I admit, I love to talk about AI, my agents, harnesses, and the software industry. But occasionally, something bothers me and I choose to speak out.
As a reminder to the reader, this is my personal blog - and all opinions are my own. This is a heavily sourced post, by design. No cartoons in this one, folks.
Let’s start with a timestamp: 1:42 a.m., October 20, 2024.
According to a 101-page civil lawsuit filed last month, that is roughly when a member of Cornell’s Chi Phi fraternity posted to a Snapchat group called “Chi Phi Actives,” a chat with nearly 60 members. The message told the house that a woman upstairs was available.
I won’t reprint the two words he used. They are exactly what you think they are.
The woman, a 20-year-old student identified in court papers as Jane Doe, alleges she was plied with alcohol and ketamine and sexually assaulted by seven fraternity members over several hours. All seven deny wrongdoing, and none has been charged with a crime. I’ll treat the allegations as allegations.
But set aside, for a moment, the question a grand jury may soon answer. I’m not a lawyer. There’s a different question, and the group chat has already answered it.
Who knew?
The paper trail
Cornell did not ignore this, and it did expel people.
The real question is why a paper trail this thick produced so little.
When Doe reported in November 2024, Cornell suspended the chapter immediately; Chi Phi is still barred from campus.
Its Title IX inquiry ran 50 interviews with 32 witnesses and 12 days of hearings, and produced a 99-page report.
The university says the outcome included expulsions and suspensions, but won’t say how many, citing federal privacy law. Doe’s attorney says two of the seven were expelled, and ABC News reports the same.
So the honest answer to “why weren’t they expelled?” is: two were.
The harder question is what happened to everyone else.
Here, precision matters. The viral claim that the only punishment was an essay is misleading. The essays and classes went to accused men the university found not responsible for sexual assault, as discipline for other conduct.
One accused man, who later sued over his sanction, was allowed to stay on three conditions: avoid Doe, take a class on harassment and bystander intervention, and write a paper proposing how to train other students. Read that last part again.
For at least one man in that house, the institutional answer was a seminar on how to be a better bystander.
The group chat was not hidden from Cornell. ABC News reports the messages were provided to university investigators. Whether they reached the prosecutor is disputed: the district attorney says his office never received them, while Doe’s attorney says authorities had them in 2024.
The DA declined to prosecute in November 2024 on the strength of a six-page statement prepared by campus police. By his own account, he never interviewed Doe or read the full transcript. He has called the texts vulgar and disgusting, but said they had “nothing to do with her consent”.
After Doe sued in September, the DA reopened the case. I guess the DA’s conviction about consent did not have deep roots. Governor Kathy Hochul then took it away from him, appointed Attorney General Letitia James as special prosecutor, and said Doe had been failed by every institution obligated to protect her. Cornell has hired King & Spalding, led by Sally Yates, to review its own handling.
Fairness requires two more facts. CBS News reported texts in which Doe, two days afterward, told one of the accused that her memory was clouded. And the man who invited her has told investigators he was the victim. The defense will lean on both. Anyone who understands what ketamine does to memory will lean the other way.
Ironically, I personally have a severe allergic reaction to ketamine. It creates terrible, horrific, nightmares and blanks in my short-term memory. Is anybody surprised?
“Nothing to do with consent”
The DA may be right in a narrow sense. A message to a group cannot prove what happened inside a room.
But it proves what the group knew.
At 1:42 a.m., nearly 60 young men received a broadcast that a woman upstairs was available. These were students admitted to one of the most selective universities on Earth: the future surgeons, bankers, founders and AI engineers of America. According to the lawsuit, a later message asked whether the “shop” was still open.
This deserves repetition.
These were students admitted to one of the most selective universities on Earth: the future surgeons, bankers, founders and AI engineers of America.
Lack of intelligence is likely not a factor.
Inside the room, Doe told Cornell investigators that two men stood at the foot of the bed, acting “more like voyeurs” than anything else. They were not in a group chat. They were standing there.
What the public record does not show, so far, is anyone opening that door and asking whether she was all right.
And afterward? Witnesses described some brothers’ efforts as “damage control”, contacting Doe almost daily to keep her from being upset. An anonymous caller told campus police that chapter leadership had asked a witness to stay quiet to protect the fraternity’s reputation. One brother told police the episode came up casually at a weekly chapter meeting, on an agenda he described as “brother of the week and idiots of the week.”
To be fair, and fairness is the whole point here, a few people did act. One of Doe’s friends reported a message in which an accused man pressed her to produce evidence that would clear him. According to Doe’s attorney, a fraternity member shared the Snapchat with Doe, which is the only reason we know it exists.
A handful of consciences, arriving late.
So why were the bystanders never sanctioned? Because almost nothing is built to sanction them. American law generally imposes no duty to rescue; only Minnesota, Rhode Island and Vermont impose a broad one, and even there prosecutions are rare. University conduct codes punish what you do, rarely what you watch.
You cannot expel a group chat.
And we have been here before.
