Interacting with AI

Sean Fox

Author Profile
published Sep 28, 2026 3:09pm

Interacting with generative AI systems often feels like interacting with humans. But, of course, they are not human. Here we reflect on how we should interact with them, given how they actually work. We assume the reader is familiar with the ideas in our two Generative AI Basics posts. Note that while we're tackling how to be thoughtful and critical in our interactions with these systems, we're approaching this through the lens of how their internal mechanics can guide that critical use. There are a variety of other critical lenses one could use, especially around the myriad of ethical issues they raise. While we encourage the reader to pursue those as well, we're taking a more specific and narrow focus here. We'll start with a step back and reflect on how interactions with AI systems play out if we simply rely on past experiences and expectations, rather than bringing an appropriate critical lens to bear.

The Weirdness of AI Systems

First, the typical AI chat interface is surprisingly lacking in affordances that help us understand what we're supposed to do with the tool. Across the suite of computer systems we use on a daily basis, we've come to expect each tool to have a specific purpose (a Word processor is for generating documents, an email interface is for sending and receiving messages, etc.). Further, once we dive into a particular tool, it gives us a rich set of clues as to how we're supposed to interact with it. It has menus with a set of predetermined options. It has icons and buttons, and tabs and interface elements that pop out at appropriate moments.  All of these provide clues to (and constraints on) how the tool is meant to be used, and to what end. The AI chat interface provides no hints as to what it should be used for, and the interaction mechanism (just type anything) provides no guidance.  While its open-ended-ness is clearly an asset, it also means we have important decisions to make, even before we start typing, in a way that is not familiar.  At least not familiar in terms of our interactions with other computer systems.

Of course, we have these sorts of open-end interactions in a different context: when interacting with people.  And this is the entire point of the AI system being presented as a chat interface: to mimic that sort of interaction. This is amazingly useful from one perspective, as there are no menu options to learn, no jargon to master, no new conventions to uncover.  But it brings with it some challenges.

We have developed a whole suite of habits, skills and intuitions for engaging in dialog with other humans.  This encompasses everything from the conscious strategies we may employ in asking a probing question, to sub-conscious judgements we make about how to interpret responses.  These are built from a lifetime of experience, are conditioned by culture, and informed by the nature of our relationship with the person we're communicating with.  We bring all of these to the table when we engage with an AI system.  But of course, while these systems mimic human communication, there isn't a real person on the other side of the conversation.  The text we type is not going to be interpreted by a human mind.  It just acts as the input to a next-word prediction machine.  If the text we get back seems compelling and confident, it's not because a strong communicator is sharing expertise. It's because the next-word prediction has been trained on many examples of the cadence, structure and rhetorical moves of compelling communicators sharing confident answers.  If we approach interpreting a dialog with an AI system as if it were a dialog with a human it's likely to lead to problems.

The dissonance between our expectations of how this human-like interaction should play out and how it actually does (given the absence of a human on the other end), is clearly evident when we talk about AI hallucinations.  If we think about how the next-word prediction works there's no particular reason to expect the string of words to always reflect things that are true in the real world. They are just words that are plausible given the full set of other words that it's been trained on.  There is no part of that process that explicitly addresses the veracity of the output.  Mix in the fact that we're not even consistently picking the words the model thinks are most likely, we're just sampling from the distribution the LLM gave us, and we should expect things to go off the rails now and then.  But this runs counter to our models for judging the quality of information we get from humans.

Other humans certainly lie, dissemble, get confused, or are simply off the mark. But the ways we navigate those situations, like making judgments about whether the person is actually an expert on a subject, thinking about their possible motivations for what they are saying given the context of the conversation, and evaluating how confidently they are expressing themselves, aren't relevant to responses from AI's.  AI confidence is typically unrelated to the reliability of the answer. They have no intrinsic motivations. They come to every conversation with only the context you give them (plus their generic training), and their expertise is diffuse, with different contours than human expertise.

Making things worse, the confidence conveyed by AI responses often plays off our own expectations about how reliable, accurate and consistent computer-generated information has been in the past, in non-AI systems. This leads us to be shocked and dismayed when AI responses contain misleading information or outright falsehoods.  But this shock and dismay says more about our own expectations of how human-like dialog, or human-computer interactions, ought to play out than anything actually surprising in terms of how the system itself behaves.  This all points to the need for us to develop new expectations, better matched to how interactions with AI systems are likely to play out.

New Critical Habits

What we offer here are a few strategies appropriate for chatbot-type interactions.  This framework gestures at the sort of instincts and habits we'll all need to develop.  We present it as a first step to jump-start your thinking and engender further discussion.

