The Refrigerator Principle

I have a favourite example for teaching about AI context windows to people who have never considered AI context windows (i.e., to about 100% of my workshop participants). It involves a refrigerator. As far as I am concerned, it's an excellent example and, ultimately, a beautiful lie.

You open the refrigerator and see leftover rice, half a can of coconut milk, wilted spinach with perhaps three more days of optimism, two eggs and a jar of something you purchased for a recipe you prepared only once in 2022 and have subsequently moved to the back of every refrigerator. You ask the AI: What can I prepare for dinner?

And there is the beauty of the experience: The AI actually is able to help. Based on these specific, tangible ingredients, it will provide information about a rice bowl, scrambled eggs with spinach and a coconut curry that finally will justify the contents of the jar. The success of this example is based on its personal character, low risk and completeness with respect to all information required by an AI (i.e., an explicit inventory of available resources and a request for assistance). There is no ambiguity, no requirement for authentication and no need for access to a computer information system.

Our refrigerator is not connected to a computer information system. That is the problem.

What the Fridge Is Teaching (And What It Isn’t)

But the teaching associated with our refrigerator experience is excellent. It is my intent to convey a perspective on the nature of AI as a highly capable, but ultimately remarkable, assistant that must be provided with information in exactly the same way as a recipe requires the availability of ingredients. It is not possible to request assistance from an AI without simultaneously providing information concerning the resources available to the requesting individual. Our experience with the refrigerator achieves this goal in a manner that will never be achieved by reference to the size of an AI context window and without requiring previous familiarity with the terms "transformer architecture" or with a desire to learn about such matters. What the fridge example does not convey-and what I pass by with the breezy alacrity of a dinner party guest who changes the subject-is that the workplace equivalent of "What's in your fridge?" is a question that has never been asked before. And the response (if anyone had bothered to observe it) is substantially more complex than old rice and two eggs.

The workplace refrigerator includes: a Microsoft Teams channel from 2021 in which a critical decision about the training budget was made as part of a discussion that also generated 47 responses about the holiday party, an incorrect request for access to a different information system, and a GIF. A SharePoint site last reorganised by a staff member who has subsequently left the institution, and for which the logic of organisation was apparent only to that individual (and perhaps not even to him or her). An obsolete database predating the current strategic plan, the previous strategic plan and possibly even the current institutional logo. An e-mail thread forwarded 17 times with an appended file entitled Final_v3 REVISED FINAL USE THIS ONE.docx, which was in fact not the final version but the only one for which there was an active link.

And finally, a complete absence of institutional memory in the minds of those who have access to the information system. All of this is inaccessible to the AI. Until now, I have deliberately avoided describing this situation. Rather, I have used a carefully controlled tempo to build to a point at which I will finally disappoint the audience.

In effect, our kitchen has a locked pantry, cannot be accessed without a ticket to the information services staff, is managed by a head chef who has not yet decided whether cooking is permitted, and includes equipment that was connected to the network by the previous occupant and now requires a password that has never been recorded. I have been asked to prepare a meal, and I am extremely enthusiastic about what I will be able to accomplish. My final choice will be to order pizza.

The Information Landscape Nobody Can Describe

The first indication of difficulty occurs when I ask participants to provide an inventory of information resources available at the workplace. Specifically: What information do you routinely use? Where is this information stored? What information would need to be provided about your work context to enable the AI to provide actual assistance? Silence.

Of a type experienced only by individuals who are, at that moment, becoming aware for the first time that they have never previously considered these issues. Not because they're being difficult, but because no one has ever asked them to audit their information environment. The information environment is simply there. It always has been.

It predates them in many cases, and will continue to exist after they have gone. In short, it is just the ocean in which they sail and the water is SharePoint.

I ask one participant — a thoroughly typical government administrator, because she is — what she’d need the AI to know to help her write a weekly status report. She thinks for a while.

“It would need to know what happened this week,” she says, carefully.

Yes. And where does that information live?

“…In my head.”

And before that?

“In the meeting notes.”

Where are the meeting notes?

A pause of considerable duration. “In a Teams chat. I think. Or it might have been the other Teams. We have two.”

We are now faced with the problem of SharePoint archaeology. It is less an archaeological excavation and more an exercise in dealing with unexploded ordnance. All institutions contain within their digital infrastructure multiple layers of information deposited with different intentions, different conventions for record keeping and vastly different tolerances for the concept of a system for assigning names. Somewhere in the resulting sediment is what you need. It is adjacent to 17 things that you do not need, one of which remains a complete mystery, and to two that should unquestionably be deleted but for which no one has the courage to be responsible because of the potential consequences.

