Emerging Approaches

What This Page Is For

This page describes shifts in how AI tools work that are likely to matter at a college over the next few years. It is deliberately written about capabilities rather than products, because product names change quarterly and the underlying shifts change much more slowly.

A caution worth stating plainly: most of what is announced does not arrive, and most of what arrives is less transformative than promised. The point of watching this space is not to adopt early. It is to recognize a genuine shift when one happens, and to avoid being pushed into decisions by enthusiasm.

What Is Actually Changing

What it is: Rather than answering from what a model absorbed during training, these tools answer only from documents you provide, and cite which source each claim came from. This method is called Retrieval Augmented Generation, or RAG.

Why it matters: It addresses the single biggest reliability problem. When a tool can only draw on what you gave it and shows you where each statement came from, fabrication drops dramatically and verification becomes fast.

Where it is already useful: Course material and study guides built from assigned readings. Committee preparation from a stack of documents. Policy questions answered against the actual policy. Onboarding material built from existing procedures.

The limitation: it can only tell you what is in your sources. If the documents are out of date or incomplete, so are the answers — and the citations will make wrong information look well-supported.

Gemini Notebook is the Maricopa-licensed example of this approach.

What it is: Tools that carry out multi-step tasks — searching, filling forms, moving data between systems, sending things — rather than handing you text to act on yourself. This is called Agentic AI with agents being created to carry out the task(s) and use the tools.

Why it matters: The value proposition shifts from "helps you draft" to "does the task." That is a substantially larger claim.

Why to be careful: Every reliability problem becomes more serious when the output is an action rather than a suggestion. A wrong paragraph is an editing problem. A wrong action taken in a student information system is an incident. Review before sending becomes review before doing, and the opportunity to catch the error is much narrower.

In a college context: anything that touches student records, financial systems, or communication going out under an institutional name needs review before it is automated, not after. This is squarely the kind of project that triggers security, data-handling, and records requirements — see Choosing the Right Tool.

What it is: Tools that handle more than text — reading a diagram, transcribing and summarizing a recording, describing an image, generating audio. This is called Multimodal AI, and many of the latest models are multimodal.

Where it is genuinely useful now: Transcription and captioning of recorded sessions. Draft alt text for images. Converting a whiteboard photo into structured notes. Making recorded material searchable.

Accessibility both ways: these capabilities can meaningfully improve access — captions, described images, material offered in more than one format. They can also produce output that looks accessible and is not. AI-drafted alt text describes what is visually present, not what the image is doing in context, and auto-captions still mishandle names, technical terms, and accented speech. Both need human review. See Accessibility.

What it is: Smaller models that run on a laptop or on institutional hardware rather than sending data to a vendor. Some of these models are called Small Language Models, or SLM, but not all local models are SLMs. Some can still be LLMs.

Why it is interesting here: It addresses the data question directly. If the model runs on your machine, sensitive information does not leave it — which changes the calculation for a lot of work that currently cannot use AI at all.

The honest state of it: capable models are smaller and faster every year, and the setup is no longer specialist work with applications like Ollama and LM Studio making it easy to download and use the models. But these models remain less capable than the large hosted ones, and running them at institutional scale is an infrastructure commitment, not a download.

Why to watch it: if the reason you cannot use AI for something is that the data cannot leave the district, this is the development that eventually changes that answer.

What it is: Not new applications, but AI appearing inside the existing ones — the LMS, the email client, the document editor, the student information system. 

Why it matters more than it sounds: It removes the decision. Nobody chooses to adopt a feature that simply appears in the sidebar, which means these capabilities reach far more people than any standalone tool, including people who have had no training and never read a page like this one. The Ask Gemini in Chrome falls in this category where anyone using Chrome can use the Gemini AI within their browser without leaving the page they are on.

What this means practically: guidance has to reach people who are not seeking it out, and default settings matter enormously. When a vendor turns on an AI feature that processes institutional data, that is a data-handling question whether or not anyone chose it.

How to Read the Announcements

A few questions that hold up regardless of what is being announced:

  • What does it actually do, stripped of adjectives? If a description survives removing the words "revolutionary," "seamless," and "intelligent," there is something there.
  • What happens to the data? Where does it go, who can see it, how long is it kept, is it used for training.
  • What does it do when it is wrong? Every system is wrong sometimes. The ones worth adopting fail visibly.
  • Who is accountable for the output? If the answer is unclear, it is you.
  • What would we stop doing? A tool that adds capability without removing work is adding work.

If something here looks genuinely relevant to your area, that is worth a pilot rather than an adoption — see Pilots and Experiments for how to run one that produces an answer.