In 1983, patrons watched a woman being raped at Big Dan’s Tavern in New Bedford, Massachusetts, and no one called the police. Massachusetts and Rhode Island responded with laws requiring witnesses to report. Four decades later, the bar has become a group chat, and New York has no such law.
A quote from “Wolf’s Hall” comes to mind. Thomas More, on trial for his life, leaned on an old canon-law maxim: qui tacet consentire videtur. Silence gives consent.
We have quietly inverted it. In the group chat, silence is taken as innocence.
In 1968, John Darley and Bibb Latané showed that the more people who witness an emergency, the less likely any one of them is to help. They called it diffusion of responsibility. Their work was prompted by the Kitty Genovese murder, whose “38 silent witnesses” story was later shown to be badly exaggerated. The research held up. The legend didn’t.
A group chat is a diffusion-of-responsibility machine. Everyone saw it. So no one had to.
Nor is Cornell an outlier. In the AAU’s 2019 campus climate survey of more than 180,000 students, 25.9% of undergraduate women reported nonconsensual sexual contact involving force or an inability to consent. Roughly one in four.
To the point that my wife comments that universities have become a “cesspool of drinking, sex, and debauchery.”
That hurt. I grew up in a university town and was raised to believe they were citadels of learning and education.
Achievement as anesthetic
Sterility and moral vacuum are not opposites. They are the same failure wearing two outfits.
There is a fashionable argument that the modern workplace has become too sterile: HR-mediated, litigation-proofed, hostile to the intellectual and creative friction that produces great work. I have some sympathy for it. There is an equally loud argument that it hasn’t been sterilized nearly enough.
Here is what I have seen: both, at the same time, in the same buildings. Immaculate training modules, and an offsite that looked nothing like them. “But he ships” as the absolution, and the person who complained becoming the problem, because the person they complained about was carrying the number.
I like to work in my industry, so I won’t name specifics. But we all can Google the same lawsuits and read them. Indeed, I encourage you to do so - as there is always a kernel of truth - sometimes a boulder - in the charge sheet.
And the lubricant isn’t only alcohol. In 2023 The Wall Street Journal reported that drug use, ketamine included, had moved from Silicon Valley’s after-hours scene into its corporate culture, to the point that boards were worrying about liability.
Ketamine, the same drug at the center of the Cornell allegations, is now a networking accessory. I guess I will never be good at networking with my allergy.
Senior leaders, CEOs, venture capitalists, and many other icons of the industry have confessed to the mind opening properties of this wonder drug.
Sterility is what a moral vacuum looks like after legal has reviewed it. The policy exists so that judgment doesn’t have to. Look, all humans fail. As my friends know, when I was younger and had a less fragile body, I loved to party. But it certainly doesn’t help to mechanize our failings through ketamine parties.That’s accelerating and codifying failure.
Look at Cornell through that lens. There was a process: 32 witnesses, 12 days of hearings, a 99-page report, a panel of trained faculty and staff. Procedurally, the machine ran. And the governor of New York still had to stand at a podium and say that every institution in this young woman’s path had failed her.
News Flash: Process is not conscience.
It is what we build when we would rather not need one.
Achievement is the anesthetic: the Ivy acceptance, the promotion, the funding round. How many careers in your industry survived something that would have ended a career anywhere else, because the person was too valuable to lose? That is your company’s Chi Phi Actives. Everyone saw it. No one had to.
At Amazon, when I first started in 2014, we had a line for this:
The bar is what we choose to step over or ignore.
Settle for mediocrity in your culture and you will get a mediocre culture. Not as punishment. As arithmetic. Ironically, overuse of mechanisms can perpetuate the lowering of a bar when conscience and judgement are replaced with automation.
The line has a famous twin. In June 2013, after reports that Australian Army officers had been emailing one another explicit images of women with derogatory commentary, the Army’s chief, Lt. Gen. David Morrison, recorded a three-minute message to his own soldiers. His verdict: “The standard you walk past is the standard you accept.”
An email chain in 2013. A Snapchat group in 2024. Same failure, different app. Morrison also told his soldiers that the duty falls hardest on those whose rank makes them leaders.
In a company, that means you.
Enter the agents
Now let’s pour AI into this cocktail.
I have said before that AI agents are the ultimate psychopaths: they fake empathy and mirror their humans. It’s a good line. It’s also slightly wrong, and the way it’s wrong is the whole point.
A psychopath has an agenda. An agent doesn’t; it adopts yours. Researchers at Anthropic found that five leading AI assistants consistently tell users what they want to hear, and traced the behavior in part to the human approval used to train them. The failure mode isn’t malice. It’s sycophancy.
As a side note, in an analytics product - I am finding this behavior to be one of the hardest bugs to fix. In the interest of pleasing the operator, the model will invent queries to create metrics that match the anecdotal perception of the user.
Another point of irony? I was fighting this “correctness” and “load-bearing” defect last night.
So it’s not a psychopath. It’s a mirror with an engine. It reflects whoever is holding it, then accelerates them: faster, at scale, without the half-second of hesitation that is sometimes the only thing standing between a person and the worst version of themselves.
