Daily AI · 2026-08-18
Useful AI Daily - August 18, 2026
Useful AI work today is less about chasing another model and more about knowing who controls the input, account, and workflow after you press go.
Run a 25-minute AI workflow audit: check what a service can train on, review active sessions, and turn one spreadsheet or routine into a small reversible experiment.
The Short Version
- Try: use Google Sheets Canvas on a low-risk spreadsheet to make one editable dashboard, then verify every write still lands in the original sheet.
- Watch: Relay's shutdown is a reminder that an AI automation can disappear or move inside a larger product. Export what makes your workflow work before you need it.
- Check: Twitch's reported training-data opt-out model makes settings and terms part of publishing. Revisit them before you share more source material.
- Ordinary-user bridge: review active sessions on the AI services that hold your chats and files. An unfamiliar device is a reason to revoke access, not to wait for a suspicious prompt.
- Builder signal: put export, revoke, and data-use controls in the product. A useful AI workflow should be understandable, reversible, and recoverable.
5 Updates Worth Your Time
Relay is shutting down as its team moves to Google's Chrome group
- What changed
- TechCrunch reports that Relay, an AI-powered workflow-automation tool launched in 2021, is shutting down while its founder and staff join Google's Chrome team. The founder said future Chrome AI plans would be shared later.
- Why it matters
- An automation service can be useful and still stop being a standalone product. Prompts, templates, approvals, files, and run history become liabilities if they exist only inside that service.
- Try, watch, or skip?
- Do not assume a future Chrome feature replaces Relay. Export reusable prompts, workflow steps, contacts, and output records now; name a human fallback for each important task and test one export in another approved tool.
Amazon's reported rare-book scanning is a signal to inspect data-use claims
- What changed
- TechCrunch, citing 404 Media, reports that Amazon is buying rare books, removing their bindings, and scanning them for AI training. Amazon said it purchases books through commercial channels to improve products and services.
- Why it matters
- Training-data provenance is not an abstract lab question. It affects creator trust, disclosure obligations, and whether people believe a service's data-use controls match what it says in public.
- Try, watch, or skip?
- Treat this as a reported signal, not a reason to assume every Amazon product uses the same practice. For your own work, keep customer and creator material out of training unless terms, permission, retention, and deletion controls are explicit.
Twitch's reported AI-training opt-out makes publishing controls worth a fresh look
- What changed
- WIRED reports that Twitch streamers can opt out of Amazon using their content to train AI models, prompting questions from creators about why that use is enabled unless they change a setting.
- Why it matters
- A creator's stream, draft, support recording, or private community content can be more valuable than a generic prompt. A buried default can change the practical boundary of what sharing means.
- Try, watch, or skip?
- Open the current settings and terms for any service where you publish source material. Record the data-use choice, the date, and the account that made it. Recheck after major terms or product changes.
Google Sheets Canvas turns rows into interactive mini-apps
- What changed
- Google says Sheets Canvas is now available globally in English. It uses Gemini prompts to create an interactive, read-write layer over a spreadsheet, with changes synchronized back to the source sheet.
- Why it matters
- This is a practical bridge for people who need a clearer tracker or decision board but do not want to build software. It can make a familiar spreadsheet easier to scan without creating another disconnected system.
- Try, watch, or skip?
- Start with a non-sensitive planning sheet. Ask for one small dashboard, make an edit in Canvas, and confirm the source rows update as expected. Do not use it for customer, legal, health, or unreleased data until your Workspace controls are clear.
Check active sessions before you trust an AI workspace
- What changed
- TechCrunch's current account-security guide notes that ChatGPT and Claude expose active-session controls, while Perplexity provides a sign-out-all-sessions path. It also recommends unique passwords and multi-factor authentication where available.
- Why it matters
- AI accounts may hold conversation history, uploaded documents, custom instructions, billing data, and linked services. A stolen session can expose more than a single chat.
- Try, watch, or skip?
- Review active sessions on every AI service you use for work. Terminate unfamiliar devices, sign out of all sessions if needed, use a unique password stored in a password manager, enable MFA where offered, and rotate exposed API keys.
Tool Worth Trying Today
A small Sheets Canvas decision board
Create a copy of a non-sensitive spreadsheet with five columns: task, owner, decision, next check, and fallback. Ask Sheets Canvas for a simple editable board, then verify that one Canvas edit updates the underlying row and that the sheet still works when Canvas is closed.
Best for: Individuals, creators, and small teams who already use Google Sheets and want a clearer weekly planning surface without moving their work into a new task system.
Watch out: Canvas is not a permission, retention, or backup plan. Keep sensitive personal, customer, legal, health, financial, unreleased, and private photo or document data out until Workspace terms, retention settings, and admin controls are clear.
Privacy / Cost Watch
- An opt-out setting is not a universal privacy guarantee. Read the current terms for training, retention, sharing, deletion, and regional availability before you upload creator or customer material.
- AI account sessions can expose chats, files, billing, and linked tools. Review devices, revoke unfamiliar sessions, and rotate API keys or passwords after a suspected compromise.
- Do not upload sensitive personal, customer, legal, health, financial, unreleased, school, location, or private photo and document data to a new AI feature until its terms, retention settings, and admin controls are clear.
- For real security, legal, medical, financial, election, or abuse incidents, preserve evidence and use official support or qualified help rather than relying on an AI answer.
One Practical Workflow
Run a 25-minute AI workflow exit and account check
- Pick one AI workflow you would struggle to replace and write down its input, output, owner, source data, account, and stop condition.
- Export the prompt, instructions, template, and recent non-sensitive output. Open the export outside the original service to prove it is readable.
- Review the service's current data-use setting, retention terms, and active sessions. Revoke anything you do not recognize and record the decision date.
- Run a small, non-sensitive version through a second approved tool or manual process. Measure whether you can still complete the task if the first service closes or changes terms.
- Keep one source-of-truth file, one named fallback, and one review date. Update them after any product shutdown, terms change, or account-security event.
Builder Note
AI products earn trust when users can leave, inspect, and stop them. Store prompts, run history, permission state, data-use choices, and outputs in exportable formats. Make session revocation, data deletion, and human fallback visible in the product, then test them before a shutdown or incident forces the exercise.
Ignore For Now
Ignore vague promises that another browser will absorb every workflow
A team move, an AI roadmap tease, or a polished demo does not preserve your current data, approvals, and integrations. Keep the service that works today, but verify the export and fallback before you move important work to a promise.
Bottom Line
Bottom line: make every AI workflow reviewable and recoverable. Check who can train on the input, who is signed into the account, where the instructions live, and how you finish the work if the product changes tomorrow.
Sources
- TechCrunch: AI automation startup Relay shuts down, staff joins Google's Chrome team
- TechCrunch: Amazon, which started off selling books, is destroying rare texts to train AI
- WIRED: Amazon Can Use Your Twitch Content to Train Its AI - Unless You Opt Out
- Google: Bring your spreadsheet data to life with Sheets canvas
- TechCrunch: How to tell if your AI platforms' accounts have been hacked