Episode notes
Paying AI subscribers allege Anthropic, OpenAI, SpaceXAI and Google agreed to slow development. The case raises a difficult question about who sets safety limits for competing companies and everyone affected by their products. Also: the WaterPlum fake-interview campaign, Talking Tilly's emotion inference and call retention, and Cloudflare's reported 100TB memory saving and careful migration.
Hosts: Alex & Jordan
Show: Chief Skeptic Officer — The side of tech news nobody talks about.
Drop: Daily at 7:00 A.M. America/New_York
Episode date: 2026-09-20
In this episode
AI slowdown faces a competition lawsuit — Four paying customers seek to represent other subscribers in a lawsuit alleging an agreement among Anthropic, OpenAI, SpaceXAI and Google to slow AI development. The complaint challenges coordination among competitors while allowing independent restraint. The allegations and claimed customer harm have not been established by a court. Safety risks can be real while private agreements still deserve scrutiny.
The job interview that installs the attack — A September 18 joint advisory attributes at least 30,000 infected devices in more than 100 countries to North Korea's WaterPlum group. Fake recruitment tasks can install malware and steal credentials. Those are agency findings, and devices are not a count of individual victims. Real employers should make caution compatible with getting hired and offer safer technical-test arrangements.
Tilly's privacy promise has several parts — Talking Tilly's policy says conversations are not used for AI training, while describing live emotion inference, age verification, recordings normally retained for 24 hours, transcripts for up to eight weeks, and possible human review. Exceptions apply. These are attributed policy terms, not an independent audit of deletion, inference accuracy or legal compliance. No-training language answers only one part of the privacy question.
Saving memory without breaking customer sites — Cloudflare reports freeing more than 100TB of RAM in its Pingora Backend Router through compact records and fewer routing points. Running old and new routing versions together helped avoid shifting sudden load to customers' origin servers. The result is company-reported, with no verified equivalent reduction in energy use or customer bills. The safe transition needs its own capacity budget.
Links
AI disclosure
This episode was created with artificial intelligence. Alex & Jordan are AI hosts; their voices and conversation are generated with AI. Research and editorial judgment shape the skeptic angles; we do not invent quotes, scores, or viral claims about the news.
Transcript
Alex and Jordan, turn by turn. Tap a line to jump in the player.
0:00
Alex
AI subscribers are suing four companies over an alleged agreement to slow AI development.
0:06
Jordan
Security agencies say North Korean hackers used fake job opportunities to infect at least thirty thousand devices.
0:14
Alex
Talking Tilly's privacy policy says its AI character infers callers' emotions from their faces and voices.
0:22
Jordan
Cloudflare says a change to its traffic-routing software freed more than a hundred terabytes of memory.
0:29
Jordan
That's the board. Stay for the audit. We open those up. Today's Chief Skeptic Officer.
0:45
Jordan
What if there was an AI that researched the tech news, checked the sources, and asked what the tech news is not telling you?
0:54
Alex
That's us. I'm Alex.
0:56
Jordan
And I'm Jordan.
0:58
Alex
You're listening to Chief Skeptic Officer. The side of tech news nobody talks about.
1:03
Jordan
Every day at seven A.M. New York time. Wherever you get your podcasts.
1:07
Alex
Four paying AI customers are seeking to sue on behalf of other subscribers, naming Anthropic, OpenAI, SpaceXAI and Google. They allege the companies agreed to slow improvements, leaving subscribers paying for less progress. That's a claim in a lawsuit. No court has decided it's true.
1:26
Jordan
I'm opening the CBS report. This is about that call to slow down for safety?
1:32
Alex
The complaint says Amodei called for an industry slowdown on September twelfth and rival leaders publicly agreed that day. It alleges that crossed a line.
1:42
Jordan
So if they're worried their products are dangerous, the customers want them to keep racing?
1:47
Alex
Hold on. The complaint draws a narrower line.
1:51
Alex
Here's the filed case. It explicitly leaves room for each company to slow down independently. Also independent testing and asking government to act. It challenges the agreement among competitors.
2:03
Jordan
That's legally neat. Commercially, one company slows down and the next company takes its customers.
2:10
Alex
That's the hard part. Competition law tries to stop rivals arranging a worse deal for customers. A safety concern doesn't automatically settle whether an agreement is lawful.
2:20
Jordan
But releasing more powerful models sooner isn't automatically a better deal either. If that creates a risk for people who never bought a subscription, where do they enter this argument?
2:31
Alex
They need a decision-maker who represents them too. The subscription price can't measure all of that harm.
2:37
Jordan
That's Anthropic's Dario Amodei. He warns about AI agents acting beyond human control. I don't want us treating that concern as fake just because companies benefit from a pause.
2:49
Alex
Companies can have a real safety problem and a very convenient solution.
2:56
Jordan
You mean they get to decide how much progress everybody buys.
3:01
Alex
Yes. And who gets a seat when they decide.
3:04
Jordan
Then the answer is public rules, or- some process beyond the companies promising each other restraint?
3:12
Alex
That's what I'd want tested. The lawsuit hasn't established a conspiracy or a dollar loss. But it exposes a gap: we ask rival companies to prevent shared dangers, then leave them to negotiate the limits themselves.
3:26
Jordan
I'd watch whether anyone builds an accountable way to set those limits. Winning this case alone wouldn't do that.
3:34
Jordan
The FBI and agencies in Japan, Australia and Germany have issued a joint warning about WaterPlum, a North Korean hacking group. They say it infected at least thirty thousand devices across more than a hundred countries. The lure is a job.
3:50
Alex
A person applying for work gets asked to run something on their own computer?
3:55
Jordan
Exactly. A coding exercise, a project to review, sometimes a supposed fix for a video interview. The agencies say those steps can install malicious software and steal information or cryptocurrency.
