These comments are specific to Current Topics in the Law and Policy of AI, a two-credit Fall 2026 course that I teach at Emory University School of Law. Unlike many law school courses, this course encourages students to use generative AI for research and writing while holding them responsible for accuracy and for producing high-quality, readable work.
I plan to continually revise this document throughout the semester, as each new paper and presentation usually triggers some new observation.
Memos
Some context: students in this course work in teams of two to prepare a 25-minute presentation and a 15–20-page research-based memorandum. Both are organized around five components: an overview of the topic and competing perspectives, technological developments, the legal and regulatory landscape, the academic literature, and the team’s own analysis.
You need to read your memos aloud
If you are using text that came from AI, you should probably stop after each sentence and ask yourself:
- What am I trying to say here?
- Why am I saying it?
- Is this the clearest way to say it?
- What does it leave unexplained?
- Do I even understand this sentence?
When you read your memo aloud, you will begin to notice the latest quirks of AI drafting. The ones that come immediately to mind are excessive use of three-part lists, “not x, but y” framing, and unnecessary or repetitive use of words such as “honestly” and “genuinely.”
Also, when you read your memo aloud, you will begin to appreciate how monotonous the cadence of AI writing can be. Variation is good, but it can’t just be random. Engaging analytical writing often moves between analytical prose and direct, sometimes even idiomatic language, but always in a way that feels intentional.
Be on the lookout for signs of incomplete editing, synthesis, or reflection
Here are some signs:
- When you tell the reader something is recent, but actually it happened two and a half years ago
- When you give the reader a fact or a statistic presented as though it is current and it is clearly not
- When you present a piece of information or a statistic as representative without explaining why it is, might be, or is not
- When you say something once and then basically say the same thing again three sentences later or in a subsequent paragraph
- When you repeat yourself in only slightly different language (okay, that one was a joke)
More generally:
- Check whether every paragraph is doing a distinct job. Redundant paragraphs are just as off-putting to your reader as redundant sentences.
- Beware of over-claiming and overgeneralization. AI writing, and quite frankly, a lot of bad writing by humans, is prone to resting broad generalizations on a narrow base of evidence.
- But you need to also beware the countervailing mistake of detail for the sake of detail. Especially in a background section where you are describing how technology works, the state of the literature, or litigation history, ask yourself, how much of this detail does the reader need to know to support the analysis that’s coming.
- Relatedly, before subjecting the reader to an avalanche of detail, tell them why it’s important and what they are supposed to get out of it.
Be on the lookout for signs of incomplete research and analysis
- When you talk about one academic paper as though it covers the field or represents the field
- When you explain the technology, especially LLMs, in terms that might have sounded reasonable in early 2024 but now seem simplistic—especially when discussing what LLMs supposedly cannot do, or when your argument is effectively “stochastic parrots … therefore”
- When you fail to grapple with the fact that LLMs are deployed as a general-purpose technology
- When you fail to address significant events bearing on your topic that have occurred in the last six months
Moving from the trees to the forest
Using generative AI in research can quickly leave you facing a daunting amount of information, and it can be difficult to take a step back and see the forest for the trees. Here is a method that might help you.
Once you have written a defensible first draft, step back and ask yourself, “How would I summarize what we are trying to say in 60 seconds?” Do not just ask the question. Answer it. Keep working until you have a 60-second summary that makes sense to someone who has not done the research you have done.
Once you have your 60-second summary, look back at your paper and consider how it could be restructured and how its signposting and transitions could be improved.
Quality control
When you think your paper is finished, open a brand-new window in your preferred LLM, with memory off, and ask questions such as these:
- What technical errors and misstatements are contained in this paper?
- Are the facts or the technological environment described in this paper out of date? As a follow-up, ask the model to design a Google (or Bing, I guess) search that could identify recent developments relevant to the topic that you might be missing.
- How complete is the survey of the relevant academic literature presented in this paper? Am I missing anything well-known and important, or anything particularly on point and recent?
- Highlight seven or eight sentences or phrases that, on reflection, sound as though they were written by AI, and explain why. It may be helpful to follow up by asking for seven more, repeating the process until you begin to see diminishing returns.
Presentations
Think about your audience
Your audience has not done the reading you have done. You need to explain the topic to them, taking for granted a high base of general knowledge and intelligence, but without assuming that they know the things you have learned in your research.
Every presentation needs an introduction
It is a good idea to begin your presentation with an introduction. In a team presentation, the introduction should ideally tell the audience who will cover each part. For this course, briefly introduce the topic or controversy, preview the three or four parts of the presentation, and foreshadow your ultimate conclusion.
These presentations are not mystery novels. It is fine to state your thesis at the beginning. In fact, it is more than fine; it is genuinely helpful.
Stay in control of your slide deck
It’s fine to use AI to prepare a slide deck, but you should be in control of the design and architecture of the slide deck.
Too much text
You should be aware that AI-generated slides—at least as of September 2026—almost always contain far too much text.
Why is too much text a problem?
First, from the perspective of effective communication, a slide jam-packed with text will either be unreadable or cause your audience to spend so much time reading that they will not listen to what you are saying. My slides often have no text at all. When they do, the text anchors the essential point rather than providing its full elaboration.
Second, too much text on a slide becomes a crutch and a substitute for having internalized the content. If you have not internalized the content, you can’t give an effective presentation, respond to your audience’s puzzled faces, signs of boredom, etc. And you also can’t stay in control of the timing.
Beware of meaningless iconography
I like using images on slides because they are engaging and, if chosen carefully, can provide an extremely compact representation of a much larger point.
Poorly chosen images plonked down merely for the sake of having an image have the opposite effect. They are the equivalent of someone storming into your conversation, half-drunk and rambling—effectively, an off-topic interjection.
Practice!
Respecting your audience means practicing your presentation.
When I give an important presentation, I spend anywhere from hours to days fine-tuning the delivery. Along the way, I usually customize the slide deck, work on the timing, and trim, edit, and rearrange the material.