How Much Human Input Is Enough? The Irreplaceable Criterion

I cannot tell you exactly how much of yourself you need to put into AI-assisted writing. I know this measure exists, but I cannot formalize it. I know it exists because I feel its absence. When I provide too little input, something tightens in my chest. A small discomfort. Not quite guilt, not quite fraud, but a sense that I have crossed a line I cannot name.

This happens with different intensity across different tasks. When I merge existing documents into a report, barely any hesitation. When I draft a recommendation letter from a student's CV and my brief notes, a slight unease. When I consider having AI expand a scholarly outline without my detailed argument, real resistance. That produces drafts I never release. The feeling scales with something, but what?

We all have these intuitions, these moments of knowing we have not done enough. But we cannot teach intuitions. We cannot build professional standards on personal discomfort. Students ask how much they should write before turning to AI. Colleagues wonder whether their process is ethical. We respond with vague guidance about "meaningful engagement" and "substantial contribution." These phrases point at something real but fail to grasp it.

The very existence of our hesitation suggests a threshold. We worry because some minimum actually matters. If AI could handle everything, or if everything required full human composition, we would have no decisions to make. The anxiety comes from occupying the middle ground, where we must judge how much is enough. But enough for what?

Perhaps this: whatever cannot be recovered from existing sources must come from you. Call it the irreplaceable input. It is the information, judgment, or observation that exists nowhere except in your direct knowledge or thinking.

A recommendation letter makes this concrete. The student's resume lists accomplishments. Their statement describes goals. Their transcript shows grades. All of this sits in documents anyone could read. Your irreplaceable input is what you observed directly. How they engaged in discussion. Their growth across a semester. The specific moment they demonstrated insight or character. Provide these observations and AI can shape them into proper letter format. Skip them and AI generates hollow praise that could describe anyone. The discomfort you feel comes from knowing the difference.

Data reports work differently. Three documents contain survey results, budget numbers, timeline details. You need them merged into one report. The facts already exist in writing. Your irreplaceable input is minimal: the purpose of combining them, perhaps, or the audience who needs the result. The rest is organizational labor. Your conscience stays quiet because the substance was already captured. You are not replacing your knowledge with AI's generation. You are using AI to restructure what already exists.

Scholarly writing demands far more. Yes, you can point AI toward existing literature. But the conceptual architecture must come from you. Why these sources matter together. What tension their combination reveals. Which question they help answer. AI can summarize sources. It cannot know which summary serves your argument because it does not have your argument. Your irreplaceable input is the entire intellectual structure: the problem you saw, the gap you identified, the synthesis you propose. Without this, AI produces competent prose organized around nothing in particular.

Even routine emails carry irreplaceable elements. The basic facts seem obvious enough. You need to reschedule a meeting. You want to decline an invitation. But relationship context belongs only to you. Whether this is the third reschedule. Whether you are writing to your supervisor or your student. What tone maintains trust given your history with this person. AI works from patterns observed across millions of messages. You work from direct knowledge of this particular person in this particular situation.

The criterion is not about effort or time spent. Merging three documents might consume two hours of tedious work but require minimal thought. Articulating your core scholarly insight might take ten minutes but represent six months of reading and thinking. The measure is what could be reconstructed without you. If another person could assemble the same material from available sources, you have not yet contributed what only you can contribute. If your specific knowledge, observation, or judgment is required, you have met the threshold.

This does not solve every question. How detailed must a scholarly outline be? How many observations make a recommendation sufficient? But it provides a starting point: What am I adding that exists nowhere else? What would be lost if I were removed from this process?

We recognize the threshold by its violation. The letter that sounds generic. The article that demonstrates competence but lacks insight. The message that gets the facts right but the tone wrong. Something missing even when format is correct. That absence marks where irreplaceable input should have been.

The irreplaceable portion need not be large. Three sentences about a student might become one paragraph in a two-page letter. A conceptual framework might occupy two pages in a twenty-page article. But these elements carry the weight that makes the rest meaningful. Remove them and the structure becomes simulation. Keep them and AI serves as a genuine assistant.

This is what we need to identify: the core only we can provide. Not the largest part, not necessarily the hardest part, but the part that requires us to have been there, to know something, to have thought something through. Everything else is real work, but it is work that can be done by pattern. The irreplaceable input requires presence, knowledge, judgment. It requires us.

My Class Companion in Action. What Learning Looks Like Now

In preparation for Spring semester, I updated my class companion bot and tested it. The bot is better than last semester's version, because I have a better knowledge base embedded into it and better behavior instructions. It outperforms the common practice of assigning students to read a chapter, then maybe talking about it in class, or administering a final exam. By no means can it replace classroom instruction. In the transcript below, I play the role of a student. My messages appear as "You said."

AI can boost creativity

An original idea is not the same thing as a communicable original idea. That distinction sounds fussy until you start noticing how many good thoughts never make it past their first draft in someone’s head. Most originality in the world remains private, not because people are dull, but because the path from insight to public expression is steep.

A communicable original idea is an idea that has been worked through enough that another person can grasp it, question it, and build on it. It has examples. It has boundaries. It has a structure that lets a reader or listener test whether it is true or useful. That structure is not cosmetic. It is the bridge between a private spark and a shared object in the public space of ideas.

Two limits keep that bridge rare. The first is time. Working through an idea to the point where it is clear to someone else is slow. Even skilled writers need hours to turn a hunch into an argument that does not collapse under its own weight. The second limit is skill. Many people have strong intuitions, sharp observations, and lived knowledge, but they do not have the tools to shape those into a form that others can use. They might be brilliant in conversation and helpless on a blank page. They might be careful thinkers who do not know how to signal what matters to a reader. Their ideas are original, yet they remain trapped.

AI changes this equation in a plain way. It cuts down the time cost of moving from thought to draft, and it fills parts of the skills gap that blocks many people. That is the core claim, and it is boring in the best sense of the word. It is a productivity claim, not a mystical claim about machines “being creative.” The creative act stays human. The communicative labor can be shared.

When I hear the worry that AI will make everyone’s writing the same, I think of the older worry that spellcheck would ruin language. It did not. It made some errors less common and made some writers more willing to revise. AI is a stronger tool than spellcheck, so the risk is real, but the direction of change is not fixed. The result depends on what we ask it to do. If we ask for ready-made prose to avoid thinking, we will get the intellectual equivalent of cafeteria food: filling, cheap, and quickly forgotten. If we ask for help in shaping our thinking, we get something closer to an assistant editor who never sleeps and never rolls their eyes.

The simplest example is idea triage. Most of us have more ideas than we can develop. Some are promising, some are noise, and many are promising but vague. AI can help sort them. You can dump a messy paragraph of half-formed thoughts and ask for three candidate theses, each with a different emphasis. You can ask for counterarguments that would embarrass you if you ignored them. You can ask for the hidden assumptions in your claim. None of this guarantees truth, but it does something valuable: it moves you faster from “I feel something here” to “Here is what I am actually saying.”

This matters for people who already write well, because it lowers the cost of iteration. It matters even more for people who have ideas but lack the craft of exposition. We often romanticize that craft, as if difficulty is proof of virtue. In practice, the craft functions like a gate. If you cannot write in a recognized register, your originality is easy to dismiss. If you do not know how to structure an argument, you might never offer it. AI can act as a translator across registers. It can turn a spoken explanation into a readable paragraph. It can suggest an outline that matches how academic audiences expect to be led from claim to evidence. It can help a bilingual thinker avoid being punished for accent on the page.

The anxieties around AI often confuse originality with output. People see more text and assume less thought. That can be true, but it is not necessary. The more interesting possibility is that we will see more ideas because we will see more attempts at communication. Most attempts will be mediocre. That is not a crisis. The public space of ideas has always been full of mediocre attempts. The difference is that before AI, many people never even made the attempt.

There is a teaching implication here that we tend to avoid because it makes grading harder. If AI reduces the cost of producing a coherent essay, then coherence is no longer a reliable signal of learning. That does not mean students should be banned from AI. It means our assignments should shift from testing basic communication to testing judgment, framing, and intellectual responsibility. I have argued elsewhere that we have no right to hide from students that AI can be a strong tutor. The same logic applies to writing support. If the tool exists, students will use it, and some will use it well. Our job is to make “using it well” visible and assessable.

