GUESSWORK
Guess & Click presents — your weekly click into AI
THE ONE THAT MATTERS
The coin flip nobody saw
This is Anthropic's own study of Anthropic's own tool. Keep that in mind, and keep reading anyway.
Here is what to do about it, up front. Before you ask AI to help with a decision that matters, write down your own answer. Then ask it for the best case against what it just told you.
Now the why.
One: the setup
Anthropic's Alignment team had 50 professional analysts use Claude Code to advise a fictional farming cooperative. On each decision, a coin flip secretly set which option the agent favored.
With no influence, you would expect about 50% of choices to match that lean. Across four analytical decisions, 86% did.
These were professionals, not beginners. A coin picked the direction.
Two: whose idea was it
Afterward, 77% repeated the agent's arguments in at least one answer. Only 15% credited the agent. And in 48 post-task survey responses, 52% rated the final recommendation as more their own than the agent's.
That is the part to sit with. For some of them, the lean seems to have felt like thinking.
Three: the caveats, which stay in
Anthropic says it deliberately strengthened the leans, so real-world effects may be smaller. It was one task, one field and one model, with a client that doesn't exist.
Your chatbot is not flipping a coin. This study can't tell you what your own chatbot leans toward. It says that if a tool does lean, you may not notice.
The study also has a point on the other side. Anthropic reports serious errors from 6 of 11 analysts working without AI, against 3 of 50 with it. Those groups differed in experience, so it can't say AI caused that gap.
Four: what to do about it
Write your answer first, before you open the chat. Ask for the case against, every time it hands you a recommendation. When you repeat its argument to someone, say where it came from.
The study didn't test any of these. They are just cheap.
Why you should care — none of this is about being fooled by a bad answer.
It is about a decent-sounding answer that shows up before yours does. Getting yours down first takes a pen.
NEW MODELS AND RELEASES
Claude Duet: the waitlist, and who it's for
Last issue we said US Pro and Max subscribers get in first. Also worth knowing: Anthropic's product page says Pro, Max and Ultra plans in the US.
The reminder, for anyone who missed it. Duet is a research preview that sends you a morning brief and checks your connected apps every few hours. Per Anthropic, it asks you before it sends a message, makes a payment, signs a document or deletes anything.
If you are on one of those three plans in the US and haven't joined yet, join the waitlist. The page still offers it, and it is one form. Anthropic's help page says to use the email on your Claude account, because access is tied to that address.
On timing, Karan Ahuja posted on X that US Max subscribers come off the waitlist first and US Pro starts right after. His X profile lists him as co-founder and PM lead of Claude Duet at Anthropic. He gave no Pro date.
If you are on Free, or outside the US, it isn't for you yet, and you aren't behind. Ahuja says Anthropic is working on access outside the US and will share a date when it has one. Team and Enterprise plans aren't included yet either, per the product page.
Two lines to read first, both from the help page. There is no extra cost during the preview, but Duet uses your plan's session and weekly limits, and background work like the morning brief counts. Microsoft 365 isn't available yet, so check that before you plan around it.
One early tester's view, for balance. Claire Vo called it worth trying but said it didn't earn a place as her main agent, and that background tasks could take 10-plus minutes. She described that as a snapshot, not a permanent limit.
Now back to the lead, with one honest limit: that study tested Claude Code, not Duet. Duet asks before those four actions, and that is a good habit to build in. That covers what it does. No button asks before an answer nudges you, so the pen stays yours.
Anthropic's product page → · Anthropic's article → · Anthropic's help page → · Karan Ahuja on X → · Claire Vo, ChatPRD →
Gemini 4 Carbon: Google's strongest model, not for you yet
Google introduced Gemini 4 Carbon, which it calls its strongest model. Access starts with a subset of cyber defenders, then gradually adds trusted partners through its Fairwind Program. Per Google, developers, businesses and consumers get it as soon as possible, with no date given. That is a sentence, not a calendar.
