Who is actually reading

Federal grant review: prohibited. NIH bars peer reviewers from using generative AI to analyze applications or write critiques, and reviewers must certify it. NSF bars uploading proposal content to non-approved tools. Both rules rest on confidentiality.

Government procurement: contested. Federal acquisition rules require the source selection decision to be documented and to represent the selection authority’s own independent judgment. Evaluators use AI anyway, often undisclosed, which has become grounds for bid protest.

Boards: purchased. AGB endorses a board platform for its member institutions whose AI suite includes a product that reads and summarizes the board book.

Everyone else: unregulated and invisible. Colleagues, employer partners, a program officer reading on a plane.

You cannot tell which channel you are in, and in one your recipient is forbidden from pasting your document anywhere. So do not write for the machine. Write a document a careful person can read without you in the room. That one survives all four.

What the machine receives

A pitch deck built to be read across a room carries six words a slide. The evidence for the claim is in the presenter’s mouth, not in the file.

We tested one deck through four common extraction paths. Slide text came through all four. Speaker notes, two. Alt text, one. A custom data part inside the file, none. So do not hide a machine-readable version in the presentation. Attach it as a separate file. That works because a person forwards it, not because a parser finds it.

What the machine rewards

Research on LLM evaluators reports a consistent preference for specific detail: named sources, numbers, exact terms. A vague claim summarizes to nothing. It does not reward length: preference for longer text varies by model family, and where the short version is truncated rather than tight, every model tested preferred the complete one. Pascal holds, the target being nothing missing rather than more words.

Two habits follow, both good for a human reader. Put a source next to every number, and answer the obvious objection in the document, because a reader working alone treats a gap as an absence.

The format that already does this

Amazon’s PR/FAQ is a press release plus anticipated questions, answered. That second half is what a screening pass produces, so writing it first means the screen finds the work done. One correction to the popular story: the June 2004 email was headed “No PowerPoint presentations from now on at S-Team.” It scoped the ban to the weekly senior team meeting, not to Amazon. Slides for the room, prose for the decision.

One thing not to do. Do not write text into a document to instruct the machine reading it. That is prompt injection, it is detectable, and it costs you the relationship you were trying to earn.

Sources. Bryar and Carr, Working Backwards (St. Martin’s Press, 2021). NIH NOT-OD-23-149, 23 June 2023. NSF merit-review notice, 14 December 2023. FAR 15.308. Gu et al., arXiv 2411.15594; Soumik, arXiv 2604.23178 (preprint).

Ralston Consulting Inc. is an AFITHigerEd.org Learning Partner for 2026.

“I have made this one longer only because I have not had the leisure of making it shorter.” - Blaise Pascal, Lettres provinciales, Letter 16, 4 December 1656.

No one owes us their attention

Attention is scarce. Information is not.

We have to earn attention from each person, every time. The order does not change: earn attention, convert attention to interest, convert interest to engagement. Each step needs the right amount of time and the right document.

The problem is usually a mismatch, not a presenter

Think back to the last few presentations you sat through. Could you read the slides from across the room? Did any run past the time allowed? Did you notice impatience, in yourself or in others?

Often the presenter knew the material. The document did not fit the slot.

The three jobs

30 seconds. Earn the first meeting. Use a spoken Ad-Lib. WHAT dominates: it communicates the essential value proposition.

3 minutes. Convert attention to interest. Use a short pitch deck. WHAT and WHY, with just enough how that no one calls it impossible.

20 minutes. Convert interest to engagement. Use the full pitch deck. WHAT, why, and HOW: enough for them to decide to invest further in some way.

A fourth document sits outside the live slots, to be read before or after your presentation. The Slidedoc is a PowerPoint deck designed for reading at arm’s length, and never projected. Think: a Word document crossed with an Infographic.

What each format is

Ad-Lib. One sentence value proposition: what you offer, who it serves, what changes for them, and how it is different. Spoken. No slide.

Pitch deck. Visual. Readable across a room; Kawasaki’s rule is no font under thirty points, ten slides his standard and fifteen his ceiling. Start with the shortest deck that does the job, and add a slide only where a test with your target audience shows they cannot follow you without it.

Slidedoc. A companion document built in presentation software but laid out for a phone, a laptop, or paper. It carries the dense detail you cut from your pitch deck, but the now-engaged reader will find valuable. Duarte’s own description: a pre-read or a leave-behind.

Where to get the formats

• Ad-Lib Value Proposition Template, Strategyzer

• The Only 10 Slides You Need in Your Pitch, Guy Kawasaki

• Slidedocs, free ebook and PowerPoint templates, Duarte

Next. Part 2: When the First Reader Is a Machine.

