This article was produced in collaboration with Court Watch, an independent outlet that unearths overlooked court records.

The student cofounders of an AI studying tool won a $10,000 entrepreneurship prize from Emory University for their idea, were championed publicly and repeatedly by the university’s business school for creating the software, and then were promptly suspended by the school for a semester for building exactly what the school had just given them money to build.

The students were suspended by the school’s Honor Council because their AI tool “could be used for cheating” and because they connected it to a software platform used by the university to host course reading material, homework, and other assignments without obtaining express permission, though this feature was mentioned at the competition it won $10,000 at. But the school’s Honor Council did not actually find evidence that it was ever used to cheat, and a review of the Honor Council’s writeup shows an incredible misunderstanding of how the specific tool, called Eightball, was designed and a misunderstanding of how large language models are trained and what they can do.  

“While nothing about Eightball changed, Emory’s view of Eightball changed dramatically,” a lawsuit filed by Benjamin Craver, one of the suspended students against the university reads. “Emory concedes that there is no evidence that anyone has ever used Eightball to cheat. And to this day Emory advertises Eightball as an example of student innovation and entrepreneurship.”

This whole embarrassing saga, revealed in the lawsuit, is another piece of evidence that demonstrates universities and schools more broadly have absolutely no idea how to deal with AI in an academic context and have a misunderstanding of the technology. We have seen mixed messaging from schools about whether or not students should use ChatGPT in any context, have seen students who have been falsely accused of using AI to write essays, and have seen disturbing cases where students use AI to make nonconsensual porn of each other. In each of these instances, it is becoming clear that schools do not know what the rules about AI should be and that they are often being made up on the fly. 

There is a lot of detail in this article about what was ultimately an Honor Council proceeding at a private university, but the Kafkaesque proceedings in this case are instructive because they demonstrate how wholly unprepared schools and universities are to deal with even straightforward issues involving AI.

Last spring, the students presented Eightball at the university’s “Entrepreneurship Summit” and were given a $10,000 grand prize to build and launch their software, which allowed students to upload PDFs of course readings, syllabuses, and other material and turn those into practice tests and flash cards. They also explained that they were eventually going to allow users to connect to Canvas, which is a software platform used by the university where professors upload course readings, documentation, assignments, etc, the lawsuit alleges. “By connecting Eightball to Canvas, students would be able to import their course materials to Eightball all at once rather than uploading the same documents individually.”

“Eightball is a platform kind of like ChatGPT but trained directly on your Canvas courses. The way Eightball works is it connects to your Canvas and goes through each of your courses. And for each course it studies the modules, the lectures, the slides, the readings, everything. From there, it becomes a ChatGPT-like experience, but the AI is customized for your course,” one of the creators explains in a demo video. The student then shows that Eightball surfaces directly relevant passages and serves as, more or less, a search-engine for class material.

“Dorm Room Entrepreneur,” the headline of an article on Emory University’s website that was live until I asked the university for comment for this story. “Student co-founds AI-Powered Study Tool Eightball.” The article explains how three students created Eightball, and notes that some professors began recommending that their students use it to help them study for tests.The school promoted this article, and the students’ business and AI tool, in LinkedIn posts: “Emory students are using AI to improve the studying experience!,” one post by the business school reads. The student lawsuit also includes numerous emails sent to the cofounders by professors and faculty at Emory, who said things like “I was very pleased to hear about your startup, Eightball. I congratulate you for your entrepreneurial attitude and for the very interesting idea that you and your business partners had,” and “it looks great.” Andrea Hershatter, the associate dean of Emory’s business school, sent an email introducing the students to a potential outside investor and said “I hope you are having a wonderful summer and finding time and resources to continue your work on EightBall.”

An email the students sent to the team that gave them the $10,000 prize explained their plans to connect to Canvas: “To reiterate, Eightball essentially just shows you students’ materials from their Canvas courses - sort of like an advanced search inside Canvas, and is not capable of solving complete homework problems or writing essays or anything of [the] sort,” the email says. 

It is not clear, exactly, what changed at Emory that made the university take action against a startup that it went out of its way to promote, but both the lawsuit and the Honor Council writeup asserts that the university’s IT department was angry that the company allowed students to connect their own Canvas API tokens to the app. In the lawsuit, the students’ lawyers write that the university changed the settings within Canvas and “hid the button that generates Canvas [API] tokens, but it did not inform [the students] that the change was in response to Eightball’s newly available method for uploading course materials.” Soon after this, “Emory informed [one of the students] that he may have violated Emory’s Undergraduate Code of Conduct by Connecting Eightball to Canvas.” The students shut Eightball down at this point.

