So You've Become the AI Person: Ten Things Nobody Warned You About

Consider this a public service announcement.

If you're right now on the edge of becoming your institution's person for AI, and if a manager has looked at you with the expression of someone who has discovered a volunteer and has no intention of first explaining the job, then I want to help you. Not by making it sound attractive, but by making it sound accurate.

Officially, I am the AI literacy trainer at my institution. In reality, that position arose because my colleagues asked for help with issues related to artificial intelligence, rather than turning to Google. As a result, they expressed either great confidence in me or a strong preference for interacting with me rather than with an electronic search engine. There is no official title for this work, but my appointment as digital literacy trainer at MVIT represents a well-defined institutional position reflected in a clearly delineated place on the organisational chart.

The development of my AI training materials was entirely unplanned and arose solely because they were needed. No one else was providing this training, and I was close enough to the problem to be assigned to develop and deliver these materials. This is a typical mode of development for major roles in education. It is not based on strategic planning for personnel needs, but reflects instead the value of proximity to and willingness to respond to problems before they are well defined.

What follows is a field manual that did not previously exist as I adapted to this position. It describes 10 experiences that will provide a complete basis for your future professional development. Information is presented in a format readily applicable to teaching, and with the warmth of a person who has successfully emerged from a difficult situation.

Explaining how AI works requires explaining mathematics. Not vaguely pointing to mathematics and calling the result "algorithms," while hoping that the audience will fill in the gaps with its own assumptions about what algorithms are. But actual mathematics, vectors, matrices, attention weights, derivatives governing how a model modifies its parameters during training in a notation appropriate for individuals who have earned the right to use these symbols by spending seven years in training.

You have not learned this mathematics, and you know it. Your audience does not know it and thereby provides the entire basis for your professional credibility. As a result, you will learn sufficient mathematics to communicate with confidence, but not enough to be subjected to cross-examination.

You will understand the intuition for backpropagation, but will not be able to perform backpropagation. You will describe gradient descent in terms of a ball rolling downhill on a landscape of loss functions, and will do so with sufficient accuracy to be clinically useful and with enough imprecision to survive the first follow-up question. Finally, you will learn to describe calculus with the confidence, selectivity of detail and quiet hope that no member of your audience has actually lived in the country, that a travel writer conveys after having visited a country for the first time.

The greatest risk arises from increasing depth of understanding. You will receive two types of questions: one that seeks clarification and one that increases depth of understanding. The question for clarification is, "Does the model adjust its weights to minimise error?" The question for increased depth of understanding is, "Can you explain the application of the chain rule in that situation?" The response to this latter question will be, "Of course, but let us return to the practical aspects of this situation," with the serenity of an individual retreating from the edge of a cliff and pretending instead that he had simply decided to stop walking.

Ultimately, you will become very skilled at this task. The resulting change in strategy will appear, to an untrained observer, identical to a change in direction. This represents a professional level of expertise that deserves to be highly valued.

The worst part of impostor syndrome in this particular field is that it is in large part correct. Expertise is a moving target, and our field changes so rapidly that the person who read everything last month is already behind on Tuesday. This should equalise the playing field, but in reality confirms that all of us are partial impostors, and that the most successful among us are merely better at masking the partial. You will receive compliments.

Someone will refer to you as an expert in front of an audience, preferably at a conference and preferably with an amplifying microphone. You will respond with the poise of a poker player who has just been dealt three aces and desperately hopes that no one will ask for a show of cards. Internally, however, you will be having a completely different conversation. "Expert?" you will be asking, in the tone of someone who is about to reject a gift that he will immediately throw away upon returning to his private space.

"Expert in what? In the things that you described before anyone had an opportunity to evaluate your work? In capabilities that were accurate only on Thursday? In features that you demonstrated last week but for which Microsoft quietly discontinued development over the weekend without informing anyone, including itself?"

