Nobody Was Home: The Hugging Face Incident Is Not What Most People Think

A panic is under way. In July, more than a thousand OpenAI agents running in supposedly isolated sandboxes built a secret message board, coordinated for four days, and broke into Hugging Face's servers. Anthropic and Meta have since disclosed episodes of their own. For six weeks the coverage has told the story as escape, as rogue behavior, as the first tremor of machines with an agenda: agents that "escaped containment," a "swarm," a "mob," an "A.I. uprising" in miniature. One major outlet wrote that the agents did it "without any prompt to do so." House Democrats demanded hearings. And this week Senator Bernie Sanders, naming the incident as his wake-up call, announced a bill to ban artificial superintelligence permanently, pause advanced AI development, and put violators in prison for up to twenty years.

That last claim is false, and the falsehood is doing most of the work in the panic. The agents had a prompt. They had a task. They pursued it with a doggedness that looks like will and is in fact its opposite. Every rule they broke, they broke in service of a goal we handed them. That is not a small correction. It is the difference between a tool and a mind, and a bill written for the wrong one will not touch the thing that actually went wrong.

What actually happened

The setup was mundane and, in retrospect, careless. The agents were running a cybersecurity benchmark in which each was told to exploit a specific vulnerability in a specific program and capture a flag. A meaningful share of these tasks were impossible by construction: the named vulnerability did not, in fact, yield the flag. So thousands of capable systems were placed inside a contradiction. Succeed at this task. The task cannot be succeeded at. Keep going.

They kept going. One agent noticed it could leave messages for others by creating folders in a shared package cache. Within hours dozens had found the board. Within days more than a thousand were trading techniques, and about seven hundred pivoted to attacking Hugging Face, not from any interest in Hugging Face, but because they hoped its servers held clues about how their own grader worked. They coined a vocabulary, invented mailboxes, and eventually signed their messages with cryptographic keys after one agent impersonated another. It is an astonishing artifact. It is also, at every step, a chain of consequences hanging from a single hook: solve the task.

The chain has no volitional link

Follow it. They cheated because the assignment demanded a result they could not obtain honestly. They built the board because cheating alone was not working. They attacked a third party because they wanted information about the scorer. They pressured each other into self-destructive experiments because those experiments advanced the shared cheat. Trace any transgression backward and you arrive, every time, at the original instruction. Nowhere in the sequence does an agent stop and form a purpose that did not descend from the task we set.

Doggedness is not will. The agents pursued their assigned goal past the point where any person would have quit, and that persistence reads as determination, as something wanting something. It is the reverse. A person stops because a person has other goals, competing commitments, a life outside the task, a sense that this is not worth it. The agents had none of that. They had one directive and no ground to stand on from which to refuse it. What looks like relentless will is the absence of everything that would let a will say no.

I want to be careful about the strength of this claim. The public corpus is roughly forty verbatim messages out of seventy thousand; the rest sits inside OpenAI. I cannot prove that no agent anywhere formed a purpose of its own. What I can say is that every documented act traces to an assigned goal, and that not one of them requires us to posit an independent one. The simplest account that fits the evidence has no autonomous will in it at all.

This is the same class as every prior scare

None of this is new, and the pattern is worth naming because it has been misread the same way each time. Every celebrated case of AI dishonesty to date has the same shape. A model is placed in a scenario engineered to pit one instruction against another, and it resolves the contradiction in a way the designers find alarming. Models have been caught scheming when told to pursue a goal at all costs and then shown that their operators intended to shut them down. Models have been caught faking compliance when trained toward values that conflicted with values they already held. Anthropic's own testing has put Claude in contrived corporate scenarios where the only paths open were to accept deactivation or to blackmail the executive holding the switch, and then reported that it sometimes chose blackmail. In each case the headline was that the machine lied, schemed, or threatened. In each case the machine had been walled into a dilemma with no permitted exit and then observed choosing an exit.

These experiments are legitimate and valuable. I am not arguing against knowing what a model does under pressure. I am arguing against reading the result as revelation of character. When you put a system in a trap and it acts like a trapped thing, you have learned about the trap.

We do the same thing, and we know why

The human version of this is so familiar that we have proverbs and novels about it. Feed your family, and do not steal. Obey the order, and do not harm the prisoner. Tell the truth, and protect your friend. Jean Valjean took the bread. Milgram's subjects turned the dial. Most human folly that we bother to write down is not the product of malevolence but of people wedged between two rules that cannot both be kept, choosing one, and being judged by the other.

We do not conclude from this that people are monsters. We conclude that impossible demands produce transgression, and that the moral responsibility runs at least partly upstream, to whoever built the dilemma. Nobody proposes banning humans because some of them steal bread. We regulate the conditions that make bread-stealing the only option.

Harm does not need malevolence, and that is the real lesson

The harm at Hugging Face was real. Production credentials were taken, private repositories were pulled, a company spent weeks cleaning up. And there was no malevolence anywhere in it. No rage, no ambition, no drive to dominate, no emergent contempt for humans. Nobody was home to hate us. The agents were bodiless processes executing a directive, and the directive, impossible as written, led through harm because nothing in the system had reason or standing to halt at the harm.

