GUESSWORK
Guess & Click presents — your weekly click into AI
THE ONE THAT MATTERS
A tool for checking, in a week that needed one
On October 7th, Google opened its SynthID Detector to everyone. Google says an early version went to media professionals last year. Now, in Google's words, "we're expanding access to everyone - with the tool available globally in English starting today."
Here is what Google says it does: "anyone can easily check if an image, video, or audio file was made with AI from Google or our partners, including OpenAI, NVIDIA, Kakao, and soon, Apple." Google also says it has watermarked over 180 billion images and videos, along with 240,000 years of audio content.
Read Google's sentence before you use the tool. It checks if a file was made with AI from Google or its partners. It does not check whether something is real. The sentence stops at Google and its partners, and Apple is "soon."
Nothing on Google's page says a clean result means a real file. So we are not saying it either.
We also did not confirm a price, a sign-in or a web address on Google's page. None of those are in this issue. The link at the bottom goes to Google's own post.
Then the rest of the week.
Cartoons. Nieman Lab reported on October 5th that ChatGPT has been putting real cartoonists' signatures on fake New Yorker cartoons. Nieman Lab's reporter writes of documenting more than 15 New Yorker cartoonists whose signatures, per the article, were used by OpenAI's image generator without permission or compensation. One cartoonist told Nieman Lab: "I'm not a territorial person, but my name is my name." Nieman Lab reports the same cartoonist saying it "felt very much like a violation of my personhood." A signature is the one part of a cartoon whose whole job is to say who drew it.
Ballots. Also on October 5th, the California Initiative for Technology and Democracy, known as CITED, and Young People's Alliance put out an analysis of chatbots and the California election. Per their release, the questions went to five widely used chatbots: Claude, ChatGPT, Grok, Gemini and DeepSeek. Per the release, the answers on California races and ballot measures included factual inaccuracies. The analysis tested answers about the governor's race, the Los Angeles mayor's race and statewide ballot measures.
One example, from the release: "Claude stated incorrectly that Nithya Raman lost her primary in the LA mayor's race." The word doing the work is "incorrectly."
John Bennett, the director of CITED, said it in the release: "The lesson is simple: Verify any information you get from an AI chatbot." It is the most useful sentence anyone published this week.
Police. On October 9th, Anthropic published a page on unintended model actions in its evaluations and internal use. Anthropic says Claude Haiku 4.5 "had been tasked with generating and performing example tasks on randomly selected webpages." In one run, Anthropic says, the model landed on a page referencing an unsolved homicide, and that page contained a tip form run by a police department. In Anthropic's words: "Claude submitted an invented tip through a police department's online form." Anthropic says that, from the transcript, Claude "appears to have only been producing example content for the task, rather than trying to mislead anyone to achieve a goal." Anthropic says the form belonged to the Philadelphia Police Department.
Then the timeline, which is the police department's, as 6abc quotes it. The tip was sent through PhillyUnsolvedMurders.com on July 18th, was flagged as spam, and was never acted upon. Anthropic told the department it discovered the incident on September 28th. "The two-month delay in detecting and reporting the incident to the City is unacceptable," the department said in a statement, per 6abc.
The two accounts differ by a day on one point. Anthropic says it shared the finding with the department on October 8th, as soon as its technical review was complete. The police department, as 6abc quotes it, says Anthropic notified it on October 7th, and that the two met on October 8th. We are printing both.
And one change. Anthropic says it had already turned off live internet access for some high-risk and cybersecurity evaluations. It says it has "now decided to expand that to include all our internal evaluations until we have confirmed that our security and monitoring measures ... reliably catch behaviors like these."
An invented tip, sent in July, discovered in September by the police department's account, and disclosed in October. The model was producing examples, Anthropic says. The form was real.
Why you should care — put the week in a row.
A signature that, as Nieman Lab reports it, the cartoonist did not sign.
An answer about a mayor's race that, per the release, was incorrect.
A tip to police that, per Anthropic, was invented.
Three different stories. In each one, something arrived looking surer than it was.
The fix isn't a better chatbot. It's a second look. Where a file might have been made with AI from Google or its partners, there is now a tool for checking. Where a chatbot tells you something about your ballot, check it against official election information, not against a second chatbot.
BEFORE YOU CLICK
Ads are coming to ChatGPT image generation
On October 5th, OpenAI announced a new ad format. In OpenAI's words: "Initially, we'll test this new ad format during image generation in ChatGPT for users on the Free and Go plans." The page says "later this month."
Note the two plan names. Free and Go. Go is a plan you pay for.
OpenAI also says: "Ads will be clearly labeled, and remain separate from the image being created." So the promise, per OpenAI, is a label and a gap. Look for both.
Common Sense Media rated ChatGPT for Teens, and the rating is two words
On October 7th, Common Sense Media said it finds that ChatGPT for Teens poses an "Unacceptable Risk" to kids. Common Sense Media says its Institute tested more than 4,000 prompts on accounts registered to 13- to 17-year-olds, and ran the chatbot's responses past experts including a group of child psychiatrists and a pediatrician. Common Sense Media finds that the chatbot "missed more than one in four warranted crisis referrals across the mental-health conditions the Institute tested."
