AI and Assessments

Why Things Need to Change

Generative AI (GenAI) can now do or assist with many real-world skills that college prepares students for, which means it’s equally proficient at producing the kind of tasks we’ve traditionally used to measure student learning. The newest models can rapidly produce passable essays, summaries, discussions, case studies, programming, and work in sciences and math, and this creates a real tension for courses that have relied on these kinds of assessments. 

This has created an uncomfortable situation for students eager to achieve top grades (or who have discovered an effective shortcut to saving time and work), and for faculty who might struggle with determining who (or what) has produced a submission and whether these submissions truly demonstrate any learning or achievement of course outcomes.

Relying on detective work and academic integrity policing has not been effective. Instead, we encourage faculty to rethink assessment design to reflect authentic expectations around the use of AI, and in a way that faculty observe concrete evidence of authentic student learning.

Setting Expectations Early

Providing clarity about expectations by making a course-level statement is a necessary first step. See Communicating AI Permissions to determine which of the three broad “AI lanes” are appropriate to your course’s outcomes. With a blanket statement made for the course, it will usually be necessary to provide more specific direction at the assignment level, e.g., “What do you mean that students are permitted to have partial use of AI when completing this project?” Setting clear expectations at the course and assignment level will be the two main AI-related requirements in Fanshawe’s Online Learning and Educational Technology (A115) policy.

Keep in mind that setting a clear course policy is a necessary first step, but this discursive (declared) change is not enough if it relies on unenforceable student compliance. Durable solutions require structural change to how assessments are designed, delivered, and assessed. The following sections provide guidance on designing assignments that work within the overall level you’ve determined for your course.
 

Lane 1: Limited. Students Produce without AI. 

Some outcomes still require students to demonstrate independent learning. This may be for concepts that need to be internalized for instant recall, or skills that need to be performed quickly and with automaticity, or early abilities that must be mastered in order to learn more complex skills.

In these cases, the challenge for faculty is to design conditions where AI use is impossible or simply not useful. This usually means shifting away from grading unsupervised take-home work and toward environments where performance can be observed. In-class writing, live demonstrations, interviews, and time-constrained tasks all make it far easier to see what a student can actually do.

It can also help to anchor work in specific contexts that AI cannot access, like recent class discussions, personal reflections tied to prior submissions, or highly localized examples. For instance, AI “knows” all about the economics of supply and demand, but it has no idea about the specific examples or activities you used in class last week! The more a task depends on what this student has done in this course, the harder it is to offload to an AI that only understands topics generically.

Limited AI Permissibility Tips:

  • Acceptable AI in this lane tends to ask for coaching instead of doing. Students can think of this as being somewhat free to converse with AI, but never to copy or transcribe AI output into their own work.
  • Provide examples of how students might use AI responsibly without having it directly influence their work. For example, could they upload your slides and ask it to quiz them on your content? Could they ask which of their project ideas might be best? 
  • Consider ways to demonstrate the strengths and weaknesses of AI generation. Having students critique, revise, or improve AI output under supervision can help them better understand its limitations.
  • Designing online evaluations that limit AI can be tricky. Fanshawe has a license for online proctoring software that integrates with FanshaweOnline’s quiz tool (and other features). Training for this product, Respondus Monitor, is available through FanshaweLearns.
  • Try asking AI (e.g., Copilot) for options. If you can describe or upload your evaluation, GenAI tools can often provide multiple ways that you can re-envision or modify assignments so that they work better in an AI-Limited context.
Lane 2. Partial. Students Assisted by AI. 

This is possibly where most courses might land, in an in-between place where student work needs to accurately represent their learning, but where the use of AI might make their work better or more productive. That means shifting attention away from the finished product alone and making sure we can observe the process used to create it.

One of the most effective ways to do this is to break work into stages. Supervised drafts, checkpoints, and revisions all give you points of visibility, while also helping students stay engaged with the task over time. Short reflections, particularly if done while observed, or presented live, or which use video or other grounding elements, can make it more obvious where students have applied their own thinking.