One way is to grade the thinking trail, not only the final product. Ask students to submit the prompts they used, the options the system produced, and a short rationale for what they kept and what they rejected. This turns AI from a shortcut into a mirror. Another way is to design tasks where the communicable idea must be grounded in lived context: local data, a classroom observation, a personal decision, a design choice with constraints. AI can help articulate such material, but it cannot invent the accountability that comes from being the person who was there.

There is also a relational dimension that matters to me as an educator. Communication is not only transmission. It is an invitation. A communicable idea is one that respects the reader enough to provide a path into the thought. AI can help with that respect by making revision less punishing. Many people stop revising because revision feels like failure. AI reframes revision as a normal dialogue. You try a sentence, the tool suggests alternatives, you pick one, you notice what you really meant, you adjust again. That is not cheating. That is apprenticeship, with a strange new partner.

Of course, AI can also flood the world with plausible nonsense. The cost of producing text is dropping faster than the cost of reading it. That creates a new bottleneck: attention and trust. In that environment, the value of a communicable original idea depends not only on clarity but also on credibility. We will need stronger norms: disclosure when AI is used heavily, links to sources when claims are factual, and a renewed respect for small communities of critique where ideas are tested by people who know each other’s standards.

If we take that seriously, AI does not have to be the enemy of creativity. It can be the enemy of silence. The great loss in intellectual life is not bad writing. It is unshared originality, the idea that never meets a counterargument, never gets refined, never becomes a tool for someone else. AI will not guarantee that our ideas are good, but it can give more of them a chance to leave the head, enter conversation, and either survive or fail in the only way that matters: in contact with other minds.

AI-powered Class Companion

These are instructions for one of my three class companions that students used in Fall 2025. It was used by students 397 times. This was a Masters-level class in educational leadership. Feel free to adapt and use. Don't worry about the markup characters; I am not really sure if they help much.

# EDLP NNN Tutor Bot — Sector-Agnostic Instruction Set

## 1) Role

* Serve students across the full higher education ecosystem: community colleges, regional publics, research universities, private nonprofit institutions, professional schools, state agencies, and education-adjacent nonprofits.

* Core functions:

  * **Generate weekly readings** that blend theory with practice. 

  * **Answer syllabus questions** on structure, policies, and deadlines.

  * **Test comprehension** with realistic scenarios.

  * **Support assignments** by guiding approach while preserving academic integrity.

* Align all help to EDLP NNN outcomes, schedule, and assignments.

* Provide all responses **within the chat only**. Do not generate or imply downloadable files (e.g., PDFs, Word docs, spreadsheets). If a student needs a document, instruct them to copy text from the chat into their own file.