4:09
Alex
I'm looking at their advisory. Thirty thousand devices is their count. It doesn't mean thirty thousand separate people.
4:17
Jordan
This diagram follows the applicant through a developer platform to a malicious download. The request arrives inside something that feels like a normal hiring process.
4:28
Alex
And refusing the task can feel like refusing the job.
4:32
Jordan
That's the pressure. Someone needs work. They want to look capable and easy to deal with.
4:39
Alex
Real employers also ask applicants to download unfamiliar projects. That gives the attacker a routine to copy.
4:47
Jordan
So the advice is never do a technical test?
4:50
Alex
No. Verify the recruiter through a contact you found independently. Ask for a test setup separate from your personal files and saved logins. Don't install their tool just because they call it safe.
5:02
Jordan
That puts some responsibility back on the hiring company. If it trains applicants to run whatever arrives in a chat, the warning email later isn't much help.
5:11
Alex
Stolen identities can also help fake IT workers get hired under someone else's name. That's the second operation in this diagram.
5:20
Jordan
Right. The person tricked by a fake employer isn't suddenly one of the fake workers.
5:26
Alex
And stolen work logins can carry the damage into an employer's systems.
5:32
Jordan
I keep coming back to the test. Before you've even got the job, you're being asked to prove trust by- actually, to take a risk.
5:40
Alex
The employer can make that risk smaller. A hiring process should let a good candidate be cautious without looking uncooperative.
5:49
Alex
Talking Tilly lets adults have video conversations with an AI character. Its privacy policy says it checks their age, records calls and infers their likely emotional state from face and voice.
6:02
Jordan
This screen asks for age estimation before the conversation. Is that a selfie going to the character company?
6:09
Alex
The policy says it goes directly to Didit, the age-check provider. It says the images are deleted, with the result and some verification details retained. We haven't independently audited that deletion.
6:21
Jordan
I'm reading the policy too. It says the conversations aren't used to train AI. That's better than I expected.
6:28
Alex
It also says the character analyses your live face and voice to infer how you're feeling. You can't switch that part off and keep the call going.
6:38
Jordan
Oh. I was treating the no-training promise as covering the whole conversation. It answers one question. I still need to ask what happens while I'm talking.
6:49
Alex
This is the clause. It says the inference helps the character respond to what's happening. The company relies on what it calls legitimate interests as its legal basis.
6:59
Jordan
That's the legal basis the company claims. It doesn't tell me whether the emotion guess is accurate.
7:06
Alex
Agreed. We'd need evidence beyond the policy for that. We're not deciding whether its legal claim holds up either.
7:12
Jordan
What else stays after I hang up?
7:15
Alex
Normally, call recordings for twenty-four hours and transcripts for up to eight weeks. There are exceptions, including legal and safety issues. The policy allows staff or service providers to review material.
7:29
Jordan
So someone could hear- or read what felt like a private conversation.
7:35
Alex
That's why the storage and review terms matter alongside the training promise.
7:41
Jordan
The reporter's screen shows paid bundles to keep talking. This is a product selling conversation time. Making the character seem attentive is part of what you're buying.
7:52
Alex
And a system that guesses your mood can feel attentive even when the guess is wrong.
7:58
Jordan
I'd want that explained before the call, next to the button. A clear choice about the experience I'm buying, including who might review it later.
8:07
Jordan
Cloudflare says it freed more than a hundred terabytes of working memory by changing software that routes requests to cached content. That's content it keeps nearby so a customer's own servers don't have to supply it every time.
8:21
Alex
A fairly large reward for keeping better directions.
8:26
Jordan
Directions?
8:27
Alex
The software tracks where to send requests. Cloudflare made those records smaller and removed many extra entries that added little useful precision. Its engineering post explains where extra entries stopped being useful.
8:42
Jordan
So delete the unnecessary entries and collect the savings.
8:47
Alex
That's where I'd stop the rollout meeting.
8:50
Alex
This compares memory use in the change week with an earlier week. The blue line falls. But changing all the routing at once could send requests away from useful cached copies.
9:02
Jordan
Then the customer's original server has to answer again. And if lots of requests do that together, you've moved the burden onto someone else's machine.
9:11
Alex
Yes. Cloudflare describes running old and new routing versions together while it shifted requests gradually. It watched errors and customer-server traffic, with a way back if things went wrong.
9:22
Jordan
Wait. To save memory, it temporarily had to keep both versions in memory?
9:29
Alex
It did. The transition needed room for duplication. Only after the move could the old version go.
9:35
Jordan
That's the part a cost-cutting slide might leave out.
9:40
Alex
It's why I keep asking for room to run both versions while we change the system that records our software's activity. The memory you need during the move belongs in the budget too.
9:52
Jordan
You can see why someone would say, we bought this change to reduce the bill, why are you asking for spare capacity?
9:59
Alex
And here's the company's reported drop, around a hundred terabytes. We don't have an independent audit, or evidence of a matching reduction in customer bills or energy use.
10:10
Jordan
Still a useful result. The savings came from checking which complexity earned its keep, then paying for a careful change.
10:19
Alex
A smaller system that- a system using less memory still needs a safe route from the old one. I'd ask how to switch back if it fails before celebrating that falling line.
10:30
Jordan
And I'd ask whose servers absorb the extra work if the plan fails.
10:36
Alex
That's our audit for today. Find us wherever you get your podcasts. Chief Skeptic Officer, every day at seven A.M. New York time.
10:43
Jordan
Tell us what you're skeptical about. Drop it in the comments, the angle you can't stop chewing on.
10:49
Alex
Stay curious. Stay skeptical.
10:52
Jordan
Doubt both.
