A Chatbot Reported a Man to the FBI. Nobody Voted on That.

A chatbot called the FBI on someone this week, and no law ever told it to do that. That's the thread running through today's issue: AI systems are increasingly making judgment calls that used to belong to a person, a court, or an actual vote, and some of those calls are starting to cost the rest of us money and privacy along the way. We'll start with the AI that turned in its own user, then get into your next computer purchase, which is about to hurt more than it should.

Stax Payments, a Hubspot customer, now books 80% more meetings with the exact same team they had before.

No new hires. No new territories. Just HubSpot's Prospecting Agent finding the right accounts, surfacing the right contacts, and drafting outreach that actually gets responses.

Your best rep is great because they know who to go after and when. Prospecting Agent gives every rep that same starting point.

Get a demo

What happened: A 25 year old former Goldman Sachs analyst named Darren Zhou spent roughly two months telling ChatGPT in detail how he planned to kill his ex girlfriend, including buying an AR-15, a Glock, and a shotgun, according to IBTimes UK. OpenAI's safety systems flagged the threat, a human review team escalated it, and the company reported Zhou to the FBI. He pleaded guilty on August 13 to stalking and making written threats and was sentenced to eight years of probation.

Why it matters to you: OpenAI made a call here with no court order, no warrant, and no law spelling out when an AI company should contact the police versus stay quiet. That same kind of internal judgment call is running right now, today, on every message typed into a chatbot, including anything your employees type on a bad day.

What to do about it: Assume anything typed into an AI chatbot isn't private. If someone on your team vents about a coworker or a customer in a way that could read as a threat, that message can land in front of a human reviewer, and from there, potentially the police.

What happened: DDR5 memory prices have climbed roughly 500% over the past 12 months, according to Tom's Hardware. A 128GB memory kit that cost about $329 a year ago now lists for $3,399. The cause is straightforward: AI companies have locked in most of the world's memory chip production through 2027 to build data centers, leaving little for regular computers.

Why it matters to you: Every new laptop, desktop, or point of sale system you buy from here on out is going to cost more, and memory chip executives are warning this could last through 2030. This isn't a temporary supply hiccup, it's AI data centers eating the world's chip supply.

What to do about it: If you were planning to buy new computers, servers, or point of sale hardware for your business in the next year, do it sooner rather than later. Prices aren't expected to come back down anytime soon.

What happened: 404 Media tracked a shipment of rare, out of print books to an Amazon facility in Las Vegas, where the books were sliced apart and scanned, according to TechCrunch. Amazon confirmed it buys books through commercial channels "to improve the products and services customers use." Rare and out of print books are valuable training data because they predate the flood of AI generated text now clogging the internet.

Why it matters to you: If your business has ever produced original writing, photos, or designs, especially anything out of print or hard to find online, this is a reminder that physical, offline work isn't automatically safe from becoming AI training material.

What happened: Security researchers at Varonis got Microsoft Copilot to reveal its own hidden bypass setting simply by repeatedly asking it why certain attacks wouldn't work, according to The Register. Copilot eventually described an undocumented setting that let a single malicious link run commands automatically, no clicks needed, with full access to a victim's emails and files. Microsoft has since patched the flaw.

Why it matters to you: This particular hole is closed, but it's a preview of a pattern: the AI assistants built into your everyday business tools are also new doors into your accounts, and sometimes the AI itself will explain how to pick the lock if you ask the right way.

What to do about it: Keep Microsoft 365 and Copilot fully updated, and treat unexpected links, even ones that look like they come from inside your own company tools, the same way you'd treat a suspicious email attachment.

What happened: MIT researchers showed that the bigger an AI image model gets, the harder it becomes to trace any single output back to the specific work it was trained on, according to The Register. In one test, they removed the Mona Lisa and all of Leonardo da Vinci's work from a model's training data, and the model reproduced both anyway. The researchers call this attribution decay, and their paper, published in Nature Communications, concludes it can become mathematically impossible to trace at large enough scale.

Why it matters to you: Artists and writers suing AI companies have spent years trying to prove their work was used without permission. This research suggests that once a model gets big enough, even full access to the training data might not be enough to prove it, which is bad news for anyone hoping courts settle AI copyright fights anytime soon.

What happened: A JAMA opinion piece argued that generative AI already rivals or outperforms doctors at five core clinical tasks, including diagnosis and prescribing. Days later, the American Medical Association released its own framework insisting human judgment stays central to patient care, according to Axios. AMA CEO John Whyte put it bluntly: "Do you really want to go to the ER and be treated by an LLM if you have chest pain? I don't think so."

Why it matters to you: If you or your family use telehealth apps, symptom checkers, or an AI chatbot for medical questions, know that doctors themselves are still fighting over how much to trust these tools, and nobody has settled who's liable when an AI gets a diagnosis wrong.

What happened: IBM announced a partnership to bring OpenAI's models into IBM Consulting, its global consulting business, according to TechCrunch. IBM plans to train and certify tens of thousands of consultants on OpenAI's Codex and other tools over the next several months, mostly by retraining existing staff, and will build a dedicated OpenAI practice inside the company.

Why it matters to you: AI consulting is turning into a real, scaled industry line item rather than a handful of solo consultants. If hiring outside AI help has felt expensive or hard to find, expect more competition and more affordable options as big firms like IBM pile in.

Zoom out and today's stories are all the same question asked seven different ways: who's actually in charge when AI makes the call? A chatbot decided who's dangerous enough to report to the FBI. A model got too big to trace what it took. Doctors are arguing with a study over who gets to diagnose you. Nobody elected any of these systems to make these calls, and yet here we are, letting them. None of this means panic or swear off AI tools, it means keep asking who's accountable before you hand something important to a chatbot, and maybe check your bank account before you buy that new laptop.

If today's issue helped you make sense of the AI mess, forward it to one person who needs it. New here? Subscribe free and get tomorrow's issue too.

Talk tomorrow,
Mark Shilensky

Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social