START HERE

Before you give AI the keys to your inbox

Connecting your email to an AI assistant sounds pretty handy. Ask a question, find the message, get on with your day.

There’s some fine print worth reading first. Google has clarified how connected information can be used, and “a summary” may contain more of an email than you’d expect.

Also today: what Anthropic’s hiring numbers tell us about junior jobs, a new way to build slides in ChatGPT, and a practical test for anyone tired of chasing delivery updates.

PRIVACY

Google’s email connection deserves a closer look

Google’s October 16 clarification says summaries, excerpts, and inferences drawn from connected content may be used to train its AI models. A subset may also be reviewed by people. And if an email is short enough, its “summary” could be the whole message.

That gives me pause. A quick question about a meeting could involve information you wouldn’t knowingly hand over for training.

Before connecting an account, read the connected-content settings and check Search Services History separately. Disconnecting an app doesn’t erase information already in your history. You can also use AI Mode and Ask Maps without connecting personal content.

This doesn’t mean Google automatically uses your entire inbox or photo library for training. The point is to understand what a connection permits before you switch it on.

Google’s clarification · Connected-content settings and details

WORK AND HIRING

Junior hiring is getting harder at Anthropic

Anthropic published a look at its own workforce on October 16. Junior candidates’ share of technical hires fell from 27% to 9%, even as total technical hiring grew.

One AI company can’t stand in for the whole job market. Still, these numbers are worth watching. Anthropic describes a growing emphasis on judgment and independence as agents handle more implementation work.

If you’re building a portfolio, show how you check the work. Explain a mistake you caught, why you chose one approach, and how you knew the finished result held up. Those details give an employer something to judge beyond a polished output.

Read Anthropic’s workforce study

MAKING THINGS

ChatGPT is getting a slide workspace

OpenAI’s October 16 release notes introduce Slides for eligible Pro, Business, and Enterprise users with access to Space. It brings presentation editing, comments, and exports into a shared workspace. Export options include PowerPoint, PDF, and Google Slides.

The useful part is being able to work on the presentation where you’re already developing the material. Start with an outline, settle the argument, then build the slides. Check figures, chart labels, and layout before you hand the deck to anyone.

Access is rolling out, so your account may not show it yet.

See OpenAI’s release notes

ALSO WORTH KNOWING

A few more things worth knowing

ChatGPT Voice has an update. OpenAI’s October 15 notes describe GPT-Live 1.1 and additional voice options, including Quip. If you use voice while walking or thinking through a problem, check your Voice settings. Availability varies. Live’s limits also differ from Advanced Voice; video and screen sharing aren’t available in Live.

Product notes · Voice features and limits

Apple tightened its rules for apps with user-generated content. Its October 16 update requires those apps to proactively prevent and remove nonconsensual intimate imagery, including AI-generated material. Apple also strengthened language about harmful features hidden during app review.

Apple’s announcement

For the developers here: Claude Code Review is generally available for Team and Enterprise customers. Check the billing settings before enabling it across a busy repository: reviews are billed separately, and new repositories default to a review after every push. A small pilot with a monthly limit is a sensible first step.

Code Review announcement · Costs and controls

AI WORKFLOWS

Try this: stop hunting through delivery emails

You have an order sheet. The supplier has sent six emails. Someone needs to know whether Friday’s installation is still happening.

This is a good candidate for AI help because the answer should be traceable to a small set of records. The assistant can do the digging; you check the evidence before promising a customer anything.

Here’s how I’d set up a first attempt.

1. Choose one supplier and one week. Give the assistant a short list of orders. Keep the first test small enough that you can check every result.
2. Use approved sources. If your workplace permits an email or Drive connection, specify which records it may read. Otherwise, provide a redacted export. Tell it to report anything it can’t access.
3. Ask for a dated answer with a source. For every order, you want the original expected arrival, the latest explicit promise, the message date, and a link or reference to that message. “Ships Friday” and “arrives Friday” are different commitments.
4. Check the awkward cases. Open an unchanged order, a revised date, and an order with no clear answer. Confirm the order number and quantity too. An unanswered question should stay unconfirmed.
5. Review any follow-up yourself. Let AI draft a clarification email. Read it before sending, and keep changes to customer commitments in human hands.

You can adapt this prompt:

Review these orders for [supplier] with arrivals expected during [week]. Use only [approved order sheet and email sources].

For every order, list its number, quantity, original expected arrival, latest explicitly promised arrival, and the date and source of that promise. Distinguish shipping dates from arrival dates. Mark missing or conflicting information as unconfirmed, and explain what needs checking.

Include every order, even if nothing changed. Report any source you couldn’t access. Draft questions for unresolved items, but leave emails unsent and records unchanged.

Only have five minutes? Pick one order whose answer you already know. Ask AI to find the latest promise, then open its source. You’ll learn quickly whether it can retrieve the right evidence.

This is a proposed workflow, not a tested case study. Judge it by the time you save after checking the results.

ChatGPT Work and Codex · OpenAI’s connection documentation

TOOL SPOTLIGHT

Tool spotlight: Parrot

For Mac users who answer the same questions on calls.

Parrot listens during a conversation and surfaces answers from material you supply, such as a FAQ or pricing document. Think of a consultant trying to recall a detail while staying present with a client.

The app is free and open source. Cloud AI services can add costs; local models through Ollama are another option. Check the current Mac requirements before installing.

There’s a privacy decision here, too. In cloud mode, transcripts and relevant document passages may go to the provider. Cloud transcription can also send call audio. Get the permissions your workplace and participants require before using it on a real call.

One detail that’s easy to miss: muting yourself in Zoom or Teams doesn’t mute Parrot’s separate microphone capture. Use Parrot’s own control.

I’d start with a mock call and a public FAQ. Turn off general-knowledge fallback so you can see whether the answers actually come from your document. Windows users can skip this one.

I haven’t tested Parrot firsthand; this spotlight is based on its site and documentation.

Visit Parrot · Read the documentation

YOUR NEXT STEP

What would you hand over?

There’s probably a small, recurring task you’d happily stop doing. Start there, with an answer you can check and a mistake you can easily fix.

What’s the task? And what would AI have to show you before you’d trust it with the job? Hit reply and tell me.

— Aric

mAIn Street · Useful AI for work and everyday life.