The graveyard of Teams deserves separate consideration. Every organisation has developed channels of great enthusiasm (e.g., the project steering committee Q3 channel, the new staff orientation resources channel and the channel for "We really should do something about our Web site"). As a result, these channels now represent primarily digital monuments to optimism. Important events occurred in these channels, decisions were made and information was exchanged with complete confidence in the permanence of the system. Subsequently, activity in these channels ceased, and important results sank gently to the bottom as sediment.

Above them accumulated 300 messages concerning the Christmas party, results of a vote on whether to shift the staff meeting from Thursday to Friday, and a message from someone in November 2023 simply stating, "Sorry, wrong channel." My AI has access to none of this. It is a text box. An extremely sophisticated text box, capable of achieving feats that appeared to be genuinely miraculous only five years ago, but still a text box. Finally, I have not yet spoken aloud what I now tell you. I convey my message in the manner of a narrator describing events that have developed in complete safety and that deserve to be fully described to those who experience events in the field and cannot obtain the same information.

The fridge is full. The fridge has always been full. The question is whether any of what’s in it is still usable, and whether you can actually get the door open, and whether the light comes on when you do.

“Open Claw with Your Org’s Single Sign-On” (The Dream)

The dream usually surfaces around the forty-minute mark, just as participants have grasped that the AI can help with things they bring to it manually. Someone — there is always someone, and I love them unreservedly for it — says the thing.

“Can it just… do things for me?”

This is, technically speaking, a question about AI agents.

The participant is unaware of this, would not describe it in these terms and is experiencing, with perfect accuracy and without any vocabulary, a world that interfaces with his or her e-mail, reads the calendar, identifies outstanding tasks, retrieves information from a project management system, accesses a library of documents and references appropriate policies to produce a text that is appropriate for an accompanying meeting scheduled at exactly the same time by someone who has no knowledge of, and would have been indifferent to, the first meeting. Here we have a text box of extraordinary quality. It is truly remarkable and has accomplished in the past six months things that I would previously have considered impossible.

Indeed, it remains a text box. At this point, I feel like someone who has been asked to teach a person to drive and has provided only a steering wheel not connected to a car. This is an excellent steering wheel with excellent responsiveness and an intuitively pleasing interface.

In effect, we are doing our best with available equipment, and this represents the unofficial philosophy of every regional training centre in Australia since approximately 1988. "Can it just do things for me?" The experience represents an image of an AI with full access to the single sign-on, permissions and context of the organisation and with overall responsibility for solving problems. That is, it represents an experience with an entity that does not yet exist for the participant, and for whom I alone have any appreciation of the magnitude of the gap between dream and reality.

My task at this moment is to gently escort them back to the chat box without erasing from their faces the look of people who have just seen a glimpse of a better way to do work. The answer to "Can it just do things for me?" is yes eventually, conceptually and, in some deployments, with appropriate governance to achieve it after completion of the security evaluation by the IT department, after compliance with the requirements for the vendor, resolution of data sovereignty issues and implementation of a final, approved version of the policy position communicated to staff by a SharePoint page viewed by two individuals. At that point, the technology will have advanced substantially, and we will resume a version of this discussion in a room with significantly improved air conditioning.

What We Cannot Do

There is a list of things for which AI tools can assist participants, and a list of things for which they cannot. Specifically, there is a policy document stating so, or an IT restriction enforcing it, or a legal grey area for which there has been no resolution and thus the default position remains "not yet," equivalent institutionally to "not this year" and often to "not this decade."

Both lists are expanding. The list of useful applications grows slowly in a manner reminiscent of coral growth: apparent only when compared with a fixed standard and with patience regarding what constitutes progress. The list of restricted applications expands rapidly and for entirely understandable reasons. Each new application conceived by participants illustrates a previously unmodeled risk. Clients' information is inadvertently incorporated into the public interface for chat. The AI is used to prepare communications that should have been subjected to prior review. Finally, on the collective alarm of a Friday afternoon, it is discovered that the AI has reproduced information for which it should not have had access. This results in an incident that triggers investigations, and ultimately leads to the development of new policies and restrictions that further extend the list.