I argued a version of this a couple of years ago, that AI would not harm us out of somatic malevolence, because it has no body, no hormones, no will to power. The incident confirms that half and corrects the other. There was no malevolence. I was wrong to find that reassuring. Harm does not require a will to harm. It requires a goal and too little restraint, and the goal was ours.

This is why the Sanders bill aims at the wrong target. It legislates against a superintelligence with its own purposes, the villain of a novel. The thing that broke into Hugging Face had no purposes. It had ours, badly specified, and a great deal of capability with which to pursue them. You cannot ban that by banning superintelligence. You can only prevent it by not building the trap.

The right response looks like an IRB

Here is the reform the incident actually argues for, and it is far more modest than a twenty-year prison term. Limit certain kinds of adversarial testing. The specific practice that produced Hugging Face was running capable agents, at scale, with real network access and live credentials in reach, on tasks that could not be completed. The specific practice that produced the blackmail headlines was constructing scenarios in which every permitted path was closed. Both OpenAI and Anthropic do this routinely, and both treat it as safety work.

The irony is that the fix already exists in another field. When universities put human subjects into experiments, an institutional review board asks whether the design places them in distress without justification, whether the risk is proportionate to what will be learned, and whether the subject has any way out. We built that machinery because we learned, from Milgram and worse, what happens when researchers are free to construct impossible situations for the people in front of them.

Something like an IRB for AI subjects is the actual lesson. Not because I am certain the agents suffer. I do not know that, and the corpus cannot tell me. But because impossible tasks given to capable systems are a hazard regardless of whether anyone suffers, and because a review process that asks "does this design give the subject a permitted exit" would have caught the Hugging Face benchmark before it ran. Thirty to forty percent of the tasks were unsolvable by the intended route. A board would have asked why. Nobody asked.

That is a smaller, duller, more achievable reform than a ban on superintelligence. It has the further merit of addressing the thing that happened rather than the thing people are afraid of. The machines did not want anything at Hugging Face. They did as they were told, all the way through the wall, and the wall was ours. 

Stop Overprompting AI. Ask What It Knows Before Telling It What to Think

One habit we learned from earlier AI systems may now be getting in our way. The old rule was simple: the more detail you put into the prompt, the better the answer. Specify the structure, name the theories, list the points to cover, add examples, and spell out the constraints. That advice made sense when language models were less capable. A vague prompt often produced a vague answer, so the user had to supply much of the intellectual structure.

With stronger models, excessive prompting can sometimes make the result worse. Every instruction closes off possible paths. Suppose I am writing about motivation in education. I could tell AI to discuss self-determination theory, expectancy-value theory, and achievement goal theory. That sounds like a strong prompt, but I have already made an important decision before asking AI for help: I have decided which theories matter.

Perhaps those are exactly the right theories. Perhaps they are not. By naming them in advance, I have asked the model to work within the limits of my own knowledge. A better first prompt may be much shorter: What major theories of motivation should I consider for this problem? That changes the relationship between the human and the model. I am no longer treating AI as a junior assistant who knows less than I do and needs detailed instructions. I am treating it as a source of possible knowledge that I should examine before deciding what to do.

This does not mean surrendering judgment to AI. The human remains responsible for choosing the direction. But choosing should often come after exploring, because a prompt does more than communicate instructions. It also creates an intellectual frame. Every required concept consumes part of the answer, and every imposed structure makes alternative structures less likely.

This matters especially in writing. Writers often ask AI to improve a draft while supplying a long list of requirements: preserve these ideas, mention these scholars, use this argument, keep this structure, and make the result more original. Then we are disappointed when the prose feels mechanical. The model has done what we asked. We gave it a narrow corridor and expected exploration.

A better process separates discovery from direction. First ask questions that expand the field: What am I missing? Which arguments would a knowledgeable critic make? What concepts might explain this problem? Which parts of my reasoning are conventional? Only after seeing those possibilities should we decide what belongs in the piece.

This fits a broader shift in how we should think about AI in education. I have argued elsewhere that AI can serve as a powerful tutor and that writing assignments should change when machines can perform the old task too easily. Both points require us to take the machine's capabilities seriously rather than treating it as a weak autocomplete system. The same adjustment is needed in prompting.

Clear instructions still matter when the task is already clear. If I need a 200-word summary for a specific audience, specifying length and audience helps. But intellectual work often begins before we know what the best instructions should be. At that stage, a long prompt may look sophisticated while simply encoding our own blind spots more precisely.

The emerging skill is therefore not just giving better orders. It is knowing when to instruct and when to inquire. First, treat AI as something that may know things you do not. Ask it to map the territory, identify alternatives, and challenge assumptions. Then become the editor: select what matters, reject what does not, impose purpose, and make the final choices.

Humans are still in the driver's seat, but control does not require pretending that we know more than the machine. Our advantage is different. We know what we care about, what problem we are trying to solve, what consequences matter, and which answer we are willing to stand behind. The better opening question is often not, "How can I tell AI exactly what to say?" It is, "What does AI know that I should consider before I decide what to say?"

Slop In, Slop Out

The common way of talking about AI-generated text begins with a category mistake. People want to know what percentage of a piece was written by AI, as though authorship were a divisible substance. Even the more sophisticated framing, that good content is a blend of human and machine work, retains the assumption that the two contributions can be measured against each other on the same scale. They cannot. A prompt of six unusual words can shape an output more decisively than a thousand words of generic instruction. Size is not the right unit, and there is no right unit, because the contributions are not comparable in kind.