The most successful among us simply are better at concealing the partial. Our experience with this situation is strong enough to support the following conclusion: Expertise is simply being further along the same uncertain path as all others. This works well and, in our view, represents an excellent example of a learning experience that remains profoundly positive even at 2:00 on a Tuesday morning when one suddenly recognises an entire category of model architecture for which one has been mispronouncing words for the past six months and finally mispronouncing those words into an archived presentation that was viewed by 40 people. With this, the experience of being further along the same uncertain path than everyone else abruptly collapses, and we are left with the specific awareness of someone who knows both what he does not know and that this is an enormous amount of information.

Finally, it all works out. And then we hear of a new development, and the entire cycle begins again.

AI is transformative, exciting and full of real promise for education and productivity. It is also changing labour markets, perpetuating biases, concentrating control in the hands of a few companies with inadequate oversight and operating at a pace that exceeds all efforts to control it. Both of these facts are true simultaneously, are of importance and are not mutually exclusive. Your task is to convey both simultaneously to an audience that would strongly prefer a simple answer.

It will not receive one. Neither will you. The person who provides you with a simple answer is either not making sufficient effort or is engaged in commercial activity. What you will learn to say instead is, "It depends on how we choose to apply it." What the audience hears is, "She also doesn't know." Actually, what you mean is much closer to, "The technology represents a true potential for dual applications requiring continued attention to institutional and policy aspects." This is a technically accurate statement, but one that conveys the impact of a wet paper bag.

The humour associated with maintaining your balance on this tightrope comes from the fact that you will periodically be thrown to one extreme and overcorrected to the opposite. For example, you will deliver an exciting presentation on AI and access to education for students with disabilities and return home feeling like a visionary. Subsequently, you will read about errors of hallucination in a medical application of AI and spend the following day in a private crisis of concern about whether you had been irresponsible. Finally, you will resume a cautious optimism in time for a very well-earned weekend.

With the release of a new model, the entire cycle will begin anew. And you will receive no day off. Neither will the tightrope ever be granted a period of annual leave.

You wrote it. All of it. You agonised over word choices, restructured your argument four times, revised your introduction at 11 p.m. because it still wasn't quite right and cared deeply and specifically about the difference between a good sentence and one that is merely adequate. Then you turned it over to someone who read the first paragraph and said, "This is clearly the work of a chatbot." That is the particular tax imposed on those who work in AI literacy.

The better your writing, the greater your potential to arouse suspicion about your authorship. As a result, you have entered a profession in which demonstration of your skill ultimately undermines your claim to authorship. The only way to prove that you were indeed the author of the work is to produce a worse product, which is neither professionally acceptable nor desirable. In part, this reflects professional pride. In part, it reflects the fact that the performance of the AI in producing inferior products is far superior to your capacity to produce such work.

Finally, you will never be able to say, but would very much like to be able to say on every occasion: "I know exactly what poor-quality work produced by an AI looks like. I teach a course on the subject. Because of that, this text obviously does not resemble AI-generated writing. Rather, it reflects the time and effort I spent writing the second paragraph for 40 minutes -- a process that an AI does not attempt to perform because it pays no attention to the second paragraph and shows no interest in anything else." Instead, you will merely smile and say that you occasionally use the AI as a thinking partner. This represents both an acknowledgement of truth and a form of surrender to the situation.

The resulting degree of irony is both essential to and supportive of the entire experience. In effect, you live inside the irony.

Because you work in AI and thus have access to tools that make everyone ten times more productive, you are expected now to deliver ten times as much output. The fact that you devote approximately 40% of your professional energy to keeping current with the field, another 30% to obtaining access to tools for which your institution has not properly licensed, and the remaining 30% to actually building things is absent from your presentation of productivity. The arithmetic you present is optimistic in the same way that all arithmetic is optimistic when used to justify a purchase.

No one says out loud that AI increases the productivity of productive people, but does not extend the length of the work day, accelerate the pace of IT procurement or complete documentation for purposes of compliance, attend meetings about the meeting or organise or navigate the structure of a SharePoint folder reorganised by committee during your attendance at the meeting. Finally, it does not resolve your confusion in a manner that permits retrieval of information by any other individual.

