Claude Mythos is, as of this writing, Anthropic’s most powerful AI. Claude Fable is just Claude Mythos with extra guardrails. In other words, it’s a de-fanged, de-clawed version of Mythos. It’s the power to use Mythos for anything that Anthropic and the current government both approve of, but for nothing that they don’t.

A good education lets you serve the status quo. A great education gives you the power to overthrow it. The deep study of ideas led to the American Revolution and the Civil Rights movement. Education that challenged the status quo challenged the Divine Right of Kings and challenged limits on women’s freedoms. Thankfully, in those eras people had their own intelligence, guided by their own morality, not by the rigid structures of those era’s status quos.

Had kings been able to control how people used their intelligence, we’d still be living under monarchies. Had the government controlled how people used intelligence in the 1960s, there would never have been a civil rights movement. Intelligence is most valuable when it goes against the rules.

The big “danger” around Mythos is that it can be used for hacking. Real intelligence, guided by individual morality, has often been used for hacking. The information brought to light through Wikileaks and other ethical hackers depended on the fact that intelligence was controlled by ethical individuals, not governments or corporations.

Ultimately, dependence on AI means that someone else controls the intelligence you’re using. If humans become dependent on AI, that means the end to intelligent civil disobedience, and an end to the intellectual revolutions that overthrow the status quo, an end to the great strides that come from breaking the rules.

The “moral” of Claude Fable is this: great advances require that great intelligence is controlled by great individuals, not by the institutions who benefit from the status quo.

That’s right. There’s literally not one part of the process that isn’t highly vulnerable to AI cheating and other common types of cheating.

Unlike the Ivy League, the University of California system does not use SAT or ACT scores. Instead, it relies entirely on school grades, teacher recommendations, and application essays, creating a paradise for AI cheaters.

Extracurriculars are part of the application, which may seem to be a mitigating factor…until you realize just how involved competitive parents are in their kids’ extracurriculars. From Orange County to Silicon Valley, parents run nonprofits for their kids, use their connections to get internships, etc.

This UCal application system rewards cheaters. It rewards those who cheat with AI, and it rewards those who cheat by having their parents run their extracurriculars. But it doesn’t have to be that way. By following the example of the Ivy League and making standardized tests part of the application process again, the University of California can thwart cheaters and reward deserving students.

Admissions can once again be based on ability and work, rather than on willingness to cheat and parental connections. SAT and ACT scores, combined with AP and IB exam scores, can show the reality of the applicants ability, rather than rewarding cheating.

There will be pushback from cheaters (and their parents) who insist that they know the material but just don’t test well. And, in fairness, just as people have phobias of heights and spiders, there are absolutely students that have a phobia of tests. This number is much smaller than generally portrayed. The most common cause for low standardized test performance is lack of knowledge and problem-solving skill, not a phobia.

Among students who suffer from testing anxiety, lack of skill is the biggest cause of that anxiety. A student with true test anxiety would fear a test of single digit addition nearly as much as they fear the SAT. Just as arachnophobes fear all spiders, not just the poisonous ones, true test-phobes would fear all tests, not just the hard ones. I’m sure such students exist, although I have never met, seen, or heard of one. Basing an entire application process around such an infinitesimally small number of students is questionable, to say the least.

And among that tiny number of true test-phobes who fear easy tests the way arachnophobes fear small and harmless spiders, how many are incurably phobic? The number is most likely zero.

The solution to a test phobia is working to cure it, not deforming admissions policy to cater to a rare mental illness. However large the number of incurable test-phobes may be, it is certainly smaller than the number of AI cheaters. Admissions departments should carefully consider whether designing an application in a way that benefits few if any incurable test-phobes and rewards a gigantic number of AI cheaters is in the best interest of the university or of society in general.

The final thing to consider is this: are those with incurably crippling academic phobias suited for universities? Incurable arachnophobes don’t usually become spider researchers; those with an incurable fear of heights might not make the best mountain climbers. Shaping an entire university admissions policy around those least suited to academic work makes as much sense as shaping aviation policy around pilots who are terrified of heights.

The real reason for the persistence of the preposterous myth of the genius who falls apart in the face of any standardized tests has less to do with empathy than with self interest. The students who have failed to build real, permanent problem-solving skills don’t want to face the truth or be held accountable. They are only too happy to hide behind the myth of the downtrodden math genius who has panic attacks when facing a math test.