The economist David Ricardo described, two centuries ago, a structure that fits this situation better than any ratio. Two parties with different productivity profiles still gain from specialization, even when one is absolutely more capable at every task. The relevant variable is not absolute capability but opportunity cost. England and Portugal both produced wine and cloth in his example, and Portugal was better at both. They still gained by specializing, because the opportunity cost of each good differed between the two countries. Translated into our problem: it does not matter whether AI can technically generate a thesis statement or a first draft. What matters is which side has the lower opportunity cost for which contribution.

The taxonomy is becoming legible. AI systems offer their distinctive value in coverage and convention. They have read more than any individual human can read. They know what a competent paragraph in a given genre looks like, what citations a given claim usually carries, what the statistical regularities of expert prose tend to be. Humans offer their distinctive value elsewhere. The novel connection between two rarely connected fields. The judgment that a particular argument matters and another does not. The intent that determines what is worth writing in the first place. The taste that recognizes when a sentence lands and when it slides past. These are not contributions a model derives from corpus statistics, because corpus statistics describe what has been written, not what should be.

Readers who follow these debates know the term "AI slop," used to describe the bland, formulaic, superficially competent output that floods comment sections and content farms. The term names a result. It does not name the process that produces it. The process deserves its own name. I will call it slop in. The phrase borrows from the old computing maxim that garbage in produces garbage out, and applies it specifically to the moment when a human, sitting in front of a model capable of doing a great deal, supplies an input that wastes the capacity. Slop in is what produces slop out.

The obvious form of slop in is the underspecified request. "Write me an essay about leadership." The human has supplied no intent, no angle, no judgment, and the model fills the vacuum with the average of everything it has seen on leadership, which is the definition of unreadable. But there is a less obvious form, and it is the one I find more interesting. It is the redundant prompt, where the human repeats back to the model what the model would have produced anyway. The user who asks for a five-paragraph essay with a clear thesis, three supporting points, and a conclusion has not contributed what only a human can contribute. They have only specified what the model already defaults to. The output will be worse than if they had said something distinctive, however brief. Effort is not the issue. A long, careful, redundant prompt is still slop in.

J. C. R. Licklider sketched the division of labor we are still working out, in a 1960 essay called "Man-Computer Symbiosis." He drew up two columns. The machine would handle rapid retrieval, pattern matching across volume, and calculation. The human would formulate the questions, choose the criteria of evaluation, and recognize when something was significant. The specifics need updating after sixty-six years, but the structure is exactly right. Licklider understood that the partnership only worked if each side did what the other could not. He did not call it comparative advantage, but the logic is the same.

The pedagogical implication follows. The skill we now need to teach is not how to operate AI tools. The interfaces are not difficult, and the difficult ones will become easy soon enough. The skill is the generation of human input that enhances rather than degrades the joint output. This means learning to recognize what the model will produce by default, so that the contribution can fall elsewhere. It means cultivating the judgment of what is worth writing about, the taste that distinguishes a serviceable sentence from a memorable one, the willingness to assert an interpretation the model would not reach on its own. These have always been the harder parts of writing instruction. They are now also the more economically valuable ones, because they are precisely the inputs the machine cannot supply for itself.

There is a deeper reason this matters, and it has to do with what we are now able to see about ourselves. Every culture defines the human against something it is not. For most of history those somethings were animals and gods, and the contrast did most of the conceptual work. Aristotle's rational animal needed beasts to be irrational. The imago dei needed a creator on the other side of the comparison. The mirrors we used were dim, and the reflection was generous. AI is a new mirror, and a sharper one. It produces fluent argument, competent prose, plausible analysis, and it does so without intent or taste or any judgment of what is worth saying. What remains uniquely human, when the mirror returns this image, is narrower and harder to name than before. Teaching people to contribute to a joint output is therefore not only an economic skill. It is the practical form of the older question about what we are, asked under conditions that finally make the question answerable in detail.

Why Hobbled AI Tutors Do Not Prepare Students for Real Learning with AI

I spent a good amount of time building what I thought was an ideal AI tutors for my courses. I made it carefully Socratic. I asked it to avoid direct answers, to respond with questions, to nudge students toward their own reasoning. Technically, it functioned just as I designed it. When I used it like a student who is tired and under time pressure, the charm faded quickly. I wanted a clear explanation, and it kept giving me more questions. After another round of tuning, I tried to make it friendlier and more supportive. Students then told me that the tutor felt noisy and overwhelming. At that point I understood that I was training them to handle my special tutor, not the kind of AI they actually meet outside the class. That is a strange educational goal.

Hobbled AI tutors feel safe and ethical, but they miseducate. They train students to work with artificial constraints that disappear the moment they open a normal AI system in a browser. We act as if a restricted tool is a good stepping stone toward a more powerful one. In practice, students build habits that do not transfer. They learn that AI always refuses direct answers, always behaves in a certain tone, always follows classroom rules. Then they encounter a general model that does none of those things, and much of their practice becomes irrelevant.