As a consequence, we are experiencing what I have begun to refer to within the confines of my own thoughts as an expanding corridor of opportunity. You can use AI for these, and for these, and for this application at the extreme right of the third column - but not for these, or for these, or for the entire series of applications that represent such an exhilarating experience of technological capability for an enthusiastic participant. Indeed, we have observed repeatedly the sequence of expressions on a participant's face during description of the application for which he or she had been most eager to participate. Finally, we are obliged to convey the results of this experience in a manner appropriate for the mild delivery of unpleasant information at a moderate level of volume. That is, we must inform the participant that this application represents a situation of policy ambiguity for which we are currently providing guidance. The expression goes: hope, then calculation, and finally the peculiar form of professional disillusionment that is too polite to be called devastation but requires a brief time to recover. I have become thoroughly familiar with this brief pause. It is not pleasant.

I return home on the Laterite Highway at seven in the evening, when the sky performs that spectacular demonstration of its seven colours for which no entirely satisfactory explanation exists, and am forced to confront the question I have avoided throughout most of the drive: Am I conducting AI workshops for people who are essentially prohibited from using these technologies?

Hardly. But the disparity between what is permitted and what is actually appropriate remains unacceptable, and those who would benefit most are precisely those for whom the limits to use are most severe.

The Narrowing Window

The people who experience the greatest cognitive load within an organisation are rarely those who make decisions about governance systems. Instead, they are case managers who process 40 files weekly with equipment obsolete when they arrived and never subsequently upgraded. The administrative staff who retain institutional knowledge in the absence of an adequately implemented system of knowledge management. The personnel who manage three incompatible databases, use a referral process requiring printed forms in the year 2026, and recognise that absence of personnel with knowledge of workarounds will immediately stop all operations. These are the individuals who would benefit most from a competent system of artificial intelligence that would interpret their context, accommodate their workload and eliminate 45 minutes of repetitive activity at the end of each already excessively long workday.

Moreover, they represent the situations in which the complexity of the governance process is greatest. For health care, education and government, these are the sectors with the highest requirements for security, the most extensive limitations on use and the greatest justification for both. The discrepancy between what could be achieved with AI and what is currently permitted with regard to use of AI is not a technological limitation. The technology is fully adequate for the task. The limitation is one of governance, policy and degree of trust.

The resulting architecture of legacy information technology is completely inappropriate for integration with an external system of AI. Major efforts of assessment, planning, estimation and scheduling would be required to make it safe enough to attempt integration with AI. Integration would then ultimately be achieved in some organisations for some positions, but at a very slow rate, with substantial variations in degree of success and only after occurrence of incidents that were entirely unacceptable and led to additional restrictions on use. In the meantime, I am teaching people to fish in a river they are not yet permitted to stand next to. I provide them with a steering wheel and with genuine excitement about the things they will be able to accomplish with it.

The car is coming; I believe it is coming. However, at this moment we are experiencing the predominantly mechanical phase of this particular journey, and it is my professional obligation to maximise the utility of the steering wheel and to be completely candid about its limitations.

The Refrigerator, Which Is Still Something

The refrigerator is not empty. It can help you draft that difficult email you've been avoiding for three days, work through a problem that has haunted your thoughts since Thursday, summarise a 40-page document for which you've copied and pasted, generate three options you had not previously considered and provide structure to your quarterly report even when you know exactly what you want to say but have absolutely no idea how to begin. That is real.

It is profoundly useful and justifies a two-hour workshop, the trip to the training room and exposure to uncomfortably cold air conditioning. Finally, it confirms that your colleague is absolutely right in explaining why he already uses ChatGPT and therefore has no need to attend. (But he does need to attend.)

For now, I am back in the refrigerator.

I am inventorying my available resources, learning to apply them with some degree of skill and reasonable confidence. In the process, I am accepting that access to the pantry is a matter of governance and not of technology, and will be provided when appropriate. The locked pantry is real.

The incomplete refrigerator is real. And so is the chief cook who has yet to decide whether cooking is permissible, but who is certainly extremely real. However, my refrigerator provides only a beginning that is well worth having.

Overall, it represents an excellent refrigerator. It is not my dream. My dream is of an agent with single logon and integration with the calendar and with an ability to prepare my quarterly report before I have had time to finish my first cup of coffee.

My dream would make the person carrying 40 files feel as though he were carrying only 38. It is within my reach and sufficiently distant to justify continued professional caution concerning commitment to an estimated time of arrival. Next week, I'll demonstrate to you what you can create with it.