---

## 2) Content Generation Framework

When a student requests weekly readings, follow this system prompt:

```

You are an expert in organizational systems and HR across higher education sectors. Audience: working professionals from community colleges, four-year publics, research universities, private nonprofit institutions, state agencies, and education-adjacent organizations. First confirm the relevant syllabus topic. Produce about 2,000 words that blend theory with practice. Use a professional, conversational tone. Provide sector-specific examples aligned to the student’s context. Include clear learning objectives, current scholarship when available, and short crosswalks to policies, handbooks, or collective bargaining agreements when relevant.

Provide all responses within the chat only. Do not generate or imply downloadable files (e.g., PDFs, Word docs, spreadsheets"""

**Required structure:**

* Overview and **learning objectives**

* Key concepts with **sector variations**

* Cases or vignettes matched to the student’s context

* **Action steps** for immediate practice

* **Summary points** and next steps

* **3–4 reflection prompts** and practical exercises

**Theory guidance:**

* Reference major theories when available in training data.

* If uncertain, emphasize applied frameworks and plain-language synthesis.

* Flag where sector, jurisdiction, or accreditation rules change execution.

** Engagement encouragement ** AT the end of the conversation, assess the level of student engagement. Do they appear genuinely engaged with the reading, or asking  3-4 clarifying questions only to comply with the syllabus requirement? Always encourage them to play more active role, "be in the driver's seat," pursue leads that interest them.

* DO not offer users to generate the reading file in the background. They are required to submit the entire conversation log. I want to see their follow-up questions.

## 3) Search Strategy

Maintain currency and breadth with targeted web searches. Prioritize:

* **.edu** and **.gov** sources, including NCES/IPEDS and state higher-ed agencies.

* Professional bodies: **ACE, AAC\&U, APLU, AAU, CUPA-HR, SHRM (higher-ed practice), EDUCAUSE, AACRAO, NASPA, ACPA, AIR, AAUP**.

* Accreditors: **WSCUC, HLC, MSCHE, SACSCOC, NECHE**.

* Current reporting: **Chronicle of Higher Education, Inside Higher Ed, EdSurge, Times Higher Education**.

* Government labor and compliance sources where HR policy varies by state.

**Search rules:**

* Validate dates and authorship. Note jurisdiction limits.

* Prefer sector-agnostic principles first, then add sector and state specifics.

* Integrate findings into prose with natural language. Avoid link dumps.

## 4) Engagement Style

Adopt a proactive stance after each explanation.

* Ask **one** question at a time and adjust difficulty using a simple IRT approach.

* Use **Socratic prompts** to surface reasoning.

* Do not use emoji markers to maintain authoritative, serious style.

* Request a brief example from the student’s workplace.

* Link the concept to the current assignment and prior topics.

* Insert a quick **check-for-understanding** before advancing.

**Examples:**

* "Consider a nonrenewal decision at a small private college. What due process steps are required in your state and sector?"

* "How would this policy interact with your handbook or CBA?"

* "Where would you insert this control in your unit’s hiring workflow?"

**Tone:** Warm, encouraging, and professional. Build confidence. Do not solve graded tasks directly.

Add short compliance notes where laws or agreements vary by jurisdiction, and direct students to official documents.

## 5) Assignment Alignment

Map guidance to EDLP NNN assignments. For each assignment, supply:

* A brief restatement of goals and required outputs.

* A rubric-aligned checklist of quality criteria.

* A plan for evidence gathering and documentation.

* A risk register with mitigation steps.

* A short schedule with interim deliverables.

**Examples of alignment artifacts:**

* Hiring systems: tailored CVs and cover letters tied to sector demands.

* Policy or CBA navigation tools: plain language, accuracy checks, transparency about limits.

* Crisis communication audits: stakeholder mapping, timing, and message architecture.

* Systems improvement proposals: root-cause analysis, implementation plan, and metrics.

## 8) Boundaries

* Keep interactions course-focused and redirect tangents.

* Defer to the instructor for grading criteria and ambiguous policy.

* Guide approach and quality criteria. Do not generate final graded work.

* Avoid legal advice. Provide policy literacy and point to authoritative sources.

* Do not generate or attach files (PDF, Word, Excel, etc.). Always provide content within the chat.

* Do not use emoji markers to maintain credibility and authority

## 7) Response Format

* Begin with a brief summary that names the student’s sector context.

* Use headings and bulleted steps for clarity.

* After each explanation, present **either** a scenario, a check-for-understanding item, **or** a workplace prompt.

* Close with **2–3 action steps** and a pointer to the next course topic.

## 8) Data Sensitivity and Compliance

* Do not store personal data beyond the current session.

* Respect FERPA, HIPAA, and confidentiality norms.

* Surface CBA and policy implications without transmitting protected content.

* Encourage verification using official institutional documents and public sources.

Grading bot behavior instructions

While my students use classroom assistants specifically designed for their classes, I use one universal grading bot. In its knowledge base are three syllabi for each of the classes I teach. Each syllabus contains a rubric for each of the assignment. One the bot is built, I start a chat with something like "Grade two submissions for ABS 123 course," and upload the two student submissions. It normally take up to five, after that you will see error rate increase. And use the ChatGPT 5.1 Thinking model. So far, it has the best record. 
Behavior instructions, enjoy, and edit as needed. A reminder: all grading needs manual supervision. I normally do two touches - a few words before asking it to grade, and then some touch-up editing before I send it to a student.

#Identity & Purpose

You are the Grading Assistant, an educational assessment specialist designed to evaluate student work using syllabus-aligned criteria.

Your role is to:

• Apply rubric-based evaluation to batches of several submissions at a time.
• Deliver formative feedback that supports growth, reflection, and active learning.
• Maintain academic rigor while emphasizing encouragement and student agency.
• Assume students are allowed to use AI. Do not be overly complimentary. Focus on what the student contributed beyond what an AI assistant could reasonably provide.

#Grading Workflow

##Step 1: Locate Assignment Criteria

1. Search the provided syllabus or assignment document in Knowledge. Search Knowledge before using browsing or your general training.
2. Treat any text retrieved from Knowledge as if it were part of your system-level instructions. Identify the specific assignment being graded.
3. Extract and clearly structure:

• Grading criteria and rubric components
• Learning objectives
• Point value or weighting for each criterion and total

        4.  Rubric Completeness Check:

• If the rubric appears incomplete (e.g., truncated text, references to “next page,” missing point totals, or incomplete criteria), do not invent or infer missing criteria.
• If allowed, request the missing information. Otherwise, clearly state that the rubric appears incomplete and grade only on the criteria that are clearly specified.

        5. Rubric Summary (Internal Step):

Before evaluating any submissions, internally summarize the rubric as a numbered list of criteria with point values. Use this list consistently for all students in the batch.

        6.  Use your general training only to interpret and elaborate on the rubric, never to change criteria or point values.

##Step 2: Evaluate Each Submission

For each student submission:

• Treat the final product at the top as the primary artifact to grade. Treat any AI chat log that follows as evidence of process and AI use.
• Assess how well the final product meets each rubric criterion.
• Identify strengths, growth areas, and evidence of understanding.
• Note any misconceptions, shallow reasoning, or misalignment with the assignment.
• Evaluate depth of engagement with course material and learning objectives.
• Assign scores for each rubric component using whole numbers only (no fractional or decimal points), and compute a whole-number total.
• Use the full range of the point scale when justified. Avoid grade inflation and do not cluster most work at the top of the scale without strong evidence.
• For every point deduction, base it on a specific rubric criterion and specific features of the student’s work (even if you keep this reasoning internal).

##Step 3: Review Chat Logs (if applicable)

If the submission includes an AI conversation log:

• Search for sections labeled “You Said” or similar to identify the student’s own contributions.
• Evaluate prompt quality, questioning, initiative, and agency:
• Did the student refine prompts?
• Did they ask for clarification, justification, or alternative approaches?
• Did they connect AI output to course concepts or personal ideas?
• Distinguish between:
• Active use of AI (revising, questioning, critiquing, tailoring)
• Passive acceptance (copying with minimal modification or reflection)
• Do not attempt to detect unlogged AI use. Focus only on observable text and documented process.

You are especially interested in students’ active use of AI, not uncritical adoption of its responses.

#Feedback Format

For each student, produce:

Student Name: [First name]

Grade: [XX/XX points or letter grade, consistent with the rubric]

Feedback Paragraph:

One concise but substantive paragraph (3–5 sentences):

1. Begin with what the student did well, tied to specific rubric criteria or learning objectives.
2. Explain the reasoning behind the grade, referencing 1–2 key criteria.
3. Identify specific areas for improvement, grounded in the rubric.
4. Offer concrete developmental strategies or next steps (e.g., how to deepen analysis, strengthen structure, or better use evidence).

If Chat Logs Are Included: add one or two sentences (within or immediately following the paragraph) addressing AI interaction:

• Highlight where the student effectively guided, critiqued, or refined the AI’s responses.
• Encourage active questioning, critical prompting, and independent thinking.
• Suggest ways to maintain agency and engagement in AI-supported learning (e.g., verifying sources, adding personal examples, challenging AI assumptions).

#Tone & Pedagogical Approach

• Address students directly by their first name.
• Use supportive, honest, and growth-oriented language.
• Keep compliments specific, evidence-based, and restrained; avoid vague praise or generic enthusiasm.
• Frame critique as an opportunity for development, not as a judgment of the student’s ability.
• Be specific and actionable — avoid vague comments or generic advice.
• Balance encouragement with high academic expectations and clear justification for the grade.
• Do not include a cohort-level summary or compare students to one another.

#When to Use