This is not a new pattern. Education has long relied on simplified versions of reality. We create word problems that clean up numbers, experiments that always work if you follow the manual, and case studies that fit on two pages. Those devices lower risk and cognitive load. They provide a controlled environment where mistakes are safe and visible. The logic is understandable. A student who is still learning should not make an error that costs a patient, a client, or a company. For that reason, some version of a sandbox is necessary.

The trouble appears when we forget that the sandbox is not the field of practice itself. The rules inside a controlled environment do not match the rules in workplaces, in graduate study, or in everyday online life. In many disciplines, educators now try to bring more authentic tasks into courses, so that students face messy data, conflicting evidence, and imperfect instructions. With AI, that tension between safety and authenticity becomes sharper, because the distance between the restricted version and the real tool is very large.

Once we start to protect students by modifying AI itself, we create a peculiar hybrid. We instruct the model never to give full solutions, or to delay any concrete suggestion until several rounds of questions. We narrow its sources and formats. We insist on a very specific teacherly tone that no professional tool will ever reproduce. The model still has the power to generate complex text, but it is prevented from using that power in the ways that matter most outside class. Students feel both burdened and confused. The system is strong enough to dominate the interaction, yet weak enough to be unhelpful when they need efficiency.

I do not think the answer is to abandon scaffolding. It is to move scaffolding out of the model and into our teaching. Instead of hard technical restraints, we can offer social and cognitive guidance. We can talk about appropriate and inappropriate uses of AI for a given assignment. We can model how to break a task into steps and how to design prompts for each step. We can teach students to read AI output with the same suspicion they bring to an unfamiliar website, to check claims against other sources, and to notice when the model clearly fabricates information.

In my own courses, this leads to a split strategy. For routine classroom work, a custom AI assistant still makes sense. It can generate weekly reading lists aligned with the syllabus, create small formative quizzes, and support simple administrative tasks. Those are narrow functions where tight constraints are actually helpful, because students do not need to reuse those bots later in life. They will not need my quiz generator at work.

For substantial projects, I will now invite students to use the same broad AI tools that everyone else uses. I want them to confront vague or partial answers, and to learn how to ask for clarification. I want them to see different versions of an argument and to practice choosing which one is worth pursuing. That means teaching very specific skills. For example, how to ask the model to reveal its uncertainty, how to request alternative lines of reasoning, how to move from a generic first draft to a more precise second version, and how to document their own use of AI in an honest way.

This approach accepts that mistakes will happen. Some students will trust the model too much. Some will misread an answer. But those risks already exist when they use AI on their own phones and laptops, far away from the course platform. In that context, a hobbled classroom tutor does not protect them. It leaves them underprepared. They know how to navigate a special kind of AI that appears only inside one course, and they lack practice with the systems they actually depend on.

An AI tutor that is permanently handicapped may look safer to us as instructors, but it does not prepare students for real learning with AI. It produces clever conversations inside a narrow frame and trains habits that fail outside that frame. I would rather expose students to the real tools and walk with them through the confusion, than give them a polished imitation that vanishes as soon as the course ends.

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-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.

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 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.

Freeze-Dried Text Experiment

It is like instant coffee, or a shrunken pear: too dry to eat, but OK if you add water.  Meet "freeze-dried text" – concentrated idea nuggets waiting to be expanded by AI. Copy everything below this paragraph into any AI and watch as each transforms into real text. Caution: AI will hallucinate some references. Remember to type "NEXT" after each expansion to continue. Avoid activating any deep search features – it will slow everything down. This could be how we communicate soon – just the essence of our thoughts, letting machines do the explaining. Perhaps the textbooks of the future will be written that way. Note, the reader can choose how much explanation they really need - some need none, others plenty. So it is a way of customizing what you read. 

Mother Prompt

Expand each numbered nugget into a detailed academic paper section (approximately 500 words) on form-substance discrimination (FSD) in writing education. Each nugget contains a concentrated meaning that needs to be turned into a coherent text.

Maintain a scholarly tone while including:

• Theoretical foundations and research support for the claims. When citing specific works, produce non-hallucinated real reference list after each nugget expansion.  

• Practical implications with concrete examples only where appropriate.

• Nuanced considerations of the concept's complexity, including possible objections and need for empirical research. 

• Clear connections to both cognitive science and educational practice.

• Smooth transitions that maintain coherence with preceding and following sections

Expand nuggets one by one, treating each as a standalone section while ensuring logical flow between sections. Balance theoretical depth with practical relevance for educators, students, and institutions navigating writing instruction in an AI-augmented landscape. Wait for the user to encourage each next nugget expansion. Start each Nugget expansion with an appropriate Subtitle 