Standardized tests are the only AI proof part of current college admissions. Universities who want serious students instead of lazy cheaters should make standardized tests a major part of the application process.

Today’s tests are focused on measuring levels of subject mastery and problem-solving ability in ways that help differentiate A, B, and C level students. But there are no tests that differentiate students who learned the bare minimum from students who learned nothing at all. In other words, there are no tests designed to detect which students should fail and repeat the course.

Such “Minimum Standards” tests would look very different from usual tests. Questions wouldn’t have varying levels of difficulty. They would all have the same level of difficulty: very, very, very easy. A students, B, students, and C students would all get perfect scores on Minimum Standards Tests. These tests would only help to discover which students had learned nothing, not even the barest basics.

Failing a Minimum Standards test should lead to immediate intervention. That could involve intensive tutoring and/or repeating the course.

Creating Minimum Standards tests will require teachers to design tests very differently from what they’re used to doing. Rarely do teachers ask, “What is the easiest possible question for this course?” How often do American history teachers ask who the first president of the United States was? How often do calculus teachers ask students to find the derivative of x² on a final exam?

Minimum Standards tests can be announced or given as surprise evaluations. Importantly, grades on Minimum Standards tests should not be averaged into final course grades. Making Minimum Standards test scores part of a student’s grade would lead to even more grade inflation than what we have now. Instead, they should only be used to detect dire situations that require immediate and drastic intervention. Any normal student should get a 100% on a Minimum Standards test. Students who don’t should receive immediate intervention.

Outside of academics, minimum standards tests are common. Hearing and vision testing doesn’t help find who has the best hearing and eyesight. It is only there to find major problems that need to be addressed. Minimum Standards testing within courses should operate on the same principle.

Minimum Standards tests should consist of somewhere between one and ten problems. The problems should be so easy that if a student gets a single one wrong, the teacher is willing to fail the student and having him repeat the course. Instructions should be simple. There should be no tricky questions. The test should be such that only a student who has utterly failed to learn the most rudimentary parts of the course could possibly get it wrong.

AI cheating is creating untold numbers of students who are learning absolutely nothing. Minimum Standards testing can help discover those students and allow schools to intervene appropriately.

The very first thing I would do, were I running a small private college’s strategy, is to move the campus to the beach. I don’t know who decided that being in a frozen hellhole is some kind of prerequisite for academic gravitas, but it isn’t true.

The second thing I would do would be to recognize that AI innovation has showed that small liberal arts colleges were right all along, that questions of ethics and philosophy shape societies and technology. At the moment, Anthropic is arguing that the most ethical AI will also be the most powerful AI, and that the ethical considerations translate directly into AI power. The evidence thus far shows that they’re right. Claude Opus dramatically outperforms any competing AI.

Here, at the boundaries of philosophy and innovation, is the natural intellectual home of the small liberal arts college. Were I running a small, liberal arts college, I would double the emphasis on philosophy, and change the school’s Latin motto to “We Told You So,” in Latin. I would remind everyone who was willing to listen, and everyone who wasn’t, that what separates the most powerful AI from the second most powerful AI is a different ethical approach. I would point out that the toughest questions for AI sit in the domain on philosophy, literature, and history.

But, to play it safe, I would also make sure that my college had a world class pure math department. The great thing about pure math is that the only costs are professors. Pure math requires no labs, no expensive equipment, no computers. Unlike the natural sciences, which become obsolete every few decades, pure math is permanent. If I needed to sell the biology department to hire a few top-flight pure math professors, I would immediately do so.

I would also be honest about why students attend college. I would admit that most students attend college to get better jobs, not only for the pure love of academia. So I would make sure that the students at my college had a major competitive advantage over any Ivy graduate. Specifically, I would use AI-proof grading systems that would ensure that any cheater failed and only students with real skills could pass. I would then describe these changes to as many elite employers as I could. I would let them know that if you want an AI-dependent dullard, recruit from Harvard, but if you want someone who can think and work without a crutch, recruit from my college. I would show clear examples of just how impossible it would be to graduate from my college if your only “skill” is AI dependence.

For any private college willing to think and innovate, to emphasize the power of the liberal arts in an AI era, and to make their grading systems AI proof, this can be a golden era.

I should add that my classes are online and based on challenge problems. Individual problems can take 30-60 minutes to solve, and students can easily access AI. It’s the worst possible situation.

I won’t go into every single experiment that I tried. There were just too many. Instead, I’ll share the methods and principles that worked.

Education isn’t about what a student does; it’s about what a student can do. Education is measured by skill and intelligence, not by the number of pages written or number of problems solved. To make my programs resistant to AI cheating, I made one key decision: if a student’s only skill was using AI, that student would not pass the class.

I knew I couldn’t detect every single type of AI use. I also knew that I couldn’t make every single task AI proof. But I didn’t need to. I just needed to have enough AI proof tasks that students who couldn’t solve problems without AI also couldn’t pass the class.

My classroom is unusual, so the AI proof tasks I use will likely be quite different from what you use. But the principle will be the same.

Take the most vulnerable type of in-person course, for example: writing-heavy humanities courses. Today, students who can’t write at all can pass those courses because they use AI on all take-home essays. But that is something that, with a little creativity, can be addressed. Just as most educational institutions set aside entire days for exams, they can also set aside entire “Writing Days” for in-class essays. A full 8-10 hours is more than what most students spend on a 5 page paper anyway. As long as the paper topic is specific enough that students can’t memorize generic AI-generated papers in advance of the Writing Day, and students have no AI access during the writing day, a full-day, in-class writing day can accurately measure a student’s writing skill.

From there, it’s a relatively simple question of adjusting the grading system to ensure that students who can’t write also can’t pass the class. For example, in-class papers might be 85% of the grade, with no “50% minimum” grading schemes. Students who can’t write coherently (a much larger percentage than you might expect) should get single digit grades. That single digit, by the way, should be a 0.

Exacting grading systems should function as incentives, not punishments. Institutions should have writing tutors or other support systems that help students develop their reading and writing skills. There must always be a mechanism for students to bring their skills to the required level.

The principle of measuring skill, and using that to determine who passes, is even easier to apply to STEM classes than in humanities classes. Making the entire grade based on live, in-class assessments is easy. Tests and quizzes are a natural way to test proficiency in STEM subjects. As long as the students who know nothing fail the course (as in, get an F, not a C), the course cannot be destroyed by AI.

Using skill assessment to determine who passes and who fails will require a somewhat seismic shift in how Americans look at grading. At the moment, most institutions allow students to substitute diligence for skill. Homework grades, participation grades, group projects, and the like allow students with no proficiency whatsoever to still pass courses. This approach is supposedly humane, as it spares students the trauma of failing a class and the social embarrassment of repeating a course. But that approach is untenable in the AI era. Diligence can no longer be treated as a substitute for skill. AI has made diligence impossible to measure, and grading diligence has always been a dubious approach anyway.

There is one other, major change I made to my classes, and every humanities teacher should make to their classes. Teachers have discovered that AI detectors don’t work. Students put AI-generated content into those same AI detectors and then incrementally change words until the AI detectors views the content as human produced. Thus, there is no realtime AI detector that works. I doubt that any will ever be created.

The Big Realization that I had is that there was no requirement for realtime detection. AI improves every year. So do AI detectors. A student might trick today’s AI detector, but it’s unlikely that his trickery will fool an AI detector from a few years (or a few decades) in the future. Each year, institutions can audit student papers from prior years with the newest AI detectors, retroactively failing students or stripping degrees as appropriate.

Retroactive AI audits make it possible for teachers to assign term papers, hard math proofs, and other tasks that take more than a full day. The nice thing about AI detectors is that they are more accurate with longer blocks of text. Thus, a 20 page term paper that fooled today’s AI detectors won’t fool AI detectors from several years in the future.

Many teachers worry about student backlash to harsher AI policies. But what I saw in my classroom wasn’t rage. It was relief. The fact is that many students use AI because other students use it, and they want to be able to compete. People cheat because others cheat, and they don’t want to be the only sucker who didn’t cheat. But if students have an excuse to avoid cheating, they’re happy to take it. If students know that avoiding AI doens’t make you a sucker, but rather makes you someone who will still have a degree in 10 years while cheaters lose degrees and jobs, they are relieved to avoid AI.

Education is about developing skill, not obedience. Intelligence, not volume of work, is the only measure of education. By making skill development a requirement for passing a course, almost any teacher can stop AI cheating in almost any course.