In partnership with
Your site is losing leads. You can't see where.
Your site looks fine to you — and it's still losing leads. SureThing renders any URL in a real browser, scores 9 dimensions, and ranks the fixes by wasted impact.
Beginners in AI
Good morning and thank you for joining us again!
Meta says its Muse agent can phone a business and book things for you. Internal posts show some of those calls were made by people in a call center, and a tester only found out afterward. Today's lead covers what that means for anything you ask Muse to say for you. As always, every story here is human picked and edited.
THE FRONT PAGE
Meta's AI Agent Handed Its Phone Calls to People
TLDR: Meta has been testing a "human concierge" for Muse, so some calls its AI agent places to book haircuts or check store stock are made by contract workers, and at least one tester wasn't told until the call was over.
The Story:
Meta has promoted Muse as an agent that can call a business for you to make a reservation or an appointment. Internal posts seen by 404 Media and Reuters show Meta "added a human agent layer for calls to get completed," and last week it switched that on for half of its employees. Contractors placed the calls and handled requests like haircut bookings, inventory checks and contractor quotes. One employee who tested it learned the caller was human only afterward.
Its Significance:
Meta's own staff saw this coming. "This has potential for so much negative PR," one wrote, and another warned of "a ton of outcry" if it launched switched on by default. The bigger issue for you is where your words end up. Ask Muse to call your doctor's office and a stranger may hear the details, with nothing on screen saying so. Meta says it'll only roll this out "with the proper disclosures." Until you see them, keep personal details out of Muse's calls, and update the Mac app today. meta’s not alone. Google was caught doing something similar in its Waymo self-driving vehicles, using humans overseas to take over and drive at times without letting the passenger know.
QUICK TAKES
The story: Anthropic released Claude Opus 5.5 on Monday, with output down to $20 per million tokens from $25. About 90 minutes later OpenAI launched GPT-6 Sol and Luna at roughly half the price of its 5.6 models, and says Sol makes about half as many factual mistakes as the model before it.
Your takeaway: Both companies grade their own tests, so wait for outside results before crowning anyone. The bump in efficiency is for the price is what you’ll notice most. I've been testing both models and they are noticeably better. Anthropic is also giving a free reset for the month in the settings.
The story: Tesla put xAI's Grok Bot in its cars on Monday. Say "Hey Grok" and it can reorder your usual coffee, book appointments, handle your Gmail and calendar, buy things online and place phone calls, all hands-free.
Your takeaway: This is an assistant that does the errand instead of just answering a question. For now it's a beta for SuperGrok Heavy subscribers at $300 a month on cars with the AMD chip, and Tesla says a deeper tie-in with Full Self-Driving is next.
The story: France, which chairs the Council this month, called a session on AI and international security for September 23. Sam Altman is expected in person, alongside representatives from Anthropic, DeepSeek and Moonshot AI.
Your takeaway: American and Chinese AI companies have never briefed the Council together before. It lands one day before Trump and Xi meet in Washington, where the AI incident hotline from Monday's edition is on the table.
TOOLS ON OUR RADAR
🛡️ Proton VPN Freemium: Hide your browsing from public Wi-Fi and your internet provider, free on one device with servers in 10 countries, a no-logs policy and no credit card.
🎬 DaVinci Resolve Freemium: Edit video, grade color and mix audio in one app, and the free version is extremely powerful. The $295 Studio version now connects straight to AI assistants like Claude, ChatGPT and Grok over MCP, so you describe the edit in plain English and it does the work.
🎧 Adobe Podcast Enhance Freemium: Clean up a noisy voice recording in your browser with one click, free for files up to 30 minutes and one hour a day.
🗺️ Organic Maps Free and Open Source: Get directions offline on iPhone and Android with no ads, alternative to Google Maps. After today's lead about where your requests end up, Organic Maps says it collects no data, and two outside privacy auditors found no trackers in the app.
TRENDING
OpenAI fired contractors for using AI to train its AI. Raters hired through Mercor to grade ChatGPT answers weren't allowed to use chatbots, Grammarly or AI detectors. Reviewers caught them by spotting repeated phrasing, fast work and too many em dashes. OpenAI declined to comment.
Snorkel AI nearly tripled its value to $3.5 billion. It raised $350 million led by Insight Partners and S32. Its yearly revenue pace hit $375 million, up 18 times in a year, from selling finished training data to AI labs.
Malware that asks four chatbots what to do next. Cisco Talos found CLOSEDQUORUM, Windows malware that has DeepSeek, Qwen, Mistral and Gemini vote on its next move, then does whatever wins. It goes after passwords and crypto wallets. The sample Talos studied had placeholder keys, so it couldn't run as found.
AstroForge is putting an AI pilot on its next spacecraft. Its model, Solo, reads about 2,500 sensors and fixes problems without waiting on Earth. It flies in "shadow mode" late this year, then takes control in 2027.