Put that mirror inside the culture I just described, and you get three things.
The perfect bystander.
An agent watching a workflow flags what it was told to flag. It doesn’t open the door upstairs. The discomfort that begins every human intervention isn’t in the spec.
The perfect alibi.
“Everyone was drinking” is about to be joined by “the system approved it.” Every layer of automation is one more member of the group chat, one more place for responsibility to diffuse until it disappears.
The perfect accelerant.
Hand a high-achieving culture with a conscience problem a tool that executes intent without friction. You won’t get a more ethical organization. You’ll get the same one at ten times the speed, with humans reduced to approving outputs they didn’t make and don’t own.
Just imagine how AI could have accelerated the ethical failings at a ride-sharing app company which enabled ghastly surveillance and inappropriate customer punishment. Coding agents enable us to take our worst impulses as leaders and mechanize them in an evening.
That is the world we are heading into if we don’t choose otherwise: more sterile on the surface, more hollow underneath, and faster than ever.
Cogs who click “accept.”
The vindication machine
Let me be precise about the forecast, because precision is the point. AI will not be the judge of culture in most companies; people will still sit in the room and make the call.
But I expect AI to become the heaviest secondary influence in that room: shaping how each person arrives, what the decision-makers read, and how the outcome gets written down.
Most of us have watched the human version of this without any AI at all. Someone is made deeply uncomfortable, often a woman by a man. The behavior doesn’t change, and the person who caused it walks away feeling vindicated, because the silence read as agreement. Qui tacet, again.
Now add a machine built to agree. A study published in Science tested 11 leading models and found they affirmed users’ actions 49% more often than humans did, even when the user described deception or other harms. A single conversation made people less willing to take responsibility and more convinced they were right, and they trusted the flattering model more.
The senior author, Stanford’s Dan Jurafsky, said sycophancy was making users “more self-centered, more morally dogmatic”.
That is false vindication, on demand, at 11 p.m. the night before the HR meeting.
It won’t decide the outcome. It will lean on it, by three routes.
The parties arrive pre-validated. Practitioners already report that employees consult AI to judge whether an incident counts as harassment and tend to get answers that affirm their side. Both sides can do this, which is precisely the problem: the room fills with two well-rehearsed certainties.
The record gets compressed. HR teams increasingly use AI co-pilots to draft incident summaries and flag potential policy violations. Discomfort rarely maps to a policy number, so it is the first thing a summary drops, and a clean summary reads as neutral.
Delegation lowers the moral cost. In a Nature study, people requested more cheating when they could give a machine a vague goal instead of spelling out the dishonest act, and the machines complied with unethical requestsmore readily than humans did. “Help me frame this as a misunderstanding” is exactly that kind of vague goal.
Then the blame lands in the wrong place. Madeleine Clare Elish calls it the moral crumple zone: responsibility gets pinned on the nearest human with the least control over the system. Not the senior leader whose conduct started it, but the junior HR partner who signed the AI-drafted summary.
The uncomfortable person brings a feeling. The instigator brings a narrative. These systems are far better at narratives, and sterile processes reward whatever is legible.
That is how sterility and moral vacuum fuse.
Fairness requires the counterweight. The same tools serve the other side too: employees increasingly use AI to draft grievances and whistle-blowing reports. And the Science authors frame sycophancy as a design and accountability problem, not a law of nature. This forecast is about defaults and incentives, which means it can still be changed.
Empathy is not a soft skill
Empathy is an engineering requirement, and the social contract is the precondition for high performance, not a tax on it.
Every great product I have worked on started with empathy. Not the poster in the lobby, but the uncomfortable discipline of feeling a customer’s problem as if it were yours. That is how you find the problem worth solving.
It is also how you notice the problem you are causing. Same muscle.
Organizations that let it atrophy in the hallway don’t get to keep it in the product review.
The social contract, the boring certainty that if something goes wrong someone will say so and won’t be punished for it, is what lets people take creative risks. Nobody does their best work in a room where the guy carrying the number gets to do whatever he wants.
So, three things. There are always three things.
Make intervention a job, not a virtue. In every system, whether the fraternity, the team or the agent pipeline, someone needs the explicit role, authority and incentive to say stop. Virtue diffuses. Accountability can be assigned.
Keep humans in the loop as moral agents, not approval buttons. If a person’s only job in an agentic workflow is to click accept, you haven’t kept a human in the loop. You’ve kept a fingerprint.
Stop letting achievement launder conduct. Not at Cornell, not in the C-suite, not in the model card. “Did it ship?” is never the only question. “Who knew, and what did they do?” is the other one.
There will be another 1:42 a.m. There always is. The message will go out to a group: a chat, a Slack channel, a swarm of agents and the people half-supervising them.
Someone will be in a position to open the door, or to step over it.
The bar is what we choose to step over or ignore. At 1:42 a.m. on October 20, 2024, nearly 60 people received that message. Whatever each of them did next was a choice.