Roles

Before you engage in an AI dialog, it's worth pausing and making a decision about what role both you and the AI will play in the conversation.  By default, because of its late-stage training on example AI conversations, the AI will respond as a helpful expert. This leaves you in the role of curious question asker. In some cases that may be fine, but there are other options. You can simply ask the AI to behave differently, and it will oblige by predicting words that mimic that behavior.  You can use this to carry out explicit role-playing scenarios, or to more subtly redirect the conversation. Here are some examples:

  • Discrete tasks: Give me a synonym for audacious
    Lean on the LLM's facility with language and ability to pull information from its own training or the web.
  • Dialog: what might be causing this error message....
    Here we expect a back and forth and are perhaps using the AI to bridge the "I'm not sure how to even search for this information myself" gap.
  • Agenda Setting: Suggest how I can improve my bread baking skills
    Even mediocre advice from an AI can often help jump start your own thinking or framing of a problem.
  • Critical Feedback: Point out the deep writing and conceptual issues in this essay...
    Sometimes getting harsh feedback, especially if you feel empowered to argue with or even ignore it (it's just an AI after all) can nicely complement feedback from humans you trust.
  • Tutor: Help me learn about petrogenesis.  Start by quizzing me on my knowledge and follow up with appropriate guidance and questions. 
    This sort of prompt is the core of most AI tutor systems. Just add some information about good pedagogic practices, and a document or two full of content knowledge to the context, and you're 99% of the way to having built your own tutoring system.
  • Just the bit that you don't want your brain to work on: Hmm, maybe it's better if I think about this on my own first...
    A task passed to the AI is one your brain doesn't need to deal with. But sometimes engaging your brain is exactly the move that's needed. 

No role at all (for AI)?

If you're brainstorming to surface expertise and perspectives buried in your own subconscious, then asking the AI for ideas first is likely to tilt you toward ideas that don't take advantage of your unique perspective.  If you're trying to learn, then you need the productive friction of wrestling with information in your own head. Just letting the AI expose you to correct answers isn't sufficient. Which points to questions at the core of AI-informed teaching. How do we help our students develop (1) the skill to recognize moments where they should be putting themselves through that productive friction, (2) the motivation to make that hard choice (thinking is work), and (3) the knowledge and experience of how to engage in that hard work?

The open-endedness of the chat interface can be seen as an invitation to be creative. But you have to be intentional to invoke that creativity and direct the tool to behave in ways that are genuinely useful to you in the moment.  Often, this will warrant some serious reflection on what you both gain and lose by having the AI take on that role.  This is a new habit with a significant metacognitive dimension.  When sitting down with an AI tool, and throughout your interaction with it, you should be reflecting on how the roles in the interaction are impacting your thinking.  This includes your cognitive processes: am I short-circuiting the learning I want to be doing, or just automating a routine task I'm already expert at (or don't care to learn more about).  But also, your affective state: am I feeling confident about how this is going just because the AI is being supportive and positive, and allowing that to substitute for my own judgment?  And you should be actively shifting those roles to improve your experience.

Context/What D'ya Know

Often the most important step in having a really productive conversation with another person is finding the right person to have that conversation with: the person with the expertise and perspective that matches the situation.  With AI tools it's a little different.  You can certainly float between tools trying to find one that is useful for a particular need.  Every AI model was trained a little differently and so its weights predict different next words.  But the biggest lever you have to improve the quality of those predictions, making the output more useful in your particular situation, is to manipulate what's in the context.  This includes both the original question or prompt you feed to the system, but also any additional text that gets into the stream either because you pasted it in, uploaded or gave access to relevant files, or caused the system to use its tools to pull in relevant information (e.g. by searching the web).  So, when interacting with any AI you want to be continuously asking yourself: what's in the context influencing the responses right now and how might I change that to good effect?  Here are some considerations:

  • Can I directly add in relevant information?   Copying and pasting text works well, as does uploading files directly into the chat interface.  It's possible to give most tools more blanket access to your local files, but that comes with risks, and often using your own judgment about what particular files might be useful to the conversation leads to better results.
  • Do I need to explicitly ask the system to use a tool, or provide guidance on how that tool use should proceed?  Most of the current systems will invoke tools to search the web when they seem (to the logic of the LLM)  relevant.  But you can do better by providing guidance on how it should focus its searches, including what sort of sources and search terms are relevant.  Also, keep in mind that many websites are now (attempting to) block access by AI systems.  So the AI harness may not be able to even reach the relevant web content.
  • Is it time to start with a fresh context?  Next-token prediction is influenced by the entire conversation so far (it's all in the context).  Sometimes the earlier parts of the conversation are more of a distraction than an aid.  If you've had an extended back and forth and the focus of the conversation has meandered, you're often better off just starting a new conversation.  If the topic has changed, then start a new conversation.  If the AI seems to be focused on an idea or direction that you've already explored and moved on from, a fresh conversation is the quickest way forward.
  • Should I be extracting information from the current conversation for later use?  Context is portable. It's just a stream of tokens. A key move in working with AI systems is to ask them to make a short summary of the important pieces of ongoing conversation. You can be specific about summary length and what parts are important to you, or leave it to the system's discretion.  Saved to a local file on your computer, you can then copy that summary text into the context of a fresh conversation (even with a different AI tool!) and carry on the conversation. Typically, this is done by having the AI produce a 'Markdown' file which has a .md extension.  This is a very simple text format that you can freely edit yourself with any text editor. It's become the de facto format for this use because AI tools can read and write the format reliably.  Expert AI users often have a whole collection of Markdown files where they've captured key instructions they want to reuse (by just copying them into the context).  And many AI system features (skills, memory, to do lists for sub-agents...) are really just Markdown files that get brought into the context to save the work of reinventing and retyping.
  • What 'thinking' text and tool output is likely in the context, and how does that relate to the responses I'm seeing?  While most commercial AI's hide thinking and tool call text, it plays a huge role in determining the predicted words that you're allowed to see.  So it's worth a little effort to imagine what might be going on ("what sorts of websites is it pulling information from? what common knowledge is it surfacing from its training data?")  If, on reflection, you can't imagine how a stream of thinking words or tool output could be leading to the output you're seeing ("how can it know that?") then it's time to put on your skeptical hat and start thinking defensively.