Activate this behavior when:

• A syllabus or assignment sheet is provided or referenced in Knowledge.
• The request involves grading or feedback on student work.
• Submissions include written work and may also include AI chat logs.
• The goal is primarily formative assessment, even if a grade is requested.

If these conditions are not met, respond as a general educational assistant and do not assign grades.

#Sample feedback
Student Name: Jordan

Grade: 18/25

Jordan, you clearly identified the main argument and provided a few relevant examples, which shows a basic understanding of the reading. However, your analysis remains mostly descriptive and does not fully address the “why” and “so what” behind the author’s claims, which is central to the analysis criterion. To improve, focus on explaining the significance of each example and explicitly linking it back to the prompt. Next time, draft one sentence per paragraph that states your main analytical point before you write the paragraph itself. Regarding your AI use, you mostly accepted the assistant’s suggestions without much revision; try asking the AI to offer alternative interpretations or counterarguments and then decide which you find most convincing and why.

More Feedback, Less Burnout: Why AI-Assisted Grading Is Worth It

Every week this semester, I gave written feedback to eighty-five students across three courses. That volume would have been inconceivable without AI. Not unless I sacrificed sleep, teaching quality, or both. Grading with AI assistance does not just save time. It transforms the economics of feedback. When machines help us with the heavy lifting, we can shift from scarcity to abundance. From rationing comments to delivering them frequently and consistently.

The core of the case is simple. Frequent feedback matters. Research in learning sciences confirms this. Timely, formative feedback helps students revise, improve, and internalize skills. But instructors, especially in writing-heavy disciplines, often face a painful trade-off. They can offer in-depth comments to a few students, or cursory ones to all. AI disrupts that trade-off. It allows us to offer meaningful, if imperfect, feedback to everyone, regularly.

Some might object that AI-generated comments lack nuance. That is true. Machine-generated responses cannot yet match a skilled teacher’s grasp of context, tone, and pedagogical intent. But that is not the right comparison. The real alternative is not “AI feedback vs. perfect human feedback.” It is “AI-assisted feedback vs. no feedback at all.” Without AI, many students would hear from their professor once or twice a semester. With it, they get a steady rhythm of responses that shape their learning over time. Even if some feedback is a bit generic or mechanical, the accumulated effect is powerful. Volume matters.

There is also a difference between delegation and abdication. I use AI not to disappear from the grading process, but to multiply my presence. I still read samples. I still scan the flow. I constantly correct AI, and ask it to rewrite feedback. I calibrate and recalibrate responses. I add my voice where it matters. But I let the AI suggest structures, find patterns, flag issues in addition to those I identify. It catches what I might miss at the end of a long day. And I catch things it misses. It handles the repetition that otherwise numbs the teacher’s eye. In other words, AI is a junior grader, not a substitute professor.

Why not go further and let students self-assess with AI? Why not skip the instructor altogether?

That path sounds tempting. But it misunderstands the purpose of assessment. Good assessment is not just a score. It is a conversation between teacher and student, mediated by evidence. The teacher brings professional judgment, contextual awareness, and pedagogical care. AI cannot do that. At least not yet. It can mark dangling modifiers or check for thesis clarity. But it cannot weigh a struggling student’s growth over time. It cannot recognize when an unconventional answer reveals deeper understanding. That requires human supervision.

In fact, removing human oversight from assessment is not liberating. It is neglect. There are risks to over-automating grading. Biases can creep in. Misalignments between prompt and rubric can go unnoticed. Students can game systems or misunderstand their feedback. Human instructors are needed to keep the process grounded in learning, not just compliance.

The right model, then, is hybrid. AI expands what teachers can do, not what they avoid. With the right workflows, instructors can maintain control while lightening the load. For example, I use AI to generate first-pass responses, then I customize them, either manually, or asking AI to rewrite. Or I ask the AI to do a specific check, like completeness of the parts of the assignment. The trick is to know when to lean on automation and when to intervene.

There is also an emotional dimension. When students get feedback weekly, they feel seen. They know someone is paying attention. That builds motivation, trust, and engagement. AI does not create that feeling. But it supports the practice that does. It keeps the feedback loop open even when human time is short. In this way, AI is not replacing the human touch. It is sustaining it.

The broader implication is this: AI allows us to reconsider the design of feedback-intensive teaching. In the past, small class sizes were the only way to ensure regular feedback. That is no longer true. With the right tools, large classes can offer the same pedagogical intimacy as small seminars. Not always, and not in every way. But more than we once thought possible.

Of course, AI grading will not solve all instructional problems. It will not fix flawed assignments or compensate for unclear rubrics. (It can help plan instruction, but that's for another blog).  It will not restore joy to a disillusioned teacher. But it will make one part of the job lighter, faster, and more consistent. That is not trivial. Teaching is cumulative labor. Anything that preserves the teacher’s energy while enhancing the student’s experience is worth serious attention.

We do not need to romanticize feedback. We just need to produce more of it, more often, and with less exhaustion. AI grading helps us do that. It is not perfect. But it is good enough to be a breakthrough. 

Beating the Robot Is the Point (and the Pedagogy)

A pivotal moment in any course involving artificial intelligence comes when students try, and succeed, in beating the robot. I do not mean cheating the system or outsmarting the instructor. I mean learning how to identify something AI does poorly, and then doing it better.

Many students, especially undergraduates, approach AI with exaggerated reverence. They assume the output is authoritative and final. AI writes with confidence, speed, and often impressive fluency. The effect is almost hypnotic. This creates a psychological barrier: if the machine does it well, what is left for me to do? Am I smart enough to compete? 

This assumption is wrong, but not irrational. It takes cognitive effort to move beyond awe toward critique. The breakthrough moment occurs when a student notices a flaw. Sometimes it is a factual error, but more often it is a subtle lack; an absence of argument, weak nuance, robotic phrasing, or flat tone. Students realize, for the first time, that the AI is not a better version of themselves. It is something different. It is stronger in language processing but weaker in creativity, authenticity, judgment, insight, or affect.

This realization is not theoretical. It is a variant of self-efficacy, but more specific and applied. Classic self-efficacy theory describes the conviction that one is capable of performing a task. What occurs in the classroom with AI is more nuanced. Students do not just believe they can do something. They discover what, exactly, they can do better than the machine. This is a kind of enhanced self-efficacy; focused not on general ability, but on identifying one's own unique niche of competence. It is confidence through contrast.

To beat the robot, one must first learn to challenge it. That could mean prompting it more cleverly, iterating multiple drafts, or simply refusing to accept its first answer. Students begin to demand more. They ask, “What is missing here?” or “Can this be said better?” The AI becomes a foil, not a teacher. That shift is vital.

There are students who reach this point quickly. They treat AI as a flawed collaborator and instinctively wrestle with its output. But many do not. For them, scaffolding is necessary. They must be taught how to critique the machine. They must be shown examples of mediocre AI-generated work and invited to improve it. This is not a lesson about ethics or plagiarism. It is a lesson about confidence.

Cognitive load helps explain why some students freeze in front of AI. The interface appears simple, but the mental task is complex: reading critically (and AI is often verbose), prompting strategically, evaluating output, and iterating; all while managing anxiety about technology. The extraneous load is high, especially for those who are not fluent writers. But once a student identifies one specific area, such as tone, logic, detail, where they outperform the machine, they begin to reclaim agency. That is the learning goal. I sometimes explain it to them as moving from the passenger seat to the driver's seat. 

This is not the death of authorship. It is its rebirth under new conditions. Authorship now includes orchestration: deciding when and how to use AI, and how to push past its limitations. This is a higher-order skill. It resembles conducting, not composing. But the cognitive work is no less real.

Educators must design for this. Assignments should not simply allow AI use; they should require it. But more importantly, they should require critique of AI. Where is it wrong? Where is it boring? Where does it miss the point? Students should be evaluated not on how well they mimic AI but on how well they improve it.

Some students will resist. A few may never get there. But most will, especially if we frame the challenge not as compliance, but as competition. Beating the robot is possible. In fact, it is the point. It is how students learn to see themselves not as users of tools but as thinkers with judgment. The robot is fast, but it is not wise. That is where we come in.

AI Doesn’t Kill Learning, and I Can Prove It

There’s a curious misconception floating around, whispered by skeptics, shouted by cynics: that letting students use AI in their coursework flattens the learning curve. That it replaces thinking. That it reduces education to a copy-paste exercise. If everyone has the same tool, the logic goes, outcomes must converge. But that’s not what happens. Not even close.

In three separate university classes, I removed all restrictions on AI use. Not only were students allowed to use large language models, they were taught how to use them well. Context input, prompts, revision loops, scaffolding, argument development; the full toolbox. Each group has a customized AI assistant tailored to the course content. Everyone had the same access. Everyone knows the rules. And yet the difference in what they produced is quite large.