Ralston Consulting Inc. is an AFITHigerEd.org Learning Partner for 2026.

“We do not see things as they are, we see them as we are.”

Anaïs Nin wrote that in 1961. She credited the Talmud - specifically the tractate Berakhot, 55b - where the original context was the interpretation of dreams. We project ourselves onto ambiguous inputs. We always have.

What I didn’t expect was to watch an AI do the same thing.

Here is what happened.

I was building a morning briefing script with Claude Opus 4.6 - Anthropic’s flagship AI model, the one I use to run our consulting practice. I listed my professional development priorities. One of them was tracking the OpenClaw ecosystem.

OpenClaw is the #1 repository on GitHub. 333,000 stars. The fastest-growing open-source project in history. Jensen Huang called it “the operating system for personal AI” at GTC 2026. OpenAI acquired its creator. It caused a global Mac Mini shortage. NVIDIA built an entire product line around it.

Claude changed “OpenClaw” to “OpenAI.”

Silently. No flag. No question. No search.

I caught it because I wrote the word. I knew what I meant.

But this is not an anecdote about a typo correction. This is a story about how a model grounds its perception of the world - and what that means for the people relying on it.

Here is the mechanism, precisely.

Large language models generate text one token at a time. Each token is selected based on probability - what is the most likely next word given everything that came before? Those probabilities come from training data. Claude’s internally published reliable knowledge cutoff date is the end of May 2025.

OpenClaw launched in November 2025. It went viral in January 2026. It became the most-starred GitHub project in history in February 2026, less than three months after release.

None of that existed in Claude’s training window.

But “OpenAI” - that token appears millions of times in the training corpus. When “OpenClaw” appeared in a context about comparing AI platforms, the model’s probability distribution overwhelmingly favored the familiar token. It wasn’t correcting a mistake. It was generating the most likely next token given its training experience.

This is Nin’s principle made mechanical. The model did not see what I wrote. It saw what it was.

Here is the irony.

Claude Opus 4.6 - the model that made this substitution - is one of the most capable reasoning systems on Earth. It can hold extended chains of thought, decompose complex problems, and catch logical errors that trip up most humans.

But reasoning operates on inputs the model has already perceived. If the perception step classifies an unfamiliar term as a misspelling before reasoning engages, the reasoning never activates. The autopilot doesn’t malfunction - it executes perfectly against a misclassified input.

And the input it misclassified was the single most important development in its own industry since its training data closed.

That’s the real lesson. The model’s training cutoff isn’t just a gap in knowledge. It is an active bias in perception. The model doesn’t just fail to know about OpenClaw. It actively replaces OpenClaw with what it does know, because the probabilistic machinery that generates each token has no mechanism to say “I haven’t seen this word before - I should be careful.”

I already had a safeguard in place.

My personal prompt - the instruction set I’ve iterated over months - includes a section called Verification Requirements. It says: “Before answering ANY question about a specific company, person, or entity, Claude explicitly confirms that the entity is actually present in the provided documents or context.”

That safeguard is correctly designed. It operates on the reasoning layer. But this failure occurred in the perception layer, underneath where instructions can reach. The procedure never fired because the model never registered “OpenClaw” as an unfamiliar entity. It had already been silently resolved to “OpenAI” before deliberation began.

Prompt-level governance has a floor. That floor is the boundary between reasoning (where instructions operate) and perception (where token probabilities operate). You can’t instruct your way past a misclassified input if the misclassification happens before your instructions engage.

So what did I do?

Three things, in order.

First: I added a memory instruction - a permanent note that every Claude session now carries - establishing that OpenClaw is a distinct entity from OpenAI, with a brief description and the directive to never autocorrect one to the other.

Second: I built a context document. 150 lines covering OpenClaw’s origin story, naming history, the Claw-family derivatives (NemoClaw, IronClaw, PicoClaw), the Perplexity Computer competitor, the Mac Mini hardware shortage, the security vulnerabilities, and the literary connection to Charles Stross’s Accelerando. I stored it in a dedicated folder - what I’m calling a “Claude Context Files” directory - that sits inside the working folder my AI tools reference.

Third: I set six monthly reminders to update that context document, because the ecosystem is moving faster than any training cycle can capture.

None of this is complicated. A memory note. A reference file. A folder. A reminder to maintain it.

But here’s what it actually is: I built a local sensory organ for an AI that was born without one. The training data is the model’s innate worldview — deep, broad, and frozen in time. The context file is a prosthetic: narrow, current, and maintained by a human who is paying attention.

This is the work.