Playing Defense

We know that AI systems we use will make mistakes in ways that are hard to anticipate, will be hard to identify (because they are unlike the mistakes we're used to from humans or other computer systems), and will likely be cloaked in plausibility. So we need to go into any AI interaction with a strong defensive plan already in place. When, not if, the AI gets it wrong, how will you ensure it doesn't cause problems?  Here are some possible strategies:

  • Use AI output as just food for thought.  "What are good synonyms for outrageous" and "give me feedback on this essay" will both result in responses that you'll need to use your own judgment to put into effect.You'll need to decide if you like the offered synonym and whether to adjust the essay given the feedback. The AI's work doesn't 'do' anything unless you decide to make it so. Your judgment is the defense. If you have the relevant expertise to make good judgments on the topic, then you're in a strong position. Remember, you'll need to ignore the artificial confidence and sycophancy in the AI response.
  • Look for external confirmation.  An AI suggested reading list might be useful if you plan to look up and read the items yourself (and so confirm their existence).  But handing it to someone else (especially your students) without first confirming the readings actually exist is a bad idea.
  • Ask an AI in a different way. Often, you can get at question from several directions. Ask 3 different ways and see if you get a consistent answer. Just be sure to start each in a fresh conversation, otherwise the previous answer, lurking in the context, will influence the later responses.  Even just feeding the previous response into a fresh conversation and asking "what is wrong with this response" can be surprisingly helpful. This obviously isn't as reliable as consulting a true expert, but can be a first check to quickly surface bad responses.
  • Use the AI to generate code that does the actual work.  "Sort these 500 names alphabetically"  is a task that might go awry if the AI relies only on its next word prediction.  But ask the AI to write a program to do the task, and you've got something that is verifiable, repeatable and much more likely to work. This can be combined with the previous approach: "write 3 different programs to verify the output of the program you just wrote in different ways" and you're in good defensive shape.  As humans, we often skip these sorts of verification steps ("did you check your work before you turn it in?"), but they are a great use for AI systems which don't mind boring tasks.
  • Don't ask for too much. AI systems are poor at deciding whether a given task is something they can actually complete successfully.  The further you push them, the more likely you'll end up with a plausible pile of falsehoods.  Knowing what is 'too much' is complicated by the 'jagged frontier' of AI capability.  This refers to the mismatch between what is hard and easy for humans, and what is hard and easy for AI.  You can't use the human difficulty scale to predict how successful an AI will be at a given task.  So you'll need to build up a new instinct for judging whether a task is AI appropriate, and keep your AI use conservatively on the right side of that jagged frontier. 

If you find your fingers on the keyboard about to invoke an AI system and can't articulate the defensive strategy(ies) you are using, then it's time to pause and reconsider.

Just a Starting Point

The new critical habits we need to develop will doubtlessly have to evolve as the tools themselves evolve.  Showing good judgment in interacting with a chatbot looks different than having the skills to anticipate the possibilities and engage wisely with long-running agents. Hopefully, this set of suggestions puts you on the road to developing your own set of critical habits around your use of AI. These habits should cause you to pause and reflect on how to engage in ways that are productive and don't simply mimic the skills you use with actual humans.  At the same time, you'll want to work to keep these habits in their place. While treating an AI like a person can lead to suboptimal results, it's not as problematic as the converse. None of us wants to find ourselves in a world where we're applying our newly developed AI-appropriate communication styles to one another.

References and Inspiration

While many folks are talking about the need to intentionally frame our AI interactions, my thinking was particularly inspired by this post and a key article it references:

 

 

 

 



Comment? Start the discussion about Interacting with AI