At Packback, we believe AI should strengthen human thinking, not replace it. We know students and educators care deeply about how educational technology is built, how student data is protected, and the environmental cost of artificial intelligence.
When students or educators scrutinize AI, ask who bears the cost, and check whether their data is being used to train commercial models, they are practicing the exact critical thinking higher education aims to cultivate. This guide outlines the technologies we use, why we use them, and the ethical safeguards built across our platform.
1. Not All AI Works the Same Way
When people hear "AI," they often picture large, energy-hungry Large Language Models (LLMs) like ChatGPT, Claude, or Gemini. But that represents only one branch of artificial intelligence.
The vast majority of what students and instructors interact with across Packback—including discussion coaching, Curiosity Scores, automated moderation, and writing mechanics—does not use Generative AI. Instead, these features rely on Rule-Based Systems and Predictive Machine Learning.
Rule-Based Systems
How they work: These systems run on explicit, predetermined logical checks ("if this, then that").
What they do: They check structured, mechanical criteria such as whether a post meets minimum length guidelines, follows citation formatting, or includes required structural elements.
Why it matters: They are completely deterministic and transparent. A rule-based check cannot hallucinate, invent false citations, or fabricate information.
Labeled in the platform as
Predictive Machine Learning
How it works: Rather than creating new text, predictive machine learning analyzes existing student writing by identifying linguistic patterns, syntactic structures, and conceptual depth.
What it does: It functions much like an advanced spam filter or diagnostic spell-checker. It assesses open-ended questioning quality, flags potential community guideline concerns, and provides instant coaching on argument structure.
Why it matters: This technology is discriminative (pattern-matching), not generative (content-creating). It cannot draft essays, originate content, or generate text of its own.
Labeled in the platform as
The Environmental Footprint of Predictive vs. Generative AI
Large generative models carry a well-documented environmental cost: training runs consume substantial electricity, inference at scale adds continuous load, and data centers draw significant water and power from local communities. Students who are concerned about the environmental impact of AI have good reason to be.
That impact is a significant part of why at Packback, we lean on predictive systems wherever they will do the job. When we choose a lighter model to handle scoring and feedback, we are making an engineering decision and an environmental one in the same breath.
Both things can be true: the environmental concerns around generative AI are real, and the footprint of the specific tool a student is using in their course may be much smaller than what they’re picturing.
To provide a sense of scale:
- A full year of Packback’s hosting emits less CO2 than the passengers on a single round-trip commercial flight between Denver and Chicago.
- Packback cut CO2 emissions from its operations by 34% between February 2023 and February 2025.
2. Generative AI for Interactive Coaching & Writing Insights
Beyond predictive tools, a few specific features on our platform use Generative AI to provide conversational coaching and content-aware insights rather than generic advice.
Voluntary Engagement: Engaging with conversational writing coaching is completely optional. Instructors can configure their courses so students can complete their work and earn full credit without touching generative features.
No Public Model Training: Any text processed to generate coaching feedback is used solely to return that immediate recommendation—it is never sent to external AI providers to train their public models.
3. Generative AI for Teach the AI & Argument Lab
Packback also uses commercial Large Language Models for a distinct assignment family: AI Integrated Assignments (such as Teach the AI, Argument Lab & Assignment Builder (Beta).
Rather than using AI to generate student work or relying on post-submission surveillance tools, these activities operate on a "Show Your Work" pedagogy:
Teach the AI: Students act as the instructor. They teach a topic to an AI "learner" that asks questions, demonstrates partial understanding, and introduces intentional misconceptions. The student’s job is to spot errors, address confusion, and explain the material in their own words.
Argument Lab: Students select a position and defend it against pre-set counterarguments presented by the AI, developing critical thinking through structured back-and-forth dialogue.
Specialized AI "Agents" Working Together
Rather than using a single open-ended chatbot, these assignments divide tasks among separate, purpose-built AI agents:
Mapping Agent: Generates concept maps or counterargument frameworks from the instructor's learning objectives, which the instructor can customize before publishing.
Conversation Agent: Engages directly with the student as the curious learner or debate challenger. Interactions are intentionally concise, designed to wrap up in about 15 exchanges to target specific learning outcomes without wasteful compute cycles.
Evaluation Agent: Reviews the completed transcript independently to tag evidence showing where the student demonstrated mastery of the instructor's learning objectives.
Fairness by Design: The AI that talks with a student during the conversation is never the AI that evaluates their performance. Isolating dialogue from evaluation ensures objectivity, makes auditing straightforward, and allows each model to be tested against strict quality benchmarks.
4. Generative AI for Assignment Builder (Beta)
Packback offers an optional generative AI tool for instructors: the Assignment Builder, launched from Create for Me on the Manage Assignments page. It is available for both Discussion Assignments and Writing Assignments (Deep Dives), and is designed to cut down on the prep work of setting up assignments one at a time.
Built from the instructor’s own materials
Instructors provide the builder with their course materials by pasting text (such as a syllabus or a short description of what they want), uploading a PDF or Word document, or both. The builder reviews what the instructor provides and generates a set of assignment previews.
Instructors remain in control
- Optional: Instructors can still build assignments by hand at any time, or combine hand-built and generated assignments in the same course.
- Reviewed before saving: Every preview can be expanded and inspected, and instructors keep only the assignments they select.
- Hidden until published: Instructors can save generated assignments as drafts, which students can’t see until the instructor publishes them.
- Fully editable: Generated assignments can be edited the same way as any manually built assignment, including titles, instructions, requirements, and rubrics.
In beta, shaped by instructor feedback
The Assignment Builder is currently in early beta, and Packback is continuing to improve it based on input from instructors using it.
5. Human-in-the-Loop: Instructors Always Decide Grades
Across every assignment type on Packback, including Discussion, Writing, and AI Integrated Assignments, the instructor makes the final grading call. Generative AI is never used to produce recommended scores within assignments. Our recommendations come from deterministic algorithms, so they are always consistent and explainable.
In AI Integrated Assignments, the Evaluation Agent serves strictly as an administrative assistant by highlighting key moments in the transcript and organizing evidence against the course rubric. It does not score student work. The instructor maintains complete control:
- Instructors define, customize, and weight the rubric.
- Instructors make the final grading call and can adjust or override any recommended score or AI-generated summary at will.
6. Our Ethical, Security, and Data Commitments
What Happens to Student Data?
Student Work is Never Used to Train Commercial LLMs: Packback does not send student writing or data back to commercial AI providers to train their public foundation models.
Internal Application Tuning Only: We only use anonymized activity logs internally to ensure our own agents are behaving properly, following instructions, and staying in character.
Strict Security Compliance: Packback adheres to leading data security, privacy, and accessibility standards, including FERPA, COPPA, SOC 2 Type 2, and WCAG AA compliance.
Still Have Questions?
We welcome open conversations about how technology is used in education. For technical questions, security documentation, or assistance configuring AI tools for your course, please consult our security overview or contact help@packback.co.