Skeptics of AI feedback often point to literary analysis as the place where automated tools should fail. Interpretation seems subjective, and a novel like Till We Have Faces, with its layered narrator and ambiguous gods, looks like the kind of text that resists simple evaluation. The skepticism is reasonable, but it overlooks how much of good feedback on literary essays is actually about structure, evidence, and reasoning.

Consider a typical student claim that Orual is simply a victim of the gods. A useful comment points out that the essay quotes only Orual's own complaint and never tests it against what Psyche, the Fox, or Bardia say about the events. That feedback requires no special insight into Lewis's theology; it requires noticing a gap in how the student uses evidence, which is exactly what a well-configured tool can do.

Where AI feedback becomes less reliable is in judging originality or the quality of a genuinely new reading. A student who argues that Orual's veil functions as a moral confession rather than a mask may be doing something interesting that a rubric-driven tool will score only as adequate. This is why teachers need to remain the final readers on high-stakes or unusual papers.

What Good Automated Feedback Looks Like

Good feedback on a Till We Have Faces essay is specific to the paragraph in front of the student. It names the move the student attempted, explains whether it worked, and suggests one concrete revision. A comment like "You quote the lamp scene but do not say what it reveals about Orual's motives" is far more useful than a generic note to add more analysis.

• Points to a specific sentence or paragraph rather than the whole essay
• Connects the comment to a named rubric criterion
• Suggests one revision the student can attempt immediately
• Distinguishes between a weak claim and weak support for a good claim
• Uses plain language a sixteen-year-old can act on without translation

The Limits Teachers Should Plan For

AI tools can misjudge irony, and Till We Have Faces is saturated with it. Orual says she wants only what is best for Psyche while the narrative shows her possessiveness, and a student who captures that gap deserves credit even if the sentence structure is unconventional. Teachers should spot-check papers that engage with irony or ambiguity, since those are the places where mechanical scoring is most likely to undervalue good thinking.

Another limit is context about the particular class. A tool does not know that your students spent a week on the Greek myth source or that you prohibited outside criticism for this assignment. Including those details in the assignment instructions and rubric helps the tool apply the right expectations, and reading a sample of results confirms that it has. Even one short paragraph of assignment context at the top of the rubric can noticeably change how the feedback reads.

Using AI as a First Reader, Not the Final Judge

The healthiest way to think about AI feedback is as a first reader that handles the repeatable work. It can check that each paper has a thesis, uses evidence from both parts of the novel, and organizes paragraphs around claims. That leaves the teacher free to respond to ideas, which is the part of literary teaching most people entered the profession to do.

Students also benefit from receiving comments quickly enough to revise. A paper returned within a day while the lamp scene and Orual's final accusation are still fresh in their minds leads to better second drafts than one returned weeks later. Combining fast, consistent first-pass feedback with targeted teacher commentary gives students more total guidance than either approach could provide alone.

Setting Up a Fair Pilot Before You Commit

Before adopting any tool for a full unit, run a small comparison. Take ten essays you have already graded, run them through the tool using the same rubric, and see where scores and comments diverge from yours. Pay attention to whether disagreements cluster around particular criteria, since that pattern usually points to a rubric problem rather than a tool problem.

Share the results with your department so the decision rests on evidence rather than impressions. If the tool agrees with experienced graders on most criteria and flags its own uncertainty on the rest, it can reasonably take over the first pass. If it does not, the pilot has cost you an afternoon and taught you exactly what to fix. Keeping the comparison spreadsheet also gives you a record to show administrators who ask how the tool was evaluated.