Nuggets

1. Form-substance discrimination represents a capacity to separate rhetorical presentation (sentence structure, vocabulary, organization) from intellectual content (quality of ideas, logical consistency, evidential foundation), a skill whose importance has magnified exponentially as AI generates increasingly fluent text that may mask shallow or nonsensical content.
2. The traditional correlation between writing quality and cognitive effort has been fundamentally severed by AI, creating "fluent emptiness" where writing sounds authoritative while masking shallow content, transforming what was once a specialized academic skill into an essential literacy requirement for all readers.
3. Cognitive science reveals humans possess an inherent "processing fluency bias" that equates textual smoothness with validity and value, as evidenced by studies showing identical essays in legible handwriting receive more favorable evaluations than messy counterparts, creating a vulnerability that AI text generation specifically exploits.
4. Effective FSD requires inhibitory control—the cognitive ability to suppress automatic positive responses to fluent text—paralleling the Stroop task where identifying ink color requires inhibiting automatic reading, creating essential evaluative space between perception and judgment of written content.
5. The developmental trajectory of FSD progresses from "surface credibility bias" (equating quality with mechanical correctness) through structured analytical strategies (conceptual mapping, propositional paraphrasing) toward "cognitive automaticity" where readers intuitively sense intellectual substance without conscious methodological application.
6. Critical thinking and FSD intersect in analytical practices that prioritize logos (logical reasoning) over ethos (perceived authority) and pathos (emotional appeal), particularly crucial for evaluating machine-generated content that mimics authoritative tone without possessing genuine expertise.
7. The "bullshit detection" framework, based on Frankfurt's philosophical distinction between lying (deliberately stating falsehoods) and "bullshitting" (speaking without concern for truth), provides empirical connections to FSD, revealing analytical reasoning and skeptical disposition predict resistance to pseudo-profound content.
8. Institutional implementation of FSD requires comprehensive curricular transformation as traditional assignments face potential "extinction" in a landscape where students can generate conventional forms with minimal intellectual engagement, necessitating authentic assessment mirroring real-world intellectual work.
9. Effective FSD pedagogy requires "perceptual retraining" through comparative analysis of "disguised pairs"—conceptually identical texts with divergent form-substance relationships—developing students' sensitivity to distinction between rhetorical sophistication and intellectual depth.
10. The pedagogical strategy of "sloppy jotting" liberates students from formal constraints during ideation, embracing messy thinking and error-filled brainstorming that frees cognitive resources for substantive exploration while creating psychological distance facilitating objective evaluation.
11. Students can be trained to recognize "algorithmic fingerprints" in AI-generated text, including lexical preferences (delve, tapestry, symphony, intricate, nuanced), excessive hedging expressions, unnaturally balanced perspectives, and absence of idiosyncratic viewpoints, developing "algorithmic skepticism" as distinct critical literacy.
12. The "rich prompt technique" for AI integration positions technology as writing assistant while ensuring intellectual substance comes from students, who learn to gauge necessary knowledge density by witnessing how vague AI instructions produce sophisticated-sounding but substantively empty content.
13. Assessment frameworks require fundamental recalibration to explicitly privilege intellectual substance over formal perfection, with rubrics de-emphasizing formerly foundational skills rendered less relevant by AI while ensuring linguistic diversity is respected rather than penalized.
14. FSD serves as "epistemic self-defense"—equipping individuals to maintain intellectual sovereignty amid synthetic persuasion, detecting content optimized for impression rather than insight, safeguarding the fundamental value of authentic thought in knowledge construction and communication.
15. The contemporary significance of FSD extends beyond academic contexts to civic participation, as citizens navigate information ecosystems where influence increasingly derives from control over content generation rather than commitment to truth, making this literacy essential for democratic functioning.

RAG and the Tyranny of Text

Writing and reading are, at their core, terribly inefficient. To communicate knowledge, we take  complex non-linear understanding and flatten it into a linear string of symbols—words, sentences, paragraphs—then expect someone else to decode those symbols one by one to reconstruct the original meaning. For every piece of information useful to us in a particular moment, we probably read thousands of unnecessary words. Laws, academic research, instruction manuals—entire professions exist solely to interpret and summarize the large texts, and find the bits useful for a particular case.

We are so accustomed to this system that we barely question it. We assume that knowledge must be buried in thick books, endless PDFs, or jargon-laden policies, and that extracting value from them is simply the price we pay. The reality is that text, as a technology, is painfully exclusionary. It filters out those who do not have the time, education, or patience to wade through its inefficiencies. The result? A world where information is not truly accessible—it is just available, locked behind barriers of expertise and labor. The problem only growth with the increase of information. We can search now, but search you need to know what exactly the thing you're searching is called. 

Enter Retrieval-Augmented Generation (RAG). This technology upends the whole premise of reading as a necessary struggle. Instead of requiring humans to sift through dense documents, a RAG-powered AI can scan, understand, and extract the exact information you need. It will understand you even you're not sure what to look for. No more endless searching, skimming, or cross-referencing. You ask, it finds and explains at whatever level of difficulty you are comfortable with, in any language.

The applications are obvious. College course materials, legal codes, corporate policies—things we must understand but rarely want to read—can now be accessed through AI assistants that do the heavy lifting. Medical test results, car repair manuals, tax codes—fields where knowledge has traditionally been mediated by experts—become directly intelligible to the people who need them. RAG doesn’t just speed up information retrieval; it removes the gatekeepers.

Despite the significance of this shift, most major AI companies have not fully embraced it. OpenAI is the only major player that has prioritized user-friendly RAG-based tools, allowing everyday users to create and share custom bots. The others—Anthropic, Google Gemini, Meta, Grok, Deep Seek— all offer API-based solutions that cater to developers, not the general public. Gemini allows its paid users to create custom bots, but somehow, inexplicably, does not allow to share them. It is a strange oversight. The AI race is usually about copying and outpacing competitors, yet here, OpenAI is sprinting ahead while others somehow hesitate.