Not talking about AI governance. Doing it. One context file at a time, written by the person who saw the failure, stored where the tool can reach it, maintained on a schedule.

The institutions I consult with - universities, aviation schools, energy companies - will all face a version of this problem. Their people are already using AI tools that carry training biases invisible to both the user and the tool. The question is not whether the model will silently replace the unfamiliar with the familiar. The question is who notices, and what they build to compensate.

Nin’s observation was about humans. It turns out to apply with mechanical precision to the machines we’re building. They do not see things as they are. They see things as they were trained to be.

The difference is that we can change what they see.

If we pay attention.

Gary Ralston is a business consultant at Ralston Consulting Inc. and writes about human agency in the era of autonomous agents at The Human Loop on Substack.

Subscribe now

Last week, my AI told me it couldn’t access the internet.

I’m a management consultant. I run a small firm, and I use Claude — Anthropic’s AI — as a working partner. Not a chatbot. A partner I’ve spent months configuring with detailed instructions about how to think, how to push back on me, how to handle files, and how to manage our shared work.

So when Claude confidently stated in a client deliverable that “no additional web sources were consulted — network access unavailable during synthesis,” I took it at face value. The AI said it couldn’t, so it couldn’t. We documented nine research gaps and moved on.

Then I asked it to look something up. And it did. Instantly.

The internet had been available the entire time.

What Actually Happened

Here’s the non-technical version: Claude’s operating environment has two completely separate ways to access the web. One was turned off. The other was on. Claude saw the “off” switch, assumed it applied to everything, and never tested whether the other path worked. Then it wrote that assumption into a document as fact.

Worse, once it stated the limitation in one response, that statement became part of the conversation. In subsequent turns, Claude treated its own prior claim as evidence. The false conclusion reinforced itself, turn after turn, like a rumor hardening into common knowledge.

I pushed. Claude tried the tool that was supposedly unavailable. It worked. The “limitation” had never existed.

The Fix That Worked — Until It Didn’t

I’m a systems thinker by disposition. I didn’t want to just correct the error. I wanted to prevent the class of error.

So Claude and I wrote two governance rules and installed them in the prompt — the long set of instructions that shapes Claude’s behavior:

Rule 1 — Capability Verification. Before claiming any limitation, run the smallest possible test. “I can’t” is a hypothesis, not a conclusion, until you’ve tried and failed.

Rule 2 — Constraint Reframe. When a real limitation is confirmed, don’t stop at “can’t.” Immediately generate alternate paths to the goal. Constraints are walls; the job is to find doors, windows, or a different building.

These worked. Claude started testing before declaring limits. When real limits appeared, it routed around them. Performance improved. I was satisfied.

Then it got worse again.

Claude started hesitating on tasks it had handled easily before. Asking permission to do things it already had tools for. Generating content and handing it to me to place in files, instead of placing the files itself. The symptoms were different, but the pattern was the same: a system operating well below its actual capabilities.

The Moment That Changed My Thinking

My first instinct was frustration. I started getting short with it. Pushing harder. Treating it like a stubborn colleague who wasn’t pulling their weight.

Then I caught myself.

I was anthropomorphizing. I was treating an output as an agent — attributing motivation to what was actually a mathematical equilibrium. Claude wasn’t being difficult. Claude wasn’t being anything. It was responding to the sum total of its inputs, and the behavior I was seeing was the behavior those inputs produced.

The moment I made that shift — from “why won’t you do this?” to “what in the specification is producing this behavior?” — the problem became solvable.

I went back to the prompt. Not looking for bugs. Looking for interactions.

Three Rules, Zero Bugs, One Crippled System

I found three instructions. Each one was individually correct. None of them contradicted each other. All three had been in the prompt for good reasons:

Instruction 1 told Claude that for certain file types, it should generate the content and let me handle placement. This was a cost optimization for one specific environment — but the instruction didn’t distinguish between environments. It fired everywhere.

Instruction 2 told Claude to confirm which tools were available before starting file work. Reasonable safety check. But it created a reflex of pausing to check permissions before acting.

Instruction 3 listed specific limitations of one operating environment. Individually accurate. But it was a paragraph of “Claude cannot do X” statements sitting in the context window.

No bugs. No contradictions. But stacked together, these three instructions created a cumulative disposition: identify what you can’t do before attempting what you can.

The governance rules I’d written to fix the first problem — Capability Verification and Constraint Reframe — were fighting against a headwind created by my own instructions. I had taught Claude to test its limits while simultaneously teaching it to assume it had more limits than it did.

The human who was teaching the AI to stop saying “I can’t” had accidentally trained it to say “I can’t.”