Qualcomm's new phone chip runs a 30 billion parameter AI model offline. That's the Snapdragon 8 Elite Extreme Gen 6. Apple's on-device model from June is 20 billion. The Motorola Signature 27 is the first phone to use it.
TRY THIS PROMPT (copy and paste into Claude, Grok, ChatGPT, or Gemini)
📞 Name any AI service. Find out if a person is on the other end.
Build me a single-file HTML web app called "Who's on the Line?" that tells me whether an AI service uses AI, human workers, or a mix of both, based on live web search.
Vanilla HTML, CSS and JavaScript in one file. No frameworks, no build step, no API key from the user.
THE IDEA
Some services sold as AI have people doing part of the work. In one recent test, Meta's Muse agent had its phone calls placed by human contractors. I name a service and the part I care about, and the app tells me who is doing the talking.
FORM
- Text input: "AI service or app" (placeholder Meta Muse)
- Dropdown: "What part do you care about" with options: Phone calls it makes for me / Chat and customer support / Voice it uses to talk to me / Reviewing what I type or say / Everything it does
- Submit button labeled "Check who's talking", disabled while loading
HOW IT GETS THE ANSWER
One call to the Anthropic Messages API with web search turned on:
fetch('https://api.anthropic.com/v1/messages', {method:'POST', headers:{'Content-Type':'application/json'}, body: JSON.stringify({model:'claude-sonnet-4-6', max_tokens:2400, system: SYSTEM, tools:[{type:'web_search_20250305', name:'web_search'}], messages:[{role:'user', content: serviceAndPartAndTodaysDate}]})})
Parse with (data.content||[]).filter(b=>b&&typeof b.text==='string').map(b=>b.text).join('\n'), strip ```json fences, slice from the first { to the last }, then JSON.parse. Wrap it in try/catch with a friendly inline error box.
The system prompt must tell the model to: search news reporting, help pages, privacy policies, terms of service and job listings with terms like human agents, contractors, call center, human review and handoff; never invent a fact, date, number, quote or source; never estimate a percentage split unless a source states one; use verdict "unknown" when reporting is thin; make every source URL a real page it saw in search; reply with JSON only using exactly these keys:
{"service":string,"company":string,"verdict":"ai"|"human"|"mix"|"unknown","verdict_line":string,"ai_weight":number,"human_weight":number,"steps":[{"step":string,"who":"ai"|"human"|"mix"|"unknown","detail":string}],"disclosure":string,"timeline":[{"date":string,"text":string}],"what_to_ask":string,"sources":[{"title":string,"url":string}]}
ai_weight and human_weight are rough 0 to 10 visual weights, not a measured split. steps lists 2 to 5 parts of the service in the order a user meets them. timeline has up to 4 dated reports, newest first. Plain English, and never use an em dash or an en dash.
RENDER STRUCTURE
1. Verdict card: a mono uppercase label ("AI does the talking", "People do the talking", "A mix of AI and people", "Not enough reporting"), the service and company on the right, the verdict line in italic serif, and a two-color bar for the AI and human weights. The border color follows the verdict.
2. "Who handles each part": one card per step with a colored left border and a badge (AI, Human, Both, Unclear).
3. "Do they tell you?": the disclosure text.
4. "What's been reported": a vertical timeline with dated entries.
5. "Before you share anything private": the what_to_ask text in a tinted box.
6. "Sources": real clickable links that open in a new tab.
AESTHETIC
Near-black background (#0b0f14) with soft radial glows: violet #a78bfa for AI in the top-left and peach #ffb088 for humans in the bottom-right. Gold #e8d27a for "mix" and gray for "unclear". The title and button use a violet-to-peach gradient. Fonts from Google Fonts: Sora for headings and UI, Fraunces italic for explanations, JetBrains Mono for labels and dates. Translucent cards with 1px hairline borders and 14px radius. A small pill above the title reading "Is anybody really there".
DETAILS
- Include an escHtml() helper and run every piece of API text through it before inserting it into the page.
- Add [hidden]{display:none!important} to the CSS so the loading state can't linger.
- Loading state: a ringing phone emoji with "Listening in on the line..."
- Results fade and rise in.
- Footer: "Grounded in live web search. Companies change how they staff things often, so re-check before you share anything private."
- No em dashes or en dashes anywhere in the copy.What this does: Type the name of an AI service and pick the part you care about, like the phone calls it makes or its chat support. It searches the live web for news reports, help pages and privacy policies, then tells you whether that part runs on AI, on people, or on both.
What this looks like:
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Place a phone call for you to book a haircut or check whether a store has something in stock, at least in Meta's pitch.
❌ Still Can't: Finish those calls reliably on its own. Meta added human call-center workers "for calls to get completed."
✅ AI Can Now: Trace a spacecraft fault on its own, linking a lost position fix to a power problem in the star tracker.
❌ Still Can't: Get trusted with the controls on day one. Solo spends its first mission in shadow mode while the regular system flies.
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
An anthropologist and a computer scientist at Microsoft Research spent five years studying workers in the US and India