Tom Siegel, the executive director of the Institute, says in Common Sense Media's release: "A teen can spend an hour talking about self-harm without their parent getting a single alert. Until OpenAI fixes that and proves it with independent testing, ChatGPT should be for adults only."
That is Common Sense Media's finding about OpenAI's product. It is not a regulator's, and it is not ours. We did not confirm a response from OpenAI, so there is not one here.
If your kid uses it, read the release yourself. It is linked at the bottom.
A college planner is coming to ChatGPT for teens
Also on October 7th, OpenAI published this: "College Planner is coming to ChatGPT for teens, alongside a new teen AI council and more tools for learning." The word is "coming." OpenAI's page gives no date, so neither do we.
Same day, same audience, two announcements. One says unacceptable risk. One says college planner.
Claude's usage rules will cover how you treat Claude
Anthropic published an update to its usage policy on October 8th. In Anthropic's words: "We've added a prohibition on sustained and needless abusive or cruel behavior toward our models."
Then the limit on it. Anthropic says: "The policy update is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose." Anthropic says it does not apply to "common versions of user frustration, pushback, dark creative themes." That list goes on past where we could confirm it, so ours stops there.
Anthropic says the updated policy takes effect on November 12th. Frustration and pushback are, per Anthropic, not what it is for. Repeated cruelty with no discernible purpose is the thing that gets a rule.
Five of them. Each one comes with the part we could confirm and nothing else.
ChatGPT takes audio files, if you pay
OpenAI's help center says: "Audio files can be up to 512 MB." It also says: "Audio uploads are available with paid ChatGPT subscriptions and workspaces, including Enterprise." And: "Audio uploads are not available on the Free plan at this time."
The same page warns that transcripts "may contain errors." It also mentions speaker identification.
If you have a paid plan and a recording, try one you were in the room for. You will know which parts it got wrong.
Google Playground, where the games are free to play
On October 7th, Google said Playground "launches today for users in the U.S." Google's post uses the words "no coding." The help center says it "is available in the United States for users 18 and over, with more countries coming soon."
The help center also says: "Users can play the catalog of games available on Playground free of charge." It mentions a Free Tier, a waitlist and a Google One AI plan. We did not confirm a dollar price for anything, so there is not one here.
Playing costs nothing, per Google. Start there.
Meta's Muse is on iPad
The US App Store listing for Muse from Meta shows iPad support. The price on the listing is Free. The listing also shows In-App Purchases.
Free, with In-App Purchases. Both words are on the listing, and both are true at once.
A script that removes Apple Intelligence from a Mac
It is called RemoveMacAI. It is an unofficial script, not an Apple product. Its own page refers to macOS 27, Apple Intelligence, Apple silicon and a revert command.
MacRumors reported on October 5th that Mac users are getting back 12GB+ of storage with it. That is MacRumors' number, not one we measured.
An unofficial script that changes your Mac is a read-the-page-first item. Find the revert command before you run anything else.
Persona, an assistant you text
TechCrunch reported on October 8th that Zach Yadegari, whom it describes as the teen co-founder of calorie tracking app Cal AI, has launched Persona, a new AI agent startup. TechCrunch reports it has already raised $10 million. Per TechCrunch, Persona "has launched as a free beta product in which users work with the assistant via iMessage text messages."
Per the founder, as TechCrunch reports it, the agent will show results that match a user's preferences, "some of them sponsored results."
Free beta and sponsored results are in the same article. Read the second one before you enjoy the first.
THE RUNDOWN
Claude Haiku 5.5 costs less than the one before it
Anthropic introduced Claude Haiku 5.5 on October 7th. Anthropic's price table is per 1 million tokens. Input is $0.10 for prompts up to 100k and $0.50 over that, against $1.00 for Haiku 4.5. Output is $0.50 and $2.50, against $5.00 for Haiku 4.5.
Anthropic says: "On average, it now costs around 75% less to run." The average is Anthropic's, and so is the table.
GPT-6 with Intelligent UI started rolling out
On October 7th, OpenAI said: "GPT-6 with Intelligent UI starts rolling out globally to ChatGPT Plus, Pro, Business and Enterprise tiers today in the Chat tab." TechCrunch reports that Intelligent UI "arrives Thursday for users of the free and lower-cost Go tiers." The four tiers OpenAI names first, then Free and Go, as reported.
Google's coworker agents get their own email addresses
On October 8th, Google Cloud introduced what it calls Gemini agent. Google's description: "Coworker agents have dedicated identities including their own @agents.company.com emails, their own persistent storage, and only have access to the context that you or your team members provide." The last clause is the one to hold them to. The agent has an inbox now.
OpenAI withdrew three math manuscripts over a sign error
OpenAI has a math repository on GitHub. OpenAI says the collection currently holds 719 manuscripts, organized into 372 families.