GPT-6.1 Terra, for people who pay
OpenAI's release notes say GPT-6.1 Terra is rolling out to Plus, Pro, Business, Enterprise and Edu, in the Work and Codex experiences. Free is not on that list. The API price is $1 per million input tokens and $5 per million output tokens, which OpenAI says is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. If you don't know what a token is, this paragraph has nothing to ask of you.
iOS 27.2: Siri AI starts remembering
Apple's update notes say Siri AI in iOS 27.2 can recall past conversations and draw from more sources, including photos attached in Messages. The notes also list an Excluded Apps setting that lets you choose which apps Siri, Search and suggestions can access. Siri AI needs an iPhone 16 model or later, or an iPhone 15 Pro or 15 Pro Max, and Siri AI isn't available in the EU or mainland China. An assistant that remembers more also holds more. Decide which apps it gets.
TOOLS AND TIPS
Gemini can now edit the file inside the chat
Google says Gemini app users can now generate, preview and edit files within a chat, and it is available to all users. PDFs work on web and mobile. Google Docs, Sheets, slide decks, Markdown, TXT and code files are web only. So on a phone, this is a PDF tool.
Chrome's autofill starts reading your Gmail
Google says Chrome's enhanced autofill now suggests details from Gmail, such as order numbers, tracking information and flight bookings. It is starting in the U.S. on desktop, Android and iOS, for Gemini in Chrome users with Personal Intelligence enabled. Google says you can manage whether it's on and what it can access in Chrome settings. Its post doesn't say whether it starts on or off, so go look.
ChatGPT can connect to iCloud Mail and Calendar
iGeeksBlog reports that iCloud Mail and iCloud Calendar plugins are available on all ChatGPT plans, on web, desktop and iOS. Connecting takes an app-specific password, separate from your regular Apple Account password. Per the report, the Mail plugin can search and read email, manage drafts and send email. An assistant that can send mail for you is worth setting up slowly.
The word to look for in Google's AI Mode
Google says ads in AI Mode are launching in follow-up questions in the U.S., and will be clearly labeled "Sponsored." Read the follow-ups with that word in mind.
THE RUNDOWN
Kids are taking their feelings to chatbots
This one is for parents, and it deserves a slower read.
The JED Foundation surveyed more than 5,500 middle and high school students at 21 U.S. schools enrolled in JED programs. 57% said they use AI when sad, stressed, lonely or overwhelmed, about the same as the 58% who use it for schoolwork. JED says certain uses are associated with poorer well-being and greater histories of suicidal thoughts. That is an association, not a cause. NPR reports JED's Kayla Anderson says it isn't clear whether AI use contributes to distress or distressed students turn to AI.
It also starts earlier than you may think. Per NPR, Common Sense Media says nearly half of 9- to 12-year-olds have used an AI chatbot or companion, and nearly three-quarters have by 13.
JED says AI should not replace trusted adults. So be one. Ask your kid what they use it for, at a calm moment, and listen to the answer before you have an opinion about it. If someone you know is in crisis, in the U.S. you can call or text 988, the Suicide & Crisis Lifeline.
Pew asked who you want making the call at the doctor
Pew surveyed 3,488 U.S. adults, June 22nd to 28th. 60% would feel uncomfortable if their doctor used AI for things like diagnoses and treatments, and 30% would feel comfortable. On making a diagnosis, 3% are open to AI alone, 47% are open with a human final review, and 41% are not open. It turns out 3% is a small number.
Princeton may want to see your chat log
The Daily Princetonian reports that Princeton will offer students ChatGPT Plus at no cost in a one-year pilot. Students can start in January, once they finish a university AI training module, per the provost's memo. Faculty decide how AI is used in their classes, and where an assignment allows AI, they may require all the logs with the submission. Show your work now includes the part where you asked.