Underlying all of this is a simple idea: a submission only has academic value if its authorship is clear. In the collectible world (from paintings to antique cars), this is referred to as provenance, literally the need for someone who is selling something to show that it truly is what they say it is. The same concept now applies in academia. If you can’t tell with confidence whether a student has completed a task authentically, the assessment has no value. Good design in this space tries to avoid academic integrity problems by ensuring the human creation process is visible.

Partial AI Permissibility Tips:

  • Students need to show enough of their process that the origin and development of their work is evident. Tools like version history, track changes, and draft submissions can support this, but students may want to suggest other ways they can demonstrate their learning as it happens.
  • Consider scaffolding in low-stakes or practice assessments before assigning high-stakes evaluations. This gives an opportunity to correct improper AI use or celebrate proper use so that students are aware of expectations early.
  • Many assignments have key components where students must not use AI. Some projects like this may need to be staged such that critical parts are done under supervision, but other parts (where AI isn’t problematic) can be done at the students’ convenience. 
  • Turnitin can be a useful tool within this lane for some assignment types. Be aware that translators, and paraphrasing and spin tools, will all be detected as high AI use (because they are); assignment descriptions should set clear expectations around this kind of software.
  • Try asking AI (e.g., Copilot) for options. If you can describe or upload your evaluation, GenAI tools can often provide multiple ways that you can re-envision or scaffold assignments so that they work better in an AI-Partial context.
Lane 3: Full. Student Co-works with AI

In some cases, using AI is virtually expected because it mirrors real-world or industry use. Here, the goal is not to restrict use but to evaluate how effectively students work with these tools. That is to say, if the outcome is demonstrated well, this is worth high marks! If the outcome is poor, we may need to evaluate how students have used AI in order to improve their success.

This requires a different kind of visibility. Instead of trying to separate student work from AI output, here we ask students to engage with AI deliberately and transparently. That might include showing the prompts they used, evaluating the quality of responses, or explaining how they revised or rejected initial outputs.

What matters most in this lane is judgment. Students need to demonstrate that they can guide AI appropriately, recognize its limitations, and apply it in ways that make sense for their field. The final product reflects both the tool and the person using it, and our assessments should account for both.

Full AI Permissibility Tips:

  • Most students assume AI output is better than what they could produce, whereas in this lane, AI output quality is highly dependent on users employing strategies to improve it. For example, the first output produced by GenAI should rarely be accepted as is.
  • In earlier program levels, students may benefit if provided with “starter prompts” that will guardrail their AI use so that output is of good quality and matches assignment expectations. As students gain experience, they should be expected to engineer their own prompts.
  • Consider scaffolding in low-stakes or practice assessments before assigning high-stakes evaluations. This gives an opportunity for feedback on how students might use AI more critically and interactively.
  • Assignment design can help correct that by making iteration, critique, and decision-making visible and assessable. If students are expected to engage with AI tools, this sort of critical back and forth process should be described in the assignment description and evaluated on the rubric or scoresheet.
  • Try asking AI (e.g., Copilot) for options. If you can describe or upload your evaluation, GenAI tools can often provide multiple ways that you can ensure that meaningful student oversight of AI production is a built-in requirement in an AI-Full context.

 

If Things Go Wrong

Even with thoughtful design, situations may arise where the authorship of a submission is unclear or expectations around AI use have not been followed. Although the goal of the three lane approaches above is to proactively minimize the likelihood of academic offences occurring (because detection and policing this after the fact is difficult), concerns may still need to be addressed through academic integrity processes. For additional details and support around this process, see AI Academic Offences.

Assessment design in an AI context is evolving quickly. It is complex and often challenging, and it looks different across disciplines. If you’d like to talk through an assignment, discuss ideas, explore options, or get a second perspective, the Centre for Teaching and Learning (CTL) is available for consultation. Contact us anytime at CTL@fanshawec.ca