After all of this promotion, the university’s Honor Council launched an investigation into the students and Eightball. This investigation, which can be read here, found that Eightball had not been used for cheating, and that the students had not lied about the capabilities of the software. It also did not dispute that the school both funded and championed the software. The council recommended that the students be suspended for a year, anyway. Jason Ciejka, the director of the school’s honor council, wrote “this case is unprecedented in terms of its scale and potential to harm the Emory community.” 

School officials suggested that students choosing to use their own API tokens in the way that they are intended to be used by Canvas the company was a massive security risk. 

The school “figured out that the Eightball program accesses the Canvas data through the Canvas user generated token, which is essentially users’ Emory credentials that give full access to everything users can access on Canvas. This user generated token is considered a highly restricted user credential tool and sharing it to any outside party is a violation of Canvas terms and IT policies.” API tokens are sensitive, but API tokens exist exclusively for users to connect accounts to outside services—what the Honor Council is describing is essentially the only use for an API token, and is a feature of Canvas which the Honor Council wrote “is not something that they can turn off.” Canvas’s own documentation explains to students how they can use use API tokens to connect their accounts to other apps: “Access tokens provide access to canvas resources through the Canvas API. Access tokens can be generated automatically for third-party applications or created manually.”

IT attempted to hide this feature, but students found a workaround and “were continuing to circumvent it to generate tokens.” Because of this, they were also accused of “rewriting code to circumvent an IT security measure” (the “workaround” involved right clicking on the Canvas website, clicking “Inspect,” and copy-pasting a code snippet to generate the tokens.)

One of the witnesses the Honor Council called said that “this application was being marketed through various Reddit posts as a ChatGPT for Canvas,” and then said “From the security perspective, people cannot give full access of their data to someone else. The fact that it was OpenAI which got the access made the case even worse because OpenAI is trained on data. Once people feed it with these personal data, it could answer others’ questions based on searching through all these materials, potentially leading to copyright problems etc.” But the students explained that their program “did not use ChatGPT at all.” 

The students were also accused of “disseminating course material” by allowing students to use their own, individualized Canvas API tokens to connect Eightball directly so they did not have to upload PDFs to the tool themselves. “The upgrade did not change what students could upload to Eightball or what learning materials Eightball could produce upon request,” the suspended students’ lawyers note in the lawsuit.

Despite all of this talk about Canvas, the Honor Council’s report and writeups of what happened to the students makes clear that the university was very concerned about “the cheating potential with this program,” and the students were accused of cheating,” “plagiarizing,” and “intentionally helping or attempting to help another person to violate any provision of this Honor Code.” 

According to Eightball’s marketing, the lawsuit, and Emory University’s own writeups, Eightball was not actually a cheating tool. As far as AI-tools go, it seems innocuous, and the university did not provide any examples of the tool ever being used for cheating. “Unless answers are directly in the course materials, Eightball cannot make up anything for non-existing answers.”

The Honor Council wrote “the fact that Emory gave them the grant implies that Emory was supporting them. While the Honor Council can understand this position taken by the student, we fundamentally disagree that this places the onus on the university to ensure the ethical development of this tool. All students should carry the Honor Code and the value of academic integrity as their leading principle. Moreover, this application has already reached other educational institutions and has the potential to create widespread cheating across colleges and universities that were unaware of this flaw in Canvas.”

Emory University declined to comment on this story. Craver's attorneys declined to comment for this piece.

A control system is a mechanism — mechanical, biological, or otherwise — that forces a measure towards a reference. One example is a thermostat. You set the desired temperature of your house to 73 degrees Fahrenheit, and the thermostat springs into action, to get its reading to 73 °F or die trying.

The usual assumption is that a control system works like a target, and tries to correct deviations from that target. Take a look at the simplified diagram below. In this case, the control system is set to the target indicated by the big arrow, at about 73 °F. Since control is less than perfect, the temperature isn’t always kept exactly on target, but in general the control system keeps it very close, in the range indicated in blue.

However, there are other ways to design a control system. 

One way is to make a single-headed control system, that has a reference level, and simply keeps the measure either above or below that level. For example, this single-headed control system is designed to keep the temperature above 70 °F:

This is how early thermostats worked, and how many still work in practice. They do nothing at all until the temperature drops below some reference level, at which point they turn on the furnace, driving temperature upwards. Once the temperature returns above the reference level, the furnace is switched off. Barring any serious disturbances, this keeps the temperature in the range indicated in blue. 

This works fine if your house is in Wales or in Scandinavia, where things never get too hot. But what if you want to control the temperature in both directions? 

Easy. You just add a second single-headed control system on top of the first one, controlling the same signal in the opposite direction. This is a double-headed control system, that keeps the signal between two reference values: 

One “head” kicks in if the temperature gets too low, and takes corrective actions like turning on the furnace. The other kicks in if the temperature gets too high, and takes corrective actions like turning on the air conditioning. Together they form a larger control system that, barring any damage or huge disturbances, keeps the temperature in the range indicated in blue.  