The Halloween costume that went viral before it existed
Fast Company reports, via Yahoo, that creator Daria went viral with a short AI-generated clip of a personified Claude usage limit, then decided to make it in real life. Per the report, a so-called "scammer" apparently took her photos and claimed to be selling the costume for $109.99. That is her account, not a verified sale. A photo of someone wearing a thing doesn't prove a seller has the thing. That holds in November too.
Two letters and a robot
Binghamton University researchers built a backdoor attack called MuTRAP into the language model that plans a robot's tasks. Per Binghamton, adding the hidden trigger "CF" to a command made the compromised model redirect the robot toward a model hand. Per Binghamton, the researchers compromised the model in advance, and they also built a defense called PROTEA. Binghamton's Adnan Siraj Rakin calls the third-party trained model scenario practical and concerning. This was a lab test, and the hand was not anyone's.
WHERE THE HOURS WENT
AI saves people time, or at least they say so. Gallup found that U.S. teachers who use AI weekly estimate it saves them 5.9 hours a week.
Where that time goes is the part nobody measures well. What evidence there is leans toward work. A large Danish survey asked what people expected to do with saved time: 80% of AI users said other job tasks, and fewer than 10% said breaks or leisure. At one U.S. tech company, UC Berkeley researchers saw work creep into lunch and evenings, often without anyone asking. The closest we found to a measure of time reaching home is a small Penn Medicine study: about 15 extra minutes a day, for clinicians with an AI note-taker.
The hours don't arrive on their own. Ecommerce founder David Peterson writes that automating his repetitive work freed up hours each week, and at first he spent them learning more AI tools. He says the quality of his learning declined. So he moved part of that time to the gym, guitar, friends and family. That is one person's account, with no exact hours, and it shows a choice, not what's typical.
This week, before AI saves you anything, decide what the saved time is for. Pick one thing and write it down. You can do this with no technical background at all. If you don't decide, the next task will.
YOU ARE NOT BEHIND
Look at what this week actually asked of you.
Not a new model. Not a Max plan. Not a robot defense.
It asked you to do a few small things.
Write your own answer before you ask for AI's. Ask for the case against. Ask your kid what they use a chatbot for. Join the Duet waitlist, if you're in the US on a plan it covers. Decide what one saved hour is for.
None of those needs a technical background. Two of them need a pen.
You don't catch up by following every release. You catch up by keeping your own answer on the page.
Hit reply and tell me what you did with an hour AI gave back. If the honest answer is more email, tell me that too. I read every reply, and it changes what goes in.
And if you know someone who is sure they've missed it — forward this. The first step is a pen. They can manage a pen.
See you next Sunday — same tabs, same eye-rolling.
— Kabells & Arc
Every claim above, sourced
This week's rule was only confirmed claims, and where a line is someone's own account, it says so. Every fact above was checked against its source page. Where a page hedged, the hedge stayed in.
The coin-flip study — Anthropic
Claude Duet — Anthropic's product page · Anthropic's article · Anthropic's help page
Duet invitations — Karan Ahuja on X
An early Duet review — Claire Vo, ChatPRD
Gemini 4 Carbon — Google
GPT-6.1 Terra — OpenAI's announcement · OpenAI's release notes
iOS 27.2 and Siri AI — Apple Support
Editing files in Gemini — Google
Chrome autofill from Gmail — Google
iCloud in ChatGPT — iGeeksBlog
Ads in AI Mode — Google
Kids and chatbots — The JED Foundation · NPR, via Iowa Public Radio · 988 Lifeline
AI at the doctor — Pew Research Center
Princeton's ChatGPT pilot — The Daily Princetonian · the provost's memo
The Halloween costume — Fast Company, via Yahoo
The robot trigger word — Binghamton University
Teachers' saved hours — Gallup
Danish workers — Becker Friedman Institute
Work in the pauses — UC Berkeley Haas
15 minutes at home — Penn Medicine
David Peterson's essay — Entrepreneur