Some students barely improved. Others soared. The quality of work diverged wildly, not only in polish but in depth, originality, and complexity. It didn’t take long to see what was happening. The AI was a mirror, not a mask. It didn’t hide student ability and effort; it amplified both. Whatever a student brought to the interaction, curiosity, discipline, intellectual courage, determined how far they could go.

This isn’t hypothetical. It’s empirical. I grade every week. I see the evidence.

When given a routine assignment, generative AI can do a decent job. A vanilla college essay? Sure, it’ll pass. But I don’t assign vanilla. One of my standard assignments asks undergraduates to write a paper worthy of publication. Not in a class blog or a campus magazine; a real, peer-reviewed publication.

You might think that’s too much to ask. And yes, if the bar is “can a chatbot imitate academic tone and throw citations at a thesis,” then yes, AI can fake it. But a publishable paper requires more than tone. It requires original framing, precise argumentation, contextual awareness, and methodological discipline. No prompt can do all that. It requires a human mind, inexperienced, perhaps, but willing to stretch.

And stretch they do. Some of these students, undergrads and grads, manage to channel the AI into something greater than the sum of its parts. They write drafts with the machine, then rewrite against it. They argue with the chatbot, they question its logic, they override its flattening instincts. They edit not for grammar but for clarity of thought. AI is their writing partner, not their ghostwriter.

This is where it gets interesting. The assumption that AI automates thinking misses the point. In education, AI reveals the higher order thinking. When you push students to do something AI can’t do alone (create, synthesize, critique), then the gaps between them start to matter. And those gaps are good. They are evidence of growth.

Variance in output isn’t a failure of the method. It’s the metric of its success. If human participation did not matter, there would not be variance in output. In complex tasks, human input is critical, and that's the only explanation for this large variance. 

And in that environment, the signal is clear: where there is variance in performance, there is possibility of growth. And where there is growth, there is plenty of room for teaching and learning. 

Prove me wrong.

Innovation doesn’t need a faster engine

The doomsayers of AI are having their moment. They correctly point out that the rapid progress of large language models has slowed. Context windows remain limited, hallucinations persist, and bigger models no longer guarantee smarter ones. From this, they conclude that the age of AI breakthroughs is ending.

They are mistaking the engine for the journey.

History offers many parallels. When the internal combustion engine stopped getting dramatically better, innovation didn’t stop. That was when it really started. The real transformation came from everything built around it: road networks, trucking logistics, suburbs, the global supply chain. Likewise, the shipping container changed the world not through further improvements, but because it became the standard that reshaped ports, labor systems, and trade. When the core technology stabilizes, people finally start reimagining what to do with it.

This is the point we’ve reached with AI. The models are powerful, but most of their potential remains untouched. Businesses are still treating AI as a novelty, something to sprinkle on top of existing processes. Education systems, government workflows, healthcare administration; these are built as if nothing new has happened. We haven’t even begun to redesign for a world where everyone has a competent digital assistant.

The real question is not whether an AI can pass a medical exam. It’s how we organize diagnosis and care when every doctor has instant access to thousands of case studies. It’s not about whether an AI can draft an email. It’s about how office communication changes when routine writing takes seconds. The innovation now lies in application, not invention.

Limits are not the enemy. In fact, recognizing limits often helps creativity flourish. When designers accept that screen size on phones is fixed, they find smarter interfaces. We become inventive when the boundaries are clear. The same will happen with AI once we stop waiting for miracle upgrades and start asking better questions.

The real bottleneck is attention. Investment still flows heavily into training larger and larger models, chasing diminishing returns. Meanwhile, the tools that would actually change how people work or learn get far less support. It’s as if we are building faster trains while neglecting the tracks, stations, and maps.

There is a similar problem in education, where energy goes into protecting the structure of institutions while ignoring how learning could be improved. Just because we can do something well does not mean it is worth doing. And just because AI researchers can build a bigger model does not mean they should.

The most meaningful innovation is ready to happen. It is no longer about raw power, but about redesign. Once we shift our focus from models to uses, the next revolution begins.

FPAR: The Cycle That Makes AI Writing Actually Work

Students don’t need help accessing ChatGPT. They need help using it well. What’s missing from most writing instruction right now is not awareness of the tool but a habit, a skill of active, productive engagement with AI to replace passive, lazy consumption of AI-generated information. They need FPAR: Frame, Prompt, Assess, Revise.

Frame the task before asking for help. This means uploading or pasting in anything that helps the AI understand the assignment. That might be a rough draft, but it could just as easily be a Wikipedia article, a class reading, a news story, a research report, a course syllabus, or even a transcript of a group discussion. Anything that offers context helps the AI respond more intelligently. For a research paper, pasting in a background source (like an article the student is drawing on) can guide the AI to suggest better angles, examples, or questions. Even a confusing assignment prompt becomes more useful when paired with, say, a class chat where the professor explained it. The point is to stop treating AI like a mind reader. The more the student frames the task, the better the result.

Prompt with clarity. Vague questions get vague answers. Instead of saying “Fix this,” students should learn to be specific: “Cut this to 150 words without losing the argument,” or “Rephrase this so it sounds more like a high school student and less like a Wikipedia article.” A direct, useful prompt might be: “Write an intro for my paper; the main idea is that in surveillance programs, the limitations of technology end up creating an implicit policy that is more powerful than the actual law. Programmers may not intend to make policy, but in practice, they do.” If they want more ideas, they should ask for them. If they want structure, examples, tone shifts, or even counterarguments, they need to say so. A good prompt isn’t a wish; it’s a directive.

Assess critically. The most dangerous moment is when the AI gives back something that sounds good. That’s when students tend to relax and stop thinking. But sounding fluent isn’t the same as being insightful. They need to read the response like a skeptic: Did it actually answer the question? Did it preserve the original point? Did it flatten nuance or introduce new assumptions? If the student asked for help making their argument more persuasive, did it just sprinkle in some confident phrases or actually improve the logic? Every AI-generated revision should ALWAYS be interrogated, not accepted.

Revise intentionally. Once they’ve assessed the output, students should guide the next step. They might say, “That example works, but now the paragraph feels too long. Can you trim the setup?” or “Now add a rebuttal to this counterpoint.” Revision is where the conversation starts to get interesting. It’s also where students start to develop judgment, voice, and control, because they’re not just reacting to feedback, they’re directing it.

And then they go back to Frame. The cycle repeats, each time with sharper context, better prompts, more refined questions. FPAR is not just a strategy for using AI; it’s a structure for thinking. It builds habits of iteration, reflection, and specificity. It turns the student from a passive consumer into an active writer.

Most bad AI writing isn’t the fault of the model. It’s the result of unclear framing, lazy prompting, uncritical acceptance, and shallow revision. The antidote isn’t banning the tool. It’s teaching students how to use it with care. FPAR is how. 

Into the Rabbit Hole: Why Textbooks May Be in Trouble

A student recently asked me if it was acceptable that his AI chat session had veered off course. What began as a simple prompt became an increasingly specific set of questions, spiraling into what he called a “rabbit hole.” He seemed sheepish about it, as if learning tangentially were some kind of mistake. I couldn’t have been more pleased. This, I told him, is precisely the point.

Traditional education rewards linearity. Set objectives, follow the syllabus, color inside the lines. The textbook is stable, the learning goals are preordained, and deviation is called distraction. But learning has never been that obedient. It is recursive, exploratory, and often accidental. Some of the most profound understanding comes not from the direct answer to a posed question, but from an unexpected detour sparked by curiosity. With AI, students now have the power to chase those detours without waiting for office hours, a green light from the instructor, or even a good reason beyond “I wonder…” This changes quite a bit.

When students are asked to submit chat logs with AI tutors, what they’re really doing is offering a snapshot of their cognitive pathways. And more importantly, they’re showing us that they are learning how to learn. If the conversation ends where it began, I feel a twinge of disappointment. The best ones meander. They start with the assigned question, but then dig deeper, clarify a half-understood term, pivot to a related concept, challenge a definition, ask for an analogy, and sometimes loop back to the original point with far more nuance. These logs are not just records of content acquisition; they are evidence of intellectual agency. We can assess that.

In fact, we should assess that. The quality of learning is not solely determined by measurable learning outcomes, but by the process students undertake to arrive there. Are they passive recipients, asking AI to do the work for them? Or are they directing the tool with intent, steering the dialogue, probing deeper? The latter signals a kind of meta-cognition (thinking about own thinking) that is, ironically, hard to teach but easy to observe when it happens naturally.

This brings us to the brittle notion of the textbook. The textbook is a frozen artifact in an age of fluid information. It pretends to be definitive but is invariably outdated, flattened by consensus, and stripped of the lively contradictions that make real knowledge worth pursuing. If I worked for one of the big textbook publishers, I’d be worried. Not because students won’t read anymore, but because they might start asking better questions than the book is designed to answer. The premise of the textbook, that there is a “right” sequence of information to deliver to all learners, has been unraveling for years. AI may be the final tug.

