min read

In late 2025, AWS and SANS published a paper on securing artificial intelligence as it starts to act rather than merely answer. It was written by people who help run security at that scale, including one of the internet's founding engineers, and it is precise, current, and correct in every line. It names the exact moment the ground shifts: the point where a system stops handing back sentences and starts taking actions on your behalf. Thirty-one pages, not a wasted claim. The paper is rigorous, written for the people equipped to read it. The challenge is carrying that rigor beyond them, into the classrooms and conference rooms where a wider group has to understand what AI can do.

The standard answer is to shorten it. An executive summary, a one-pager, a deck of the deck. Every organization that has ever needed a wide audience to understand a deep subject reaches for this, and underneath the reflex is a quiet assumption that understanding is a volume problem: too many pages, not enough time, reduce the pages. A shorter technical document is still only a document; understanding requires more. The executive who reads the one-pager can now repeat the conclusions, which is a different thing from seeing them, and the gap between the two is where bad decisions live.

Bar chart titled Sensitive conversations. Subhead: 57% of employees who use AI at work entered sensitive or high-risk information into a public AI assistant. Four bars show the type of information shared: personal data 31%, unreleased product or project details 29%, customer information 21%, confidential company financial information 11%. A footnote below reads 68% accessed public AI assistants through personal accounts, only 24% received mandatory training.
Artificial intelligence may improve productivity, but workers are feeding it more than prompts. Source: TELUS Digital/Pollfish, January 2025.

Consider how anyone actually comes to know a system. The knowledge arrives through use, not through description alone. You know a car because you have driven one, felt it respond, misjudged a corner once. If that is how understanding forms, then the honest way to explain a system is to build a working model of it and let people in.

That was the brief when I was engaged, after the paper was finished. The discipline is strict: you are not allowed to know anything the source does not. Every claim traces to a page. I introduced no new facts. What I built was the paper as a place you could stand in, one continuous system that changes state as you move through it, instead of a stack of slides describing it from outside.

Interactive

A visual walkthrough of how security responsibility changes when AI moves from answering questions to taking actions.

20–30 minutes
300-level
Nine stations

Enter the live walkthrough
A live, interactive walkthrough — not a recording. Click into it to drag the responsibility boundary and watch the controls respond, or use expand to open it full-screen.

The paper's real subject is responsibility. When an AI acts, who owns the action? On the page, that is an org-chart question. In the interactive experience, that distinction became a single line you could drag with your own hand: the border between the provider's territory and yours. Pull it one room deeper into your own operation and the meter climbs into amber, and three more controls light up as yours to run. Nobody had to be told where the responsibility sat. They found out by moving it, and a thing you have felt move is a thing you can still act on in a meeting six months later. A summary of the same material would have been read, agreed with, and gone by lunch.

The audience for this kind of understanding is suddenly enormous. When AI only answered, the people who had to grasp it were the people building it. Now the circle takes in everyone whose name is on the consequences: the executive who authorizes an agent to touch live systems, the counsel who signs off, the board that hears about it afterwards. The price of shallow understanding is rising too. A misunderstanding about AI used to produce a bad sentence. Now it produces a bad action, and the action has already run before anyone reads it. Nodding along has become expensive.

Expertise and compression are abundant. The harder craft is converting verified knowledge into an experience that leaves a working model in a thousand heads, at full rigor.

The people who wrote that paper did the hard part. They saw the shift early and set it down exactly. What remains is the growing, because AI that acts will be governed by the people who can picture it, and a picture cannot be handed over in summary form. It has to be built where the reader is standing. Right now we are producing the machines much faster than we are producing that.

  1. Landing screen of the interactive explainer The AI That Learned to Act. A headline reads the same, with a subhead about how security responsibility changes when AI moves from answering questions to taking actions. A three-by-three grid of nine stations sits to the right, from User and Agent through to Human, grouped into app, knowledge, action, and trust layers.
  2. The explainer with the Agent station selected in a row of nine. An inspector panel on the right, headed Agent / App, explains that the agent plans the work and routes calls to the model, data, and tools, and warns that a hijacked agent can chain together actions you never authorized.
  3. The explainer showing a trust-boundary slider dragged to scope 3, labelled pre-trained builder. The nine stations are mostly washed amber to show responsibility shifting to the builder. A panel headed Buyer vs Builder explains that scopes 1 to 2 are a buyer's mindset and scopes 3 to 5 are a builder's.
  4. The explainer at scope 5, self-trained builder, with every station washed amber to show full responsibility. Small CAP markers sit above the Tool and Action stations. A panel headed Control reads Permission boundaries. Throttles. Alarms, and explains that controls bound what agents may do and cap automated loops.
  5. The explainer with the Model station ringed by a dashed authorization boundary labelled outside. A panel headed LLM Reality 2 reads No guard inside the model. Guard the doors around it, explaining that in current state-of-the-art LLMs authorization cannot live inside the model, so the boundaries go around it.
  6. The explainer with three lenses shown above the stations, For AI, For Security, and From Threats, and the Monitor station selected. Tags read AppSec feedback, threat detection, and ATT&CK mapping. A panel headed AI for Security reads The same AI accelerates the defense, citing real-time AppSec feedback and OCSF plus MITRE ATT&CK threat detection.
  7. The closing screen of the explainer, beat 40 of 40. All nine stations are shown together. A panel headed The Complete System reads The AI That Learned to Act, with the line Before your AI acts, know what it can touch, what it can do, and who decides. A source note credits the AWS and SANS whitepaper, November 2025, with a link to read it.
1 / 7
The explainer opens on the shift it exists to teach: AI that no longer just answers but acts. Nine linked stations run from the user's goal to human review.