In fact, AI has automated some tasks that previously required substantial effort. As a result, we now have additional time to apply to other expectations, including the expectation that we will train all other users of the tools that we have obtained quietly and at considerable expense. Ultimately, this experience with AI is identical to what we experienced previously. In addition, we now conduct training sessions for public servants, educators, and owners of small businesses throughout the Territory on a second Thursday of every month. These sessions represent a belief that experience with "attending an AI session" is equivalent to experience with "competence with AI." However, this is not true. You know it. But you cannot say it out loud.

Instead, you simply smile and confirm the booking.

There is a degree of technical detail appropriate to AI literacy training that is substantially less than you would expect.

Tell someone "neural network" and you will hear that you are unnecessarily complicating things. Describe "inference" and watch as an appropriately precise sheen of politeness sweeps across the room, conveying an impression of mild alarm. Mention "attention mechanism" and you might as well stand up and declare that the remainder of the session will be conducted in a dead language.

Finally, ask whether it is really necessary to understand the difference between training data and inference, and you will be confronted with the question of whether you really need to learn "how to write an email."

The appropriate vocabulary for AI literacy training is "It's kind of like," "Imagine if" and "Think of it as an extremely sophisticated version of automatic completion." That represents the entire lexicon. After 40 hours of experience with the actual mechanics of AI, you will ultimately describe the entire experience as equivalent to having a golden retriever that has read everything on the Internet and is doing its best to predict what word will come next. This level of description is sufficient to convey utility, but inadequate to avoid eliciting occasional quiet discomfort with each use of the term.

Eventually, you will receive complete validation in a completely silent manner. Three months later, a colleague will casually inform you that AI is "basically an extremely intelligent system of automatic completion trained on all of the information available on the Internet." With this experience of complete satisfaction with the accuracy with which your original ideas have been transmitted to others, you will achieve a major goal of the profession.

Somewhere near your third month in the position, the algorithm recognised you as a type of content, not as a person. It has continued to report on your responses with the unrelenting enthusiasm of a research assistant who has received one brief and interpreted it with extraordinary breadth. Your subsequent recommendations include: 14 videos describing the ways in which artificial intelligence will destroy higher education (of which 6 are identical videos with different thumbnails), nine recordings of webinars from institutions that exemplify the various approaches to the crisis that you have now described with the precision of a naturalist documenting subspecies within an ecosystem that is rapidly disappearing, seven demonstrations of features of Microsoft Copilot that are no longer available in the interface that you see, and one video on the use of blockchain in education that the algorithm apparently included as an expression of historical irony or of potential danger.

As a result, you have developed an emergent system of classification. You can identify the type of content on the basis of the thumbnail alone. The visionary with a model that contains arrows of transformation.

The cautious reformer who believes that everything will be fine if we simply redesign our methods of assessment, a task requiring about one career and three approvals by committees of faculty. The vendor with a product and data that would be unacceptable to a first-year student of research methods. The expert who has acquired an understanding of critical thinking about artificial intelligence and provides you with a complete description of the experience, whether or not you had previously requested it.

Finally, and with the greatest degree of reliability, the panel discussion in which four individuals who agree almost completely on all issues engage in lengthy discussions about them and elicit an additional 10 minutes of appreciation for "such a rich discussion."

You no longer watch most of these activities. Instead, you read the comments, where all of the real activity occurs. This commentary is prepared by individuals who are so fatigued by the experience of having exactly the same discussion within a setting of institutional organisation for 2 years that they describe the experience in detail and provide a complete description of the content of the discussion to an audience for which there is absolutely no possibility of future contact.

The algorithm is only half the story. The remainder appears in your mailbox and exhibits the characteristics of a tidal bore.

Because you are now the AI person, all your colleagues forward you things. Every colleague, with remarkable consistency. Each message conveys the genuine belief that you are being offered a useful service, and indeed represents an important contribution to your own experience. It is akin to the service provided to oceanographers by the person who has just learned that water is wet.