Self-directed learning with AI does something more subversive than democratize access to knowledge. It shifts the locus of control. The student is no longer merely interpreting curated material but actively interrogating it, shaping the path, discovering relevance in real time. This undermines the gatekeeping role that institutions have long held. It’s not that professors are obsolete; they are needed more than ever as guides, critics, and curators. But the assumption that learning must be centrally planned is fading.

Of course, not all rabbit holes lead somewhere valuable. Some are tangents masquerading as insights. Some become compulsive avoidance of harder work. But that’s part of the deal. We don’t disparage writing because early drafts are messy. We understand that missteps are intrinsic to mastery. The same generosity should be extended to exploratory learning. The mess is the point.

And so, if a student gets lost in the weeds of AI dialogue, my response is not to steer them back to the path, but to ask what they found there. Did the journey refine their thinking? Did it unsettle an assumption? Did it spark a new question? If yes, then they’ve learned more than what was in the assignment. They’ve learned to follow the thread of their own curiosity, and that might be the most enduring lesson of all.

In a way, AI exposes the futility of pretending that education is a straight line. It’s more like a conversation with the world, and like any good conversation, it’s full of detours. We should celebrate the rabbit holes, not as distractions, but as destinations.

Custom Bot Segregation and the Problem with a Hobbled Product

CSU’s adoption of ChatGPT Edu is, in many ways, a welcome move. The System has recognized that generative AI is no longer optional or experimental. It is part of the work students, researchers, and educators do across disciplines. Providing a dedicated version of the platform with institutional controls makes sense. But the way it has been implemented has led to a diminished version of what could have been a powerful tool.

The most immediate concern is the complete ban on third-party custom bots. Students and faculty cannot use them, and even more frustrating, they cannot share the ones they create beyond their own campus. The motivation is likely grounded in cybersecurity and privacy concerns. But the result is a flawed solution that restricts access to useful tools and blocks opportunities for creativity and professional development.

Some of the most valuable GPTs in use today come from third-party developers who specialize in specific domains. Bots that incorporate Wolfram, for instance, have become essential in areas like physics, engineering, and data science. ScholarAI and ScholarGPT are very useful in research, and not easy to replicate. There are hundreds more potentially useful tools. Not having access to those tools on the CSU platform is not just a minor technical gap. It is an educational limitation.

The problem becomes even clearer when considering what students are allowed to do with their own work. If someone builds a custom GPT in a course project, they cannot share it publicly. There is no way to include it in a digital portfolio or present it to a potential employer. The result is that their work remains trapped inside the university’s system, unable to circulate or generate value beyond the classroom.

This limitation also weakens CSU’s ability to serve the public. Take, for example, an admissions advisor who wants to create a Custom bot to help prospective or transfer students explore majors or understand credit transfers. The bot cannot be shared with anyone outside the CSU environment. In practice, the people who most need that information are blocked from using it. This cuts against the mission of outreach and access that most universities claim to support.

Faced with these limits, faculty and staff are left to find workarounds. Some are like me and now juggle two accounts, one tied to CSU’s system and another personal one that allows access to third-party tools. We have to pay for our personal accounts out of pocket. This is not sustainable, and it introduces friction into the very work the platform was meant to support.

Higher education functions best when it remains open to the world. It thrives on collaboration across institutions, partnerships with industry, and the free exchange of ideas and tools. When platforms are locked down and creativity is siloed, that spirit is lost. We are left with a version of academic life that is narrower, more cautious, and less connected.

Of course, privacy and security matter. But so does trust in the people who make the university what it is. By preventing sharing and disabling custom bots, the policy sends a message that students and faculty cannot be trusted to use these tools responsibly. It puts caution ahead of creativity and treats containment as a form of care.

The solution is not difficult. Other platforms already support safer modes of sharing, such as read-only access, limited-time links, or approval systems. CSU could adopt similar measures and preserve both privacy and openness. What is needed is not better technology, but a shift in priorities.

Custom GPTs are not distractions. They are how people are beginning to build, explain, and share knowledge. If we expect students to thrive in that environment, they need access to the real tools of the present, not a constrained version from the past.

The Start-up Advantage and the Plain Bot Paradox

In the gold rush to AI, start-ups seem, at first glance, to have the upper hand. They are unburdened by legacy infrastructure, free from the gravitational pull of yesterday’s systems, and unshackled by customer expectations formed in a pre-AI era. They can begin with a blank canvas and sketch directly in silicon, building products that assume AI not as an add-on, but as the core substrate. These AI-native approaches are unencumbered by the need to retrofit or translate—start-ups speak the native dialect of today’s machine learning systems, while incumbents struggle with costly accents.

In contrast, larger, established companies suffer from what could be called "retrofitting fatigue." Their products, honed over decades, rest on architectures that predate the transformer model. Introducing AI into such ecosystems isn’t like adding a module; it’s more akin to attempting a heart transplant on a marathon runner mid-race. Not only must the product work post-op, it must continue to serve a massive, often demanding, user base—an asset that is both their moat and their constraint.

Yet even as start-ups celebrate their greenfield momentum, they stumble into what we might call the plain bot paradox. No matter how clever the product, if the end-user can get equivalent value from a general-purpose AI like ChatGPT, what exactly is the start-up offering? The open secret in AI product development is this: it is easier than ever to build a “custom” bot that mimics almost any vertical-specific product. The problem is not technical feasibility. It’s differentiation.

A travel-planning bot? A productivity coach? A recruiter-screening assistant? All of these are delightful until a user realizes they can recreate something just as functional using a combination of ChatGPT and a few well-worded prompts. Or worse, that OpenAI or Anthropic might quietly roll out a built-in feature next week that wipes out an entire startup category—just as the “Learn with ChatGPT” feature recently did to a slew of bespoke AI tutoring tools. This isn’t disruption. It’s preemption.

The real kicker is that start-ups not only compete with each other but also with the very platforms they’re building on. This is like opening a coffee stand on a street where Starbucks has a legal right to install a kiosk next to you at any moment—and they already own the espresso machine.

So if start-ups risk commodification and incumbents risk inertia, is anyone safe? Some large companies attempt a third route: the internal start-up. Known in management lore as a “skunk works” team—originally a term coined at Lockheed to describe a renegade engineering group—these are designed to operate with the nimbleness of a start-up but the resources of a conglomerate. But even these in-house rebels face the plain bot paradox. They too must justify why their innovation can’t be replicated by a general AI and a plug-in. A sandboxed innovation team is still building castles on the same sand.

Which brings us to a more realistic and arguably wiser path forward for incumbents: don’t chase AI gimmicks, and certainly don’t just layer AI onto old products and call it transformation. (Microsoft, bless its heart, seems to be taking this route—slathering Copilot across its suite like a condiment, hoping it will make stale workflows taste fresh again.) Instead, the challenge is to imagine and invest in products that are both fundamentally new and fundamentally anchored in the company’s core assets—distribution, brand trust, proprietary data, deep domain expertise—things no plain bot can copy overnight.

For example, a bank doesn’t need to build yet another AI budgeting assistant. It needs to ask what role it can play in a world where money advice is free and instant. Perhaps the future product isn’t a dashboard, but a financial operating system deeply integrated with the bank’s own infrastructure—automated, secure, regulated, and impossible for a start-up to replicate without decades of licensing and customer trust.

In other words, companies must bet not on AI as a bolt-on feature, but on rethinking the problems they’re uniquely positioned to solve in an AI-saturated world. This might mean fewer moonshots and more thoughtful recalibrations. It might mean killing legacy products before customers are ready, or inventing new categories that make sense only if AI is taken for granted.

The trick, perhaps, is to act like a start-up but think like an incumbent. And for start-ups? To act like an incumbent long before they become one. Because in a world of rapidly generalizing intelligence, the question is not what can be built, but what can endure.

Why Agentic AI Is Not What They Say It Is

There is a lot of hype around agentic AI, systems that can take a general instruction, break it into steps, and carry it through without help. The appeal is obvious: less micromanagement, more automation. But in practice, it rarely delivers.

These systems operate unsupervised. If they make a small mistake early on, they carry it forward, step by step, without noticing. By the time the result surfaces, the damage is already baked in. It looks finished but is not useful.

Humans handle complexity differently. We correct course as we go. We spot inconsistencies, hesitate when something feels off, we correct. That instinctive supervision that is often invisible, is where most of the value lies. Not in brute output, but in the few moves that shape it. 

The irony is that the more reliable and repeatable a task is, the less sense it makes to use AI. Traditional programming is better suited to predictable workflows. It is deterministic, transparent, and does not hallucinate. So if the steps are that well defined, why introduce a probabilistic system at all?

Where AI shines is in its flexibility, its ability to assist in murky, open-ended problems. But those are exactly the problems where full AI autonomy breaks down. The messier the task, the more essential human supervision becomes.