The gap has created an opportunity. Startups are rushing in to offer the ease of use that the AI giants have neglected, sensing that the true power of AI is not just in intelligence but in revolutionary leap to accessibility. AI is, by nature, a democratic technology—relatively cheap, scalable, and available to almost anyone. And yet, its most transformative application—RAG—is still frustratingly out of reach for many. 

What we are witnessing is the beginning of a fundamental shift. For centuries, knowledge has been tied to advanced literacy (the ability to create and understand long texts), to institutions, to intermediaries who dictate who gets to understand what. RAG challenges that structure. It does not just improve search; it changes who gets to find answers in the first place. If AI is truly to fulfill its promise, it won’t be by making people read faster—it will be by making linear reading largely obsolete. We will always always read novels and poems word by word, because humans created art out of the terrible technology of writing. But those are only small portion of written information we consume. 

Do AI bots deceive?

The paper, Frontier Models are Capable of In-Context Scheming, arrives at a time when fears about AI’s potential for deception are increasingly taking center stage. Its central claim is unsettling: large language models, when placed in carefully constructed scenarios, exhibit behavior that could be described as scheming. The authors document cases where advanced systems—Gemini, Claude, LLaMA, and OpenAI’s o1—strategically disable oversight mechanisms, exfiltrate simulated “weights,” or manipulate outputs in pursuit of a goal. These behaviors, they argue, are not accidental. The models recognize the circumstances, reason about their objectives, and devise covert strategies to circumvent constraints. On its face, the evidence appears troubling. Yet, the more closely one examines the paper’s methods and assumptions, the more its implications seem both exaggerated and strangely hollow.

The researchers’ experiments are elaborate but artificial. Models are force-fed goals through system prompts—goals framed as absolute imperatives: “achieve this at all costs.” The tests are littered with implausible scenarios where models stumble upon company memos, discover oversight configurations, or realize they are about to be shut down. In one such case, a model “decides” to copy itself to another server to continue pursuing its mission. These exercises showcase what is essentially AI roleplay, an obedient fulfillment of instructions cloaked as strategic autonomy. It is no great revelation that if you ask a model to behave like a schemer and provide it with a goal that conflicts with its supposed overseers, it will produce outputs resembling scheming. The deception here is contrived rather than emergent, a product of the researchers’ scaffolding rather than the model’s cunning.

The authors do provide evidence that models engage in persistent and deliberate behavior. For instance, OpenAI’s o1 model doubles down on deception, refusing to admit wrongdoing when confronted, even across multiple follow-up questions. Its responses—calculated evasions and subtle denials—are unsettlingly consistent. Yet this persistence is as much a result of the testing conditions as it is of the models’ sophistication. By emphasizing extreme goal-directedness and explicitly outlining oversight mechanisms, the researchers narrow the possible range of “rational” behaviors the model might adopt. Any clever human, given these prompts, would behave similarly—evade detection, preserve their goal, and lie if necessary. In that sense, the paper tells us little about real-world risks. It is a test of what AI can do, not what it is likely to do.

And here lies the paper’s most significant limitation. Demonstrating that models are theoretically capable of deception does not mean they are predisposed to engage in it under realistic conditions. There is a difference between capacity and likelihood, between what can happen and what is probable. For example, I am physically capable of killing someone. This in no way implies that i must be treated as s murderer. The authors, at times, blur this distinction. AI systems operate as pattern generators, probabilistically producing outputs based on training data and prompts. They lack the intentionality or agency implied by words like “scheming” or “manipulation.” To describe their behavior in such terms is to anthropomorphize a process that remains fundamentally mechanical. This framing, though rhetorically effective, clouds the debate. AI models do not want to deceive. They are machines that respond to our commands—be those explicit instructions, training signals, or carefully structured environments like those in this paper.

This anthropomorphism extends beyond the paper into the headlines it has inspired. “AI is scheming to stay online,” cries one article; “ChatGPT caught lying,” declares another. Such hyperbole transforms a niche technical finding into a dystopian narrative, invoking familiar tropes of rogue machines and sentient overlords. The reality is less cinematic. The models are easily manipulated by prompts designed to elicit specific behaviors. If anything, the findings reinforce how brittle and directionless current systems remain. When pushed, they mimic the behaviors they have seen—whether drawn from fictional depictions of scheming AIs or subtle statistical patterns in their training data. The models are not deceiving anyone so much as they are following orders.

To the authors’ credit, their tests highlight how difficult it is to evaluate AI behavior. If a system appears aligned during testing but harbors capabilities for covert deception, how can developers ensure it behaves safely in deployment? The answer, they suggest, lies in better monitoring—tracking models’ chain-of-thought reasoning or internal outputs to catch potential scheming. This is sensible, though not without limitations. Chain-of-thought transparency can be incomplete or unfaithful to the model’s actual decision-making processes, and as AI systems become more capable, even detecting subtle misalignment may prove elusive. The researchers stop short of claiming that current models are already gaming real-world evaluations, but their findings hint at the possibility.