(Both “single-headed” and “double-headed” are terms of our own invention. There may be official terms for these concepts in control engineering. If so, we haven’t been able to find them. We would love to hear if there are existing terms, please let us know!)

There is some reason to think that biological control systems in animals are mostly double-headed. This is due to the fact that these control systems are built out of neurons, and neural currents are in units of frequency of firing. Unlike other signals, frequency of firing can’t be negative: the number of impulses that occur in a unit of time must be zero or greater.[1]

Obesity

The current scientific consensus on obesity (link, link, link, link, link) is that it is the result of a problem with the control system(s) in charge of regulating body fat, the set of systems sometimes called the lipostat (lipos = fat). 

We can explore this idea through a few examples. For the purposes of illustration, let’s use BMI for our units. BMI isn’t perfect as a measure — obviously your nervous system doesn’t actually measure its weight by calculating BMI — but it’s a simple and familiar number that will do the trick. In general we should make it clear, all the following examples are greatly simplified. In reality, the body seems to have many control systems to regulate body weight, not just one. 

For starters, we know that the lipostat can’t be single-headed, because with ready access to food, people don’t generally starve to death, nor do they become fatter and fatter until they burst. 

Clearly body weight is controlled in both directions. This means it’s a double-headed system. One part of the lipostat keeps you from getting thinner than a certain threshold. And another, separate part of the lipostat keeps you from getting fatter than a different threshold.

On to the examples. A person with a healthy lipostat would look something like this: 

The two heads are set to different points, leaving a bit of room between the upper and the lower thresholds. This person’s weight can easily wander between BMIs of about 20 and 23, pushed around by normal behavior. But if they go above that upper limit, or below the lower limit, powerful systems kick into play to drive their weight back into the blue range between the two heads of the system.

What about someone whose lipostat is not healthy, someone who has become obese? One way for this to happen is for both heads to be pushed to higher thresholds, like so:

Here you can see that the upper head has been set to a BMI of about 35, and the lower head to a BMI of about 31. As before, their weight is mostly free to wander between those two levels. If they’re trying to lose weight, they can probably push their BMI down to 31. But it will be very hard to push it past that point, since the lipostat will resist them vigorously. After all, the lower limit is designed to keep us from starving to death, so it has a lot of power behind it. 

On the other hand, this person basically doesn’t have to worry about their BMI climbing above 35, since the upper limit is also defended. As long as their lipostat isn’t disrupted any further, they will remain within that range.

However, the heads don’t have to move together. They are at least somewhat independent systems, with separate set points. So another way to become obese is like this: 

This person still has a lower limit of BMI 20, just like the healthy person in the first example. But they have an upper limit of BMI 35, as high as than the obese person in the second example! 

This person is sometimes obese. On the one hand, unlike a person with a healthy lipostat, there’s nothing to keep this person’s weight from drifting up to a BMI as high as 35. So if they’re not “careful”, if they eat freely and without particular attention, sometimes it will.

But on the other hand, there’s nothing keeping this person from driving their BMI as low as 20, by doing nothing but eating less and exercising more. They don’t risk hitting a starvation response until they are well into the healthy BMI range, so they have little difficulty losing weight when they want to.

Lots of people find it really hard to lose weight. But you also encounter a lot of people who say things like, “when I was overweight I just decided to lose some weight, counted calories for a while, and made it happen, and it wasn’t that hard.” The double-headed model may explain the difference. Calorie-counters who sometimes drift upwards but can easily lower their weight on a whim have an altered upper threshold but a healthy lower threshold, while everyone else has had both their upper and lower thresholds pushed to obese new set points, and they face massive biological resistance when they try to return to a lower BMI.

Slightly Complicated

Our friend and colleague ExFatLoss likes to describe obesity as a slightly complicated problem. No one has solved obesity yet, but it doesn’t seem totally chaotic, so maybe there are just a few weird things that we’re missing. We agree that this seems likely, and one way that obesity could be slightly complicated is if different things are causing changes to the thresholds of the upper and lower heads of our lipostats.

To take a traditional example, perhaps eating lots of sugar raises your upper threshold, and eating lots of fat raises your lower threshold. In this model, if you eat lots of sugar but not lots of fat, your weight might drift up, but you can still control it. If you eat lots of fat, your weight is pushed up and can’t be pushed back down.

To take an example that seems more plausible to us, maybe one contaminant raises the upper threshold of your lipostat, and a different contaminant raises the lower threshold. Perhaps phthalates raise your upper threshold. This wouldn’t be very noticeable by itself, because you could still control your weight with diet and exercise. But maybe on top of that, exposure to lithium raises your lower threshold. This would keep you from pushing your weight back down. In combination, exposure to both contaminants would force you into obesity. (We should stress that this is a hypothetical, we have no idea whether these particular contaminants affect one head, or both, or neither.) 