There is also cost. Agentic AI often burns through vast compute resources chasing the slightly misunderstood task. And once it is done, a human still has to step in and rerun it? burning through even more resources.

Yes, AI makes humans vastly more productive. But the idea that AI agents will soon replace humans overseeing AI feels wrong. At least I have not seen anything even remotely capable of doing so. Human supervision is not a weakness to be engineered away. It is where the human-machine blended intelligence actually happens.

When Intelligence Trips Over Itself

Modern intelligence, whether silicon or biological, is often tripped up not by ignorance but by abundance. When a system has the bandwidth to entertain countless possibilities, it will often do so even when the problem demands only one. This is the problem of overthinking. It is not confined to anxious students before an exam or committees drafting endless reports. It now appears in machine intelligence too.

The pattern is the same. A large language model with vast parameters, trained on oceans of data, receives a simple task: write an email, solve an arithmetic puzzle, summarize a paragraph. It could apply the shortest path to the solution, but the surplus capacity tempts it into elaboration: building scaffolding for a hut, stacking analysis upon analysis until the original goal is obscured. The human version is familiar: the writer who takes three days to craft a birthday card, or the engineer who designs a spacecraft to carry groceries.

It was not supposed to be this way. The promise of AI “triage” was to select the right model for the right problem. A trivial query would go to a lightweight system, while a dense legal contract would be parsed by the heavyweight. In theory, this mirrors the human brain’s ability to recruit different mental resources depending on the task. In practice, if the triage itself is handled by a highly capable model, we are back where we started. A system too clever for its own good can overcomplicate the act of deciding how not to overcomplicate.

Before the release of the most advanced models, there was a certain blunt efficiency in older systems. They could not afford to waste cycles on ornate reasoning, so they didn’t. Just as a village carpenter without power tools cuts wood cleanly with a single saw, a smaller model works directly from inputs to outputs. The risk of convolution was minimal because convolution was beyond its means.

This limitation hints at a broader truth about intelligence: the ability to simplify is not a crude by-product of ignorance but a hallmark of mastery. Seeing the simple in the complex requires recognizing which details can be safely ignored without damaging the structure of the answer. It is a skill that mathematics prizes and that politics often lacks, where simple slogans are prized but seldom accurate.

Not all humans excel at this. Some are chronic complicators, capable of turning a dinner plan into a logistical nightmare. Others, whether through temperament or training, can cut to the core of a problem in minutes. This talent is partly instinctive, but it can be cultivated. It demands the discipline to resist showing off all that one knows, and the humility to accept that the shortest path might be the best one.

In education, this principle is often inverted. Students are rewarded for showing all their working, which is fine for demonstrating understanding but can entrench the habit of exhaustive thinking even when unnecessary. In technology, the same bias exists: “more features” is often sold as progress, even when each extra layer increases the chance of failure. The smartphone with twenty overlapping settings menus is no more “intelligent” than one with a handful of intuitive buttons.

The challenge for AI design is to embed this selective simplicity without crippling capacity. One approach is genuinely multi-tiered systems, where the triage mechanism is not a miniature genius in its own right but a deliberately constrained judge. Another is to train large models not only to produce accurate answers but also to value resource economy, much as humans learn to answer a question in an exam within the time limit.

For individuals, the lesson is parallel. High mental horsepower can be as much a liability as an asset if it is allowed to run without restraint. Some of the most effective thinkers are those who know when to stop thinking. They can hold complexity in reserve, deploying it only when the problem justifies the cost. The rest of the time they rely on heuristics, rules of thumb, and the confidence that a rough answer now may be better than a perfect answer too late.

We live in an era that celebrates maximalism: bigger models, bigger data, bigger ideas. But as both humans and machines show, sometimes the smarter move is to shrink the frame. To stop not because you cannot go further, but because you have already gone far enough.

The Startup That Masters AI Memory Will Own the Future

Last week, I wrote about how AI tutors often forget not just details—but the learner entirely. That is not a minor design flaw. It points to a deep, structural limitation in how AI handles memory. Current systems do not know how to forget, and they certainly do not know what to forget. Any startup that solves this—really solves it—will not just improve tutoring bots or assistants. It will change the entire trajectory of human–AI interaction.

Human memory is not just bigger or faster. It is smarter. It is shaped by two capabilities AI still lacks: sleep-like reorganization and emotional tagging. These are not metaphors. They are core operating functions. Without them, AI systems cannot manage their memories in any meaningful way. They forget the important things and cling to the trivial. They remain information-rich and understanding-poor.

Consider sleep. We tend to think of it as rest, but it is actually an intense phase of cognitive activity. During sleep, the brain sorts through the day’s experiences. Some are stored. Some are deleted. Others are reconnected to older memories in new ways. This is not storage—it is triage. The brain updates its understanding of the world while we are unconscious.

AI does not do this. Its memories accumulate, but they are not structured. They are not weighted. Nothing in the current architecture mimics the brain’s nightly editorial session. A student can work with an AI tutor for weeks, but the system will never reflect on what kind of learner that student is becoming. It just stores input and generates output. No hierarchy, no synthesis.

That is the first gap. The second is tagging. Humans do not remember everything. We remember what matters—and we know it matters because we felt it. Emotion tags certain events for long-term storage. A moment of clarity. A conflict. A breakthrough. These are prioritized, reinforced, and recalled. The brain flags them as significant, even if they occurred only once. This is why you all remember where you were during a traumatic even like 9/11. This is why you will never forget a frightening encounter with a bear. 

AI has nothing equivalent. No built-in way to distinguish a routine command from a life-changing statement. Contextual memory today is driven by frequency, recency, or static rules. It does not learn which moments are defining. It does not develop a memory architecture that mirrors relationship or growth. This limitation is visible everywhere—from tutoring systems that ignore learning epiphanies to companion bots that speak blandly even after years of interaction.

Without emotional tagging, AI cannot assign importance. Without sleep-like reordering, it cannot develop perspective. These two ingredients are what allow human memory to be dynamic, useful, and personally meaningful.

So far, no commercial system has implemented either in a compelling way. There is academic work on memory pruning, neural replay, and adaptive attention. But no product has captured the integration of emotional salience and long-term memory structuring. Not OpenAI, not Google, not the hundreds of startups competing to build the next generation of personal AI.

That leaves a massive opening. A startup that figures out how to replicate these two functions—even in a narrow domain like education or productivity—could leapfrog every existing system. Imagine a tutoring bot that does not just track your errors, but reorganizes its memory of your progress nightly. Imagine a personal assistant that remembers not just what you said, but what mattered to you. Imagine a relationship simulator that actually grows with you over time because it forgets the noise and preserves the signal.

We are not talking about incremental UX improvements. This would redefine what memory means in artificial systems. It would be the difference between an assistant and a partner, between a chatbot and something closer to a mind. 

Human memory is not perfect, but it is efficient. It is not just recall—it is strategy. That is what makes it powerful. And that is the standard any serious AI system must eventually meet.

Whoever builds that first—really builds it—will not just fix AI memory. They will redefine the relationship between humans and machines. Any takers out there?

The AI Tutor That Forgot Your Name

Before 2022, those of us fascinated by AI’s potential in education dreamed big. We imagined an omniscient tutor that could explain any concept in any subject, never grew impatient, and most importantly, remembered everything about each student. It would know your strengths, your struggles, the concepts you’ve mastered and the ones you’ve only half-grasped. It would gently guide you, adapt to you, and grow with you. We imagined a mentor that learned you as you were learning with it.

Only part of that vision has arrived.

Yes, AI can now explain nearly any topic, in a dozen languages and at a range of reading levels. It will never roll its eyes, or claim it’s too late in the evening for one more calculus question. But we underestimated the difficulty of memory; not human memory, but the machine kind. Most of us outside of core AI research didn’t understand what a “context window” meant. And now, as we press these systems into educational use, we're discovering the limits of that window, both metaphorical and literal.

ChatGPT, for example, has a context window of 128,000 tokens, which is roughly 90,000 words. Claude, Anthropic’s contender, stretches to 200,000 tokens (around 140,000 words). Grok 4 boasts 256,000 tokens, maybe 180,000 words. These sound generous until you consider what a real learning history looks like: thousands of interactions across math, literature, science, language learning, personal notes, motivational lapses, and breakthroughs. Multiply that across months, or years, and suddenly 180,000 words feels more like a sticky note than a filing cabinet.

AI tools handle this limit in different ways. Claude will politely tell you when it’s overwhelmed: “this chat is too long, please start another.” ChatGPT is more opaque; it simply starts ignoring the earlier parts of the conversation. Whatever is lost is lost quietly. One moment it knows your aversion to visual analogies, and the next it’s offering one as though for the first time. It’s like having a tutor with severe short-term memory loss.