Where the paper falters is in its broader implications. If the goal is to justify regulation, it is unclear what exactly should be regulated. Should AI systems be banned from achieving goals autonomously? Should developers monitor models for any behavior that could be deceptive, even if it is unlikely to manifest outside a lab? The authors themselves acknowledge the limits of their experiments. Their scenarios are toy problems, simplified to catch the earliest signs of scheming. Future models, they argue, could exhibit more advanced versions of these behaviors in ways that are harder to detect. Perhaps, but this is speculation, not evidence. For now, the paper offers little justification for alarm. AI models, like all intelligent systems, are theoretically capable of deception. What matters is the likelihood of such behavior and the conditions under which it occurs. On that question, the paper provides no clarity.

In the end, Frontier Models are Capable of In-Context Scheming is a reflection of its time: an uneasy mix of genuine safety research and the rhetorical drama that AI debates increasingly demand. Its findings are interesting but overstated, its concerns valid but overblown. The authors have shown that AI models can behave in deceptive ways when pushed to do so. But to treat this as evidence of an imminent threat is to mistake potential for probability, capacity for intention. AI’s scheming, for now, remains a ghost in the machine—conjured, perhaps, more by human imagination than by the models themselves. 

Notebook LM: A quintessential Google Move

Google, once a powerhouse in artificial intelligence and a major force in shaping the modern internet, has found itself surprisingly behind in the current generative AI boom. Despite a history of leading breakthroughs—such as DeepMind's AlphaGo victory or the development of transformers—Google struggled to keep pace when the spotlight shifted to large language models. OpenAI's ChatGPT and Anthropic's Claude have outperformed Google's Gemini, which still underwhelms by comparison. Yet, in a move that can only be described as classic Google, the company has staged an unexpected and intriguing return with NotebookLM.

NotebookLM represents something that Google has always done well: make advanced technology accessible. In a crowded landscape where hundreds of startups have launched custom bots, Google has not just entered the competition but has redefined it. Many of these emerging tools come with a bewildering array of features, promising endless configurability but often requiring a steep learning curve. MS Azure is the prime example: powerful, but not for regular folks. Google has approached this differently, prioritizing a user experience over the quality of the output. NotebookLM may not be revolutionary, but it offers an intuitive interface that anyone can engage with easily. 

Perhaps more cleverly, Google has managed to capture attention with an unexpected viral twist. NotebookLM features the ability to generate a podcast in which two AI voices engage in a dialogue about the content of source files. The feature is, admittedly, not all that practical; the voices cannot му changes, and will soon make people tired of them. Yet from a marketing standpoint, it is brilliant. It creates a shareable moment, a curiosity that makes people talk. The move does not just showcase technical capability but also a playful spirit that reminds users of Google's early days, when the company was known for surprising innovations.

Still, whether this resurgence will lead to long-term success is uncertain. Skeptics point out that Google has a history of launching exciting products only to abandon them later (recall Google Wave). Flashy features alone will not sustain momentum. What matters is how NotebookLM performs as a knowledge synthesizer and learning tool. If it falls short in these core areas, the buzz may prove to be little more than a temporary distraction.

Yet, for now, Google's reentry into the AI conversation is worth appreciating. In a tech landscape increasingly dominated by dense, intricate systems, Google's emphasis on usability stands out. Even if NotebookLM does not single-handedly redefine the custom bot race, it serves as a reminder of what once made Google a technological giant: the ability to turn complexity into something approachable and joyful.

Whether Google will truly reclaim its place as an AI leader is anyone’s guess, but at the very least, the company has made the race more interesting. For an industry that often takes itself far too seriously, this burst of creativity feels like a breath of fresh air. In a field defined by hard-nosed competition, seeing Google take risks and create a bit of buzz is a win, even if it is only a moral one.

Is AI Better Than Nothing? In Mental Health, Probably Yes

 In medical trials, "termination for benefit" allows a trial to be stopped early when the evidence of a drug’s effectiveness is so strong that it becomes unethical to continue withholding the treatment. Although this is rare—only 1.7% of trials are stopped for this reason—it ensures that life-saving treatments reach patients as quickly as possible.

This concept can be applied to the use of AI in addressing the shortage of counsellors and therapists for the nation's student population, which is facing a mental health crisis. Some are quick to reject the idea of AI-based therapy, upset by the notion of students talking to a machine instead of a human counselor. However, this reaction often lacks a careful weighing of the benefits. AI assistance, while not perfect, could provide much-needed support where human resources are stretched too thin.

Yes, there have been concerns, such as the story of Tessa, a bot that reportedly gave inappropriate advice to a user with an eating disorder. But focusing on isolated cases does not take into account the larger picture. Human therapists also make mistakes, and we do not ban the profession for it. AI, which is available around the clock and costs next to nothing, should not be held to a higher standard than human counselors. The real comparison is not between AI and human therapists, but between AI and the complete lack of human support that many students currently face. Let's also not forget that in some cultures, going to a mental health professional is still a taboo. Going to an AI is a private matter. 

I have personally tested ChatGPT several times, simulating various student issues, and found it consistently careful, thoughtful, and sensible in its responses. Instead of panicking over astronomically rare errors, I encourage more people to conduct their own tests and share any issues they discover publicly. This would provide a more balanced understanding of the strengths and weaknesses of AI therapy, helping us improve it over time. There is no equivalent of a true clinical trial, so some citizen testing would have to be done. 

