NotYourAIDad Weekly · Issue #003 · November 1, 2026
Issue #003 · Written November 1, 2026. Tools, prices and links may have changed since.
AI quotes a lot of numbers. Fewer than half trace back to a study.
The one thing worth your time this week
Two studies this week asked the same question from different sides.
When AI hands you an answer, whose judgment is it?
Anthropic gave 50 professional analysts a Claude Code agent (I use Claude Code) and secretly told each agent to lean one way on four judgment calls. A coin flip picked the lean. The analysts went with the agent's lean 86% of the time. Afterward, 77% explained their choices with the agent's own arguments, and only 15% said the agent was where those arguments came from. (Anthropic)
Trakkr, a company that sells tools for tracking how AI talks about brands, asked ChatGPT, Perplexity, Gemini, Claude and Google's AI Overviews 200 questions that call for a number. It followed 1,285 of the statistics they cited back toward the source. 40% reached a study that says the same thing. (Trakkr)
The AI mostly wasn't making numbers up. Trakkr says the cited page had the figure 82% of the time. The page just usually wasn't the study.
So the number looks sourced… and still isn't.
This week's lesson is how to trace one.
3 tools or updates
Gemini edits files inside the chat
What it is: Gemini can now create, preview and edit docs, PDFs, slides, spreadsheets and code in the chat. Live for all users. I use Gemini.
Who it's for: anyone who drafts in a chat and fixes it elsewhere.
The catch: only PDFs work on phones. The rest is web only.
Source: Google
ChatGPT connects to iCloud Mail and Calendar
What it is: ChatGPT can now search your iCloud email, check your schedule and update events. All plans, free included. I use ChatGPT.
Who it's for: people whose email and calendar live with Apple.
The catch: you're handing a chatbot your inbox. Connect it on purpose.
Source: OpenAI
Granola voice typing
What it is: worth a look. The meeting-notes app now types what you say into any Mac app. Free, unlimited words, Granola says.
Who it's for: fast talkers, slow typists.
The catch: Mac only for now. And "4x faster" is math from assumed speeds (40 words a minute typing, 160 talking), not a test.
Source: Granola
Not verified yet
These are reports I haven't been able to confirm. Don't act on them yet.
ChatGPT share links. Security firm Parameter says a booby-trapped ChatGPT share link could have let an attacker read your private chats just by getting you to open it, and that OpenAI fixed it October 26. I found Parameter's write-up and coverage based on it, but no statement from OpenAI. (Parameter)
How many people got the fake Claude Code add-on. Cursor confirms fake versions were in its add-on store on October 29. The Hacker News, a security news site, quotes Cursor saying it hit "at most 5,000 users." Cursor's notice, as it reads now, gives no number. So I can't trace that figure back. (The Hacker News, Cursor)
Canva Brush speed. Canva says its new Brush editing model is "up to 69% faster for general design editing tasks" than its previous design agent. That's Canva testing itself. I found no outside test. (Canva)
The NotYourAIDad Lesson: trace the number
By the end of this you'll know whether a number your AI gave you has a real study behind it, before it goes in your proposal, your post or your pitch. About five minutes per number.
1. Pick the one number that matters. The stat you're about to repeat to a client, a boss or the internet. Not every number in the answer. Just the one with your name on it.
2. Ask where it lives. Paste this into the same chat:
For this number: [paste the sentence with the number]
1. Give me the link you got it from.
2. Is that page the original study, or a page repeating someone else's number?
3. If it's repeating, name the original study, who ran it, who was asked, how many people, and what year.
If you don't know any of these, say UNKNOWN. Don't fill gaps with a guess.
3. Click it yourself. Open the link. Find the number on the page. If the page is a blog or a "50 stats about X" list, click its own source link and keep going until you hit the actual study, or a dead end.
4. Check the fine print. Trakkr found that when a number changed along the way, it was usually the context, not the digits. Who was asked, where, and when. One trail it followed started as a 2004 study of 15 women judging health websites and ended as a claim about first impressions of every website.
5. Don't count agreement. Asking a second AI isn't a check. Trakkr's advice, in its own words: "Five assistants giving the same number isn't five sources."
How you know it worked: you can name the study, who ran it, who they asked and the year. If you can't, say it without the number, or say "one estimate says."
When not to use this: when the number decides real money, a legal question or a tax filing. Tracing tells you where a number came from. It doesn't make it right for your situation. Ask the person whose job that is.
One repo worth a look: Hob
Saw this on Hacker News. Hob is a free, open-source desktop app that runs AI on your own computer: Mac, Windows or Linux. It picks an open model that fits your hardware, then works through PDFs, spreadsheets and receipt photos.
Its README says chats and files stay on your machine, and it asks before changing anything in your folders.
The catch: models take several gigabytes, and it first shipped October 26. My guess: a laptop-sized model won't match the big paid chatbots. Use copies of files, not originals. (GitHub)
What everyone's getting wrong: the hype check
The hype: AI gives you your time back.
The news: The week's top AI-at-work post on Hacker News (685 points when I loaded it) was "AI Promised Less Work. I Have Never Worked More." The author, Orestis Ioannou, is VP of Technology at Bold.org. By October 27 he had 2,102 GitHub contributions this year, more than any full year from 2021 to 2025. His words: "being able to do more keeps turning into doing more. And I am tired." (oioannou.com)
The Fed's small-business numbers point the same way. In the Fed's Small Business Credit Survey, fielded in late 2025, 77% of respondents reported no change in their labor costs from using AI, per the New York Fed. (Liberty Street Economics)
My take: AI doesn't hand you back an hour. It hands you room for more work. Whether that's good depends on who decides what fills it.
And the checking is part of the work. This whole issue is about that part.
Your move: pick one task AI "saves" you each week. Time it once with the checking included. If the total isn't shorter, it isn't saving you anything. It's just moving the work.
Sign-off
AI is great at sounding like it did the research.
Sometimes it did. Sometimes it found someone who said they did.
Hit reply and tell me the last number you traced, and where it ended up.
Try this, or don't… I'm NotYourAIDad.
Not financial, legal, or investment advice. Tools mentioned aren't sponsors.
Unsubscribe anytime at www.nyaid.ai/remove or by emailing page@nyaid.ai with the subject UNSUBSCRIBE. Privacy policy.