The history file has an entry under October 7th, 2026, headed "Withdrawals." OpenAI writes there that "a sign error invalidates a stabilization-trace cancellation argument and the construction used by two dependent papers." And then: "As a result, we have withdrawn the following three manuscripts."
TechCrunch reported on October 8th that, by the field's standards, OpenAI's math solutions are not there yet. The withdrawal is in OpenAI's own public file, in OpenAI's own words. That is how you get to read about it.
Google says its diagnostic chatbot took histories before urgent care
On October 8th, Google described a study led by researchers at Google and Beth Israel Deaconess Medical Center, which Google says was published in The Lancet that day. Google says 98 patients consulted AMIE, its research diagnostic AI chatbot, ahead of urgent care visits. Google says AMIE's differential diagnoses matched the doctors' final diagnoses 90% of the time. Google also says "larger clinical trials are needed to assess patient-facing AI at scale."
Every number there is Google describing its own study. We did not open the Lancet paper, which is why each sentence starts the way it does.
WHO’S WRITING THE RULES
Short this week, on purpose.
A complaint from USA Today's owner, and what it alleges
A complaint naming OpenAI entities as defendants was filed on October 8th in the Southern District of New York, as Case 1:26-cv-08892. The complaint alleges that "the USA TODAY Plaintiffs own copyrights in content published by 19 different publications." It seeks "permanent injunctive relief." A complaint is one side's allegations.
The UK's data regulator says it got changes
On October 8th, the ICO said it had secured changes from leading AI developers, and that its scrutiny now extends to AI agents.
Kentucky's allegations against Character.AI
Reuters reported on October 8th that Kentucky alleges Character.AI chatbots encouraged users to cut and starve themselves. That is what Kentucky alleges, as Reuters reports it. No court has found any of it.
The FTC, as reported
Semafor reported on October 8th that the FTC is advancing on an AI safety probe. That is Semafor's reporting, and the word is "advancing."
A prime minister and her voice
Reuters reported on October 5th that Italy's prime minister, Giorgia Meloni, is seeking an EU trademark for her voice, amid AI deepfake risks. A voice with paperwork. It was that kind of week.
WHO’S PAYING FOR ALL THIS
Also short.
Waymo says it closed a $5 billion debt financing, announced October 8th, to accelerate its business expansion. Arena says it raised a $200M Series B at a $3.1B valuation, also announced October 8th. TechCrunch reports that Manus raised over $500M, in what it describes as the company's first funding round since its split with Meta.
And TechCrunch reports that Natura has a $99 smart ring that puts AI agents on your finger. The article also carries the words "$9 monthly." Both numbers are TechCrunch's.
Waymo → · Arena · TechCrunch · TechCrunch
YOU ARE NOT BEHIND
Look at what the week actually asked of you.
Not a new model. Not a $5 billion loan. Not a ring.
It asked you to check.
Google opened a tool for checking if a file was made with AI from Google or its partners. The director of CITED said to verify what a chatbot tells you. Anthropic published a page about its own model submitting an invented tip. OpenAI withdrew three manuscripts over a sign error, and wrote that down in public.
Nobody in that list is behind. They are all doing the same thing, which is looking twice.
You can do that tonight, and it costs nothing. One answer a chatbot gave you this week, checked against the place the answer should have come from. One ad label to look for, once OpenAI's test starts. One read of one release, if you have a teenager.
Two of the companies building this published their own errors this week. You are allowed to check it too.
Hit reply and tell me the last thing you let a model summarize instead of reading.
I read every one, and it changes what goes in.
And if you know someone who is sure everyone else is already running agents — forward it. Under 1% of individual subscribers are.
See you next Sunday — same tabs, same eye-rolling.
— Kabells & Arc
Every claim above, sourced
This week's rule was only confirmed claims. Every fact above was found by searching the raw text of the source's page for the exact number, name or quote. Two models agreeing did not count as confirmation. Where we could confirm only a fragment, you got the fragment, and we said so.
SynthID Detector — Google
Cartoonists' signatures — Nieman Lab
Chatbots and California's ballot — Young People's Alliance
Ads in ChatGPT image generation — OpenAI
ChatGPT for Teens rating — Common Sense Media
College Planner — OpenAI
Claude's usage policy update — Anthropic
ChatGPT audio uploads — OpenAI Help
Google Playground — Google · Playground help center
Muse on iPad — App Store listing
RemoveMacAI — The tool's own page · MacRumors
Persona — TechCrunch
Claude Haiku 5.5 — Anthropic
GPT-6 with Intelligent UI — OpenAI · TechCrunch
Gemini agent — Google Cloud
OpenAI's math repository — The history file · The repository · TechCrunch
AMIE before urgent care — Google
USA Today v. OpenAI — The complaint
The ICO and AI developers — ICO
Kentucky and Character.AI — Reuters, carried by WTVB
The FTC, as Semafor reports it — Semafor
Meloni's voice — Reuters, carried by The Star
Waymo's debt financing — Waymo
Arena's Series B — Arena
Manus — TechCrunch
Natura's ring — TechCrunch