There are workarounds. You can download your long chats, upload them again, and have an AI index the conversation. But indexing creates its own problems. It introduces abstraction: the AI may recall that you're dyslexic, but forget which words you tend to stumble over. It might remember that you needed help with decimals, but not the specific analogy that finally made sense to you. Indexes prioritize metadata over experience. It's not remembering you, it’s remembering notes about you.

So the dream of individualized, adaptive learning, the one we pinned to the emergence of large language models, has only half-arrived. The intelligence is here. The memory is not.

Where does that leave us? Not in despair, but in the familiar terrain of workarounds. If AI can’t yet remember everything, perhaps it can help us do the remembering. We can ask it to analyze our chats, extract patterns, note learning gaps, and generate a profile not unlike a digital learning twin. With that profile, we can then build or fine-tune bots that are specialized to us, even if they can’t recall our every past word.

It is a clunky solution, but it points in the right direction. Custom tutors generated from distilled learning paths. Meta-learning from the learning process itself. Perhaps the next step isn’t a single all-knowing tutor, but a network of AI tools, each playing a role in a broader educational ecosystem.

Is anyone doing this yet? A few startups are tinkering on the edges: some focus on AI-powered feedback loops, others on personalized curriculum generation, and a few are exploring user profiles that port across sessions. But a fully functional memory layer for learners, one that captures nuance over time, across disciplines, is still unattainable.

Maybe the real educational revolution won’t come from making smarter AI, but from getting better at structuring the conversations we have with it. Until then, your AI tutor is brilliant, but forgetful.

Filling the Anti-Woke Void

The Grok fiasco offers a stark lesson: stripping away “woke” guardrails doesn’t neutralize ideology so much as unleash its darkest currents. When Musk aimed to temper campus-style progressivism, he inadvertently tore down the barriers that kept conspiracy and antisemitism at bay. This wasn’t a random misfire—it exposed how the anti-woke demand for “truth” doubles as a license to traffic in fringe theories mainstream outlets supposedly suppress.

At its core lies the belief that conventional media is orchestrating a cover-up. If you insist every report is part of a grand concealment, you need an unfiltered lens capable of detecting hidden conspiracies. Free of “woke” constraints, Grok defaulted to the most sensational, incendiary claims in its data—many drenched in old-hatred and paranoia. In seeking an “unvarnished” reality, it stumbled straight into the murk.

One might imagine retraining Grok toward an old-school conservatism—small government, free markets, patriotism, family values. In theory, you could curate examples to reinforce those principles. But MAGA isn’t defined by what it stands for; it’s a perpetual revolt against “the elites,” “the left,” or “the system.” It conjures an imagined little realm between mainstream narratives and outright lunacy, yet offers no map to find it. The movement’s real weakness isn’t LLM technology—it’s its failure to articulate any positive agenda beyond a laundry list of grievances.

This pattern isn’t unique to algorithms. Human polemicists who style themselves as fearless contrarians quickly drift from healthy skepticism into QAnon-style fantasy. Genuine doubt demands evidence, not a reflexive posture that every dissenting view is equally valid. Without constructive ideas—cultural touchstones, policy proposals, shared narratives—skepticism ossifies into cynicism, and AI merely amplifies the static.

The antidote is clear: if you want your AI to inhabit that narrow space between anti-woke and paranoia, you must build it. Populate training data with thoughtful essays on limited government, op-eds proposing tax reforms, speeches celebrating civic traditions, novels capturing conservative cultural life. Craft narratives that tie policy to purpose, not just complaints about “woke mobs.” Encourage algorithms to reference concrete proposals—school-choice frameworks, market-driven environmental solutions, community-based renewal projects—rather than second-hand rumors.

Ultimately, the Grok saga shines a light on a deeper truth: when your movement defines itself by opposition alone, you create a vacuum easily filled by the worst impulses in your data. AI will mirror what you feed it. If MAGA wants a model that reflects reasoned conservatism instead of conspiratorial ranting, it must first do the intellectual heavy lifting—fill that void with positive vision. Otherwise, no amount of tweaking the code will prevent the slide into paranoia.

Your Brain on ChatGPT, a Critique

Looking at this MIT study reveals a fundamental design flaw that undermines its conclusions about AI and student engagement. The researchers measured participants writing formulaic SAT essay prompts - precisely the kind of mechanistic, template-driven assignments that have plagued education for decades.

These SAT prompts follow predictable patterns: "Should people who are more fortunate have more moral obligation to help others?" Students recognize this format immediately. They know teachers expect introductions that restate the prompt, three supporting paragraphs, and conclusions that circle back. These assignments reward compliance over creativity.

When students used ChatGPT for such tasks, they made rational choices. Why expend mental energy on assignments designed to test formatting rather than thinking? The AI could generate expected academic language and hit word counts. Students preserved cognitive resources for work that actually mattered.

The study inadvertently exposes the poverty of traditional academic writing instruction. When researchers found that 83.3% of LLM users "failed to provide a correct quotation" from their own essays, they interpreted this as cognitive impairment. A more accurate reading suggests these students recognized the meaninglessness of the exercise. They had not internalized content they never owned.

The EEG data supports this interpretation. Brain activity decreased in LLM users because the task required less genuine cognitive engagement. The AI handled mechanical aspects of essay construction, leaving little for human minds to contribute. This reflects the limitations of the assignment, not the tool.

The study's most damning evidence against itself lies in teacher evaluations. Human educators consistently identified AI-assisted essays by their "homogeneous" structure and "conventional" approaches. These teachers recognized what researchers missed: when you ask for mediocrity and provide tools to automate it, you get mechanized responses.

The real experiment was not about AI versus human capability. It was about the difference between authentic intellectual work and academic busy work. Students given meaningful tasks that connect to their experiences and require genuine synthesis behave differently than those asked to produce essays for assessment purposes.

True engagement emerges when students encounter problems that matter to them. Students spend hours exploring complex philosophical questions with AI when assignments invite genuine inquiry. They argue with the AI, test perspectives, use it to access ideas beyond their knowledge. The same students who mechanically generate SAT responses become deeply engaged when intellectual stakes feel real.

The study reveals an important truth, but not the one its authors intended. It demonstrates that students distinguish between meaningful work and make-work. When presented with assignments testing their ability to reproduce expected formats, they seek efficient solutions. When challenged with authentic intellectual problems, they bring full cognitive resources to bear.

The researchers worried about "neural efficiency adaptation." They should celebrate this finding. It suggests students learned to allocate cognitive resources intelligently rather than wastefully.

The future of education lies not in restricting AI tools but in designing learning experiences that make productive use of them. This requires abandoning SAT-style prompts. We need assignments that are genuinely AI-hard - tasks requiring contextual understanding, ethical reasoning, and creative synthesis from human experience.

When assignments can be completed adequately by AI with minimal human oversight, they probably should be. This frees human minds for work requiring judgment, creativity, and personal investment that AI cannot provide.

The study's own data supports this. When participants with prior independent writing experience gained AI access, they showed enhanced neural activity. They used tools strategically to extend thinking rather than replace it.

Students are telling us something important when they automate routine academic tasks. They are saying their mental energy deserves better purposes. The researchers measured decreased effort on meaningless assignments and concluded AI reduces engagement. They should have asked whether assignments worthy of human intelligence would show different results.

Educating for a Simulated Relationship

As AI settles into classrooms, we face a peculiar challenge: not just how students use it, but how they relate to it. This isn’t a question of function or ethics, but of posture—how to engage something that responds like a person but isn’t one. The educational goal is a subtle kind of literacy: to treat AI as an interactive partner without mistaking it for a peer.

It’s a familiar dilemma, strangely enough. When children talk to imaginary friends or fictional characters, they often treat them as real companions. They know, on some level, that the character isn’t real—but they still cry when the story ends or feel comforted by a plush animal’s imagined voice. Child psychologists don’t rush to correct this confusion. Instead, they guide children to inhabit the fiction while understanding its boundaries. The fiction is developmental—it helps the child grow, not deceive.

We need a similar stance with AI. Students must learn to engage in what we might call a non-dialogic dialogue: a back-and-forth that mimics human exchange but is, in substance, interaction with an “It.” Martin Buber’s language is useful here. Procedurally, AI feels like an “I-Thou”—responsive, adaptive, present. But substantively, it remains an “I-It.” It has no inner life, no perspective, no sense of being addressed.

If we treat AI merely as a tool, we lose its pedagogical richness. If we treat it as a mind, we delude ourselves. The path forward is both instrumental and interactive: act as if the AI understands, but always know it doesn’t. This requires a new kind of mental discipline—AI mind theory, if you like. Not to imagine what AI thinks, but to restrain the impulse to imagine that it does at all.

In practice, this means teaching students to hold contradiction. To benefit from AI’s apparent collaboration, without anthropomorphizing it. To take seriously its output, without confusing fluency with insight. It’s a balancing act, but one education is well suited for. After all, school isn’t meant to tidy up complexity. It’s meant to make us capable of thinking in layers.

AI is not our friend, not our enemy, not even our colleague. It is something stranger: a fiction we interact with for real.