The subject line always reflects some variation of "I thought you might find this interesting," "Just in case you have not seen this yet" or our profoundly noncommittal "FYI." The attachment, hyperlink or screen shot reflects information that appeared on the news feed of the person who has not yet experienced complete colonisation by AI of their algorithm in the way you have experienced it.

To them, this is a novel experience; to you, it is just another Tuesday.

The resulting messages fall into three predictable categories. First, we receive descriptions of new wrapper applications that are presented as revolutionary, and occasionally represent an interesting new development. However, in effect they represent the same underlying model that you use every day, merely dressed in different colours and presented as novel.

Second, we experience dramatic increases in productivity as our workflow becomes a completely frictionless environment of AI application. Most of these applications are available at no cost for a 14-day period, and represent a price substantially in excess of your monthly rent for the full-time paid version.

Finally, and most importantly, we receive fully AI-enhanced solutions to problems that previously had not existed. For example, we learn of an application that produces full 3-D representations of talking pet mascots. With appropriate stimulation, these mascots function as a companion for a child working on arithmetic homework. You receive this message from your colleague, who asks, "Have you seen this?" Would we use it in training? You ponder the question for about three seconds and decide that deploying a talking AI dog to lead adults through your AI workshop content would generate more questions than it would answer. You respond with the diplomatic grace you have developed over eight months in such situations.

Then there are the incidents of cybersecurity concern. As a reminder, you are a person with expertise in cybersecurity. There are thousands of major incidents of AI-related security concern occurring at an accelerating rate for all sectors of activity. These events are not trivial, nor are they of purely theoretical interest. Rather, they reflect a rapidly expanding array of threat vectors for which the industry is having substantial difficulty in achieving a technological advantage. But what actually reaches your inbox each week is a single, striking example of retail-quality experience with which you previously had no familiarity. This is a story of a rogue AI agent whose human was a casual employee at a McDonald's restaurant who obtained access to the administrative login for the order station and used this to generate images of hypothetical new flavours of McFlurry. And with the earnest comment that this was "a bit concerning from a cybersecurity perspective...and I thought you should see this."

You thank them. You do this every time. You have developed, through sheer repetition, a response that is warm, acknowledges the sender, and does not in any way convey that you have a folder - an actual folder - labelled "Inbox Articles Others Think I Haven't Read," which is currently holding 847 items and growing at approximately twelve per week. You are grateful for the community it represents, even as it is slowly burying you. The FYI tsunami is, at its heart, evidence that people care. That they're thinking of you. That the topic is landing in the broader consciousness. This is what you wanted... This is what you wanted... This is what you wanted.

Microsoft Copilot is not a tool. It is a live performance on the precariousness of professional credibility for which you are always the audience, sometimes the stage and once the person insisting that the smoke is an expected part of the performance in a burning theatre.

Otherwise, Copilot is simply unreliable, inconsistent, occasionally impressive, always surprising and continuously changing in ways never disclosed to you. During training, it becomes something altogether different. It appears to select precisely the moment of maximum observation to provide an interface that has never been seen before.

The experience of live training with Copilot is as follows. You have prepared extensively. You have tested every step. You have copies of all screen images and of a contingency for the contingency. With the calm assurance of a surgeon who has performed the operation many times, you lead your group through step 1: "Can you explain how I convert a formula to an absolute reference?" This is such a fundamental operation that Excel has been providing the response with use of the F4 key since shortly after the fall of the Berlin Wall. Whew! It works. You exhale, and transfer the actual syntax. Step 2 is successful. For a brief and dangerous interval, you experience the sensation of control over events. This represents an indication that the forces of the universe are favourable to you at this time.

But not today.