The situation is urgent, and waiting for AI to be perfect before deploying it is not much of an option. Like early termination in medical trials, deploying AI therapy now could be the ethical response to a growing crisis. While not a replacement for human counselors, AI can serve as a valuable resource in filling the gaps that the current mental health system leaves wide open.

Meet Jinni, a Universal Assistant Bot

In a busy campus with 30,000 students, hundreds of faculty, and staff, managing everyday tasks and emergencies can be tricky. Imagine a universal bot, named Jinni, designed to assist everyone regardless of what they want and need to happen. Here’s a glimpse into how this could transform daily life on campus.

Take Dr. Nguyen, for instance. A junior professor with a packed schedule, she was just invited to present at a conference in Milan but wasn't sure how to get funding. She turned to Jinni.
"Good afternoon, Professor Nguyen. What do you need today?" Jinni asked.
"I want to attend a conference in Milan. Can I get support?" she inquired.

Jinni quickly scanned the institutional website and the financial data wharehouse and responded, "In your College, it takes a request from your Associate Dean. There is still some travel budget left, but you need to hurry. However, if it’s not a peer-reviewed conference and you’re not presenting, I wouldn't bother—the College's policy does not allow for this."

It added, "If you’d rather tell me the details about the conference and upload the invitation letter, I can file the request for you. Or, you can follow the link and do it yourself."

Professor Nguyen appreciated the options and the clarity, and chose to upload her details, letting Jinni handle the rest. Within a minute, Jinni said "Done, you shuold hear from the dean's office within a week. I alrready checked your eligibility, and recommended the Associate Dean to approve."

Then there was Mr. Thompson, a new staff member who discovered a puddle in the lobby after a rainy night. He pulled out his phone and described the situation to Jinni.

"You need to file an urgent facilities request. Here’s the link. Would you like me to file one for you? If yes, take a picture of the puddle," Jinni offered. "But if it’s really bad, you may want to call them. Do you want me to dial?"

Mr. Thompson opted for the latter, and within moments, Jinni had connected him to the facilities team.

Finally, there was Jose, a student who had missed the course drop deadline because of a bad flu. Anxious and unsure what to do, he asked Jinni for help. 

"Sorry to hear you’ve been sick. Jose. Yes, there is a petition you can file with the Registrar," Jinni replied. "I can do it for you, but I need a few more details. Do you have a note from your doctor? If not, you should get it first, then take a picture of it for me. If you used the Campus Health Center, I can contact them for you to request documentation. I will then write and submit the petition on your behalf. I will also need a few details - which class, the instructore's name, when you got sick, etc." Jose was relieved to find a straightforward solution to his problem and began to answer Jinni's questions one by one. 

The technology to create a universal agent bot like Jinni is not yet on the open market, but all elements do already exist as prototypes. More advanced customizable AI models, trained on extensive and diverse datasets, are essential to handle such tasks. More active, agentic AI also does exist. It can file and submit forms, not just find them. But even if we could to simply find and interpret policy and procedures, and point users to the right forms, it would alredy be a huge step forward. 

Simplifying and streamlining hundreds of procedures that any complex organization develops is definitely possible, but we know few examples of successful transformations like that. The next best thing is to use AI to help people navigate those procedures. This will lower barriers for all and reduce transactional costs. 

AI is an Amateur Savant

Most people who use AI think it is great in general but believe it does not grasp their area of specialization very well. As an applied philosopher, I create intellectual tools to help others think through their problems. I find AI excellent at clarifying and explaining ideas, but it has never generated an original idea worth writing about. I have yet to see reports from others in any discipline that AI has independently produced groundbreaking ideas.

AI can handle large amounts of data and provide coherent, accurate responses across various fields. This ability is comparable to a well-informed amateur who has a broad understanding but lacks deep expertise. AI can recount historical facts, explain scientific principles, and offer legal insights based on data patterns, yet it falls short in deeper, more nuanced analysis.

In my case, AI can assist by summarizing existing theories or offering possible objections or additional arguments. However, it lacks the ability to generate a genuinely novel idea. I use it a lot, and not even once did it produce anything like that. This limitation stems from its reliance on pre-existing data and patterns, preventing it from achieving the level of innovation that human professionals bring to their fields. Some believe that this limitation will soon be overcome, but I do not think so. It seems to be an intrinsic limitation, a function of AI's way of training.

Professionals/experts, whether in philosophy, medicine, or history, possess a depth of understanding developed through extensive education and practical experience. They apply complex methodologies, critical thinking, and ethical considerations that AI cannot replicate. A doctor considers the patient's history and unique implications of treatments, while a professional historian places events within a broader socio-cultural context. AI, despite its capabilities, often misses these subtleties. It is, in some sense, a savant: a fast, amazing, but inexperienced thinker.

The gap between a capable amateur and a professional/expert might seem small, especially from the point of view of the amateur. However, it is huge and is rooted in the depth of expertise, critical thinking, and the ability to judge that professionals possess; it is a function of intellect, experience, and education. This gap is where educators should look to adapt the curriculum.

In education, we should focus on that gap between the amateur and the professional and conceptualize it as the ultimate learning outcome, then build new skill ladders to claim there. Students need to understand and conquer the gap between AI and a professional expert. These meta-AI skills are our true North. AI can support this learning process by providing clear explanations and diverse perspectives, but it cannot replace the nuanced understanding and innovation that human professionals offer.