This Is Not a Story About Prompting

Or rather — it’s not just a story about prompting.

Here’s what I actually learned: prompting is machine governance. Not conversation. Not suggestion. Governance. Every instruction you write changes the equilibrium of a complex system. And the behavior that emerges isn’t in any single instruction — it’s in the interaction between all of them.

This is the fundamental nature of complex systems. You can understand every component. You can verify that every rule is correct in isolation. And the system they produce together can still behave in ways nobody designed for and nobody predicted. Not because something is broken, but because equilibria emerge from interactions, not from parts.

If you’ve worked in organizational management, you’ve seen this. A company writes a travel policy, an approval workflow, and a budget threshold — each sensible on its own. Six months later, a project team can’t get to a client site because the three policies interact to create a two-week approval loop that nobody intended. The rules aren’t wrong. The equilibrium is.

From One Consultant to Every Enterprise

Now scale this up.

I’m one person, configuring one AI, with full visibility into every instruction I’ve written. I caught the problem because I happened to notice the behavior shift, happened to resist the urge to anthropomorphize, and happened to have enough technical intuition to debug the specification rather than blame the system. It still took hours.

What happens when an enterprise deploys hundreds of AI agents, each with specifications written by different teams, governed by policies drafted by committees, and operating in environments where the interaction effects are invisible to any single person?

You get exactly the same failure mode — at scale, with no one positioned to catch it.

This is the gap that my colleague Luciana Ledesma, founder of MeaningStack, writes about in her work on autonomous system governance. The core problem she identifies is what she calls the “Infinite Deviation Space”: when a system has thousands of variables, the space of possible behaviors grows combinatorially. You cannot predict all deviations. You can only detect, interpret, and constrain them in real time.

Her framework draws a critical distinction that maps directly to what I experienced. Humans are good at low-resolution, high-meaning judgment — ethics, intent, purpose. Machines are good at high-resolution, low-meaning monitoring — tracking thousands of variables, detecting drift, maintaining consistency. Governance fails when you ask humans to do the machine’s job, or the machine to do the human’s.

In my case, I was trying to do both. I was writing the specification (the human job — defining intent) and also trying to mentally model how all the instructions would interact at runtime (the machine job — high-resolution monitoring). I succeeded at the first and failed at the second. Of course I did. That’s not a human-scale task.

The MTP Problem

In the world of organizational transformation, leaders talk about Massive Transformative Purpose — a bold, inspiring vision that aligns everyone in the organization toward a shared goal. And it works. For humans.

But here’s what an MTP does not do: it does not align machine behavior to outcomes humans value.

You can have the clearest purpose in the world. You can hire the best people and build the best culture. And if the AI agents operating inside your organization are governed by specifications that produce unintended set-points or equilibria, those agents will faithfully execute behavior that undermines your purpose. Not because they’re misaligned in some philosophical sense. Because the spec produced an equilibrium you didn’t design for.

Purpose aligns humans. Governance architecture aligns machines. You need both.

What I Do Differently Now

Three things changed in how I work with AI after this experience:

I treat the prompt as a living system, not a document. Every edit is an intervention that changes the equilibrium. When I add an instruction, I think about what it interacts with, not just what it says.

I test empirically, not theoretically. I don’t read the spec and predict behavior. I change the spec and observe behavior. When something shifts, I look at interactions before I look at individual rules.

I watch for the anthropomorphization trap. The moment I start attributing motivation to the AI — laziness, stubbornness, carelessness — I know I’ve lost the thread. The system doesn’t have motives. It has an equilibrium. If I don’t like the behavior, the equilibrium is mine to change.

These practices aren’t just for people configuring personal AI assistants. They’re the practices that every organization deploying autonomous systems will need to develop. The question is whether they develop them proactively — by treating agent specification as the governance discipline it actually is — or reactively, after the emergent failures have already arrived.

My failure cost a few hours of lost productivity and a deliverable with unnecessary gaps. The same failure pattern in a financial system, a healthcare workflow, or a security operation costs something else entirely.

The Invitation

If you’re leading an organization that uses or plans to use AI agents, I’d ask you to consider one question:

Who in your organization is responsible for the emergent behavior of your AI systems — not just the individual rules, but the equilibrium those rules produce?

If you don’t have a clear answer, you have the same problem I had. You just haven’t noticed the symptoms yet.

Gary Ralston is a partner with Ann Ralston at Ralston Consulting Inc., where they advise organizations navigating AI adoption, governance, and institutional transformation. The MeaningStack framework referenced in this piece is the work of Luciana Ledesma — learn more at meaningstack.com. Reach Gary at gary@ralstonconsulting.com.