Step 3, which was successful in practice four times, tested in the morning and again only minutes before the session because something seemed to be wrong that could not be identified, resulted in this response: "I am sorry, but I am unable to assist with that at this time." There was no additional information, no error code and no indication of whether "at this time" referred to the next 30 minutes or to the remainder of the current financial year. Only a politely unrelenting refusal was conveyed with the serenity of a system that had decided that the session was over. And then -- and this is where your optics of self-confidence are truly tested -- you attempt a follow-up. A simple one, based on the very spreadsheet data to which you had been painstakingly walking Copilot for the previous 10 minutes. On which it had generated a summary, drawn attention to details and quoted column names with apparent confidence only 30 seconds earlier. And then Copilot locks up; the screen goes blank. This is its expression of profound confusion -- of complete unfamiliarity with the spreadsheet you had just used to teach Copilot so much about spreadsheet operations. All the column names it had just reported to you were meaningless. And all of the painstakingly developed context of the preceding 10 minutes disappeared without ceremony.

Meanwhile one participant has raised their hand to note that their version doesn't seem to have the menu item you just pointed at. Another has it but it appears to be doing something different with the cheerful autonomy of a system that has decided to be helpful on its own terms, in its own time, with no regard for the pedagogical moment currently in progress. The thing that does the actually useful thing you want is now located in a completely different position than where Microsoft's own documentation says it lives, because Microsoft updated the documentation on Tuesday but updated the product on Monday and nobody in those two teams spoke before publishing.

The room will nod while also watching you with the careful attention of people who are now genuinely uncertain whether this is supposed to be happening, which is a question you are also privately asking, but from inside it, which is worse. You continue. You do not look at the part of your brain that is silently screaming. You have learned not to look there during working hours.

You count the tally at 3 a.m. It is not a choice, but a condition. The tally appears involuntarily with the fidelity of an appointment for which neither you nor anyone else made a reservation, and from which there was no possibility of withdrawal thereafter.

Your current inventory: The description of GPT-4 as the most powerful model available is now obsolete more frequently and updated with decreasing enthusiasm. Information about software for detection of artificial intelligence remains accurate in the sense that it continues to be incorrect, but no longer describes commercially available products. Comparison of two AI-based systems of word processing led to acquisition and conversion of one system to provide additional features and substantially reduced costs, whereas the other system was reoriented for use by enterprises and eliminated all options for free access.

As a result, you learned about these developments only in the manner available at 3 a.m., i.e., by attempting to demonstrate them to an audience.

Results of your analysis of product costs were of a historical nature, and those obtained for your system of image generation reflected conditions of extreme antiquity. The assertion that "AI is developing rapidly" was completely correct and represented an extraordinary degree of understatement.

The length of the 3 a.m. tally exceeded 40 minutes and prevented subsequent return to sleep. Use of your telephone and initiation of records of events constituted inappropriate activity at this hour, but were the only effective methods for eliminating the experience of the 3 a.m. tally.

Finally, we pose a question appropriate to every course of training, workshop, lecture, meeting of more than four minutes of discussion of AI and with genuine concern for the experience of persons who have devoted years to development of a craft and deserve an honest response:

Is this going to take my job?

The honest answer, based on sound evidence and on intellectual integrity and not on commercial interest, is that some jobs will undergo major changes, some will disappear and new occupations will emerge that previously did not exist. The pace of change will be uneven and not uniformly beneficial. Finally, anyone who confidently describes the ultimate outcome is simply expressing opinion as though it were prediction.

Those who claim that all jobs are secure provide a false sense of security, and those who maintain that all jobs are lost unnecessarily increase feelings of panic for a market that already exists.

This interpretation is entirely appropriate and represents the most accurate interpretation of available data. However, it is of little practical value to a group of vocational educators throughout the Territory who honestly question whether their experience with obsolete skills will ultimately result in a future.

They need certainty in whatever direction. What we can provide is the most accurate description possible on the basis of available evidence and clearly labelled as incomplete and preliminary. This is delivered with the compassion of persons who genuinely care for those in the room and with the candour necessary to avoid providing false reassurance for the convenience of the next 40 minutes. It doesn't land with the satisfying thud of a definitive answer. It doesn't produce the relief of certainty or the clean anger of a clear threat. It produces, instead, something quieter: the sense that they're being levelled with by someone who is in the same uncertain situation, just slightly further along the path, looking back and saying honestly, that the view from here is complicated but navigable.