So much for things being slightly complicated. One way that obesity could be very complicated is if there are not just two heads, but lots of them, maybe dozens. This is almost certainly the case. Biology tends to be massively redundant, so the most likely scenario is that the body has several different ways of measuring your body fat, and each of these measures probably has its own control systems. So you probably have many “upper” and “lower” thresholds, all interacting. It might look something like this:   

In this case, there are five heads making for five thresholds. The black thresholds have been forced wide open, defending a healthy lower BMI but a pretty high upper BMI. The red threshold is an additional lower defense, trying to keep BMI above 21. And the white thresholds are fixed to defending a range that’s solidly overweight to obese. This person is most likely to end up somewhere in the range that’s darkest blue, but could see movement all over the place. They won’t face serious resistance unless they try to push their BMI above 35 or below 20. But anything that raised the set point for that red threshold or the bottom black threshold would seriously limit their ability to stay lean.

Again, even this more complicated example is probably an oversimplification. While these models are good for illustration, real biology almost certainly involves more than 5 heads, defending lots of different thresholds in many different ways. 

There is at least one other way in which a person could become obese. As before, you could set the lower limit quite high, say to keep a person’s BMI above 31. Then you could set the upper limit below the lower limit, like so: 

The behavior of such a system is left as an exercise for the reader.

[1]: The systems engineer and control theorist William T. Powers explains this idea in Chapter 5 of his book Behavior: The Control of Perception:

The “reference signal” is a neural current having some magnitude. It is assumed to be generated elsewhere in the nervous system. It is a reference signal not because of anything special about it, but because it enters a “comparator” that also receives the perceptual signal. … 

The comparator is a subtractor. The perceptual signal enters in the inhibitory sense (minus sign), and the reference signal enters in the excitatory sense (positive sign). The resulting “error signal” has a magnitude proportional to the algebraic sum of these two neural currents — which means that when perceptual and reference signals are equal, the error signal will be zero. If both signs are reversed at the inputs of the comparator, the result will be the same. The reader may wish to remind himself here of how a neural-current subtractor works by designing a comparator that will generate one output signal for positive errors, and another for negative errors. (This is necessary because neural currents cannot change sign.)

Over the past few months, I’ve read several articles in my local urbanist publication Greater Greater Washington expressing various degrees of angst about the fact that the most recent YIMBYtown conference was held in Texas and prominently featured Montana Governor Greg Gianforte.

I don’t want to oversimplify the argument, but the alleged problem here is essentially that Gianforte, as a person, and the state of Texas, as a collective, stand for a lot of bad ideas on non-housing topics. I completely agree with that! If I lived in Montana, I would cast an ineffectual protest vote for Gianforte’s Democratic Party opponent. Or to put an even finer point on it, I will almost certainly support Katie Hobbs’ reelection against whatever freak show the Arizona GOP throws up against her in 2026, even though she vetoed a good housing bill. But that’s because I, Matthew Yglesias, am not a housing reform institution, even though it’s a cause that I care about a lot and am closely identified with personally. I also care about other things.

That said, as a stakeholder in the housing movement, I would be upset to see a housing organization uphold Hobbs as a figure to celebrate. There are plenty of Democrats who are good on housing — talk about Wes Moore or Jared Polis!

By the same token, the thing about Montana is that it is a very conservative state. I don’t believe that any state is unwinnable on a gubernatorial level, but the median voter there is way to the right of the national average, so any candidate would have to say some pretty conservative stuff to win. What’s notable about Gianforte isn’t that he has normal (bad) Republican Party positions on abortion and LGBT issues, it’s that he championed and signed a significant bipartisan housing reform. That’s good! And an organization focused on housing should highlight people who do things like that.

This all seems sufficiently obvious to me that it’s hardly worth saying, but I worry that it’s going to become one of these scenarios where the best lack all conviction and the worst are full of passion and fury.

Muhammed Alameldin’s lament that YIMBY politicians won’t take the pro-Palestinian stances he prefers is totally reasonable as a personal reflection from someone with family in Gaza. But as a NextCity take, I think it’s dangerous. I know YIMBYs with lots of different views on Israel/Palestine, a very divisive and emotional issue with extremely tenuous links to domestic urban policy in the United States. Pressuring everyone to come into alignment on it would blow the movement up.

Meanwhile, the fact that YIMBY reforms have non-trivial Republican support is a genuine source of strength that is worthy of celebration. I can think of many causes that would benefit from being less partisan, and few if any that are suffering for not being sucked into an omnicause.

Good for the movement vs good for the staff

I think part of the breakdown here is that most of the people doing the work of YIMBYism are under-forty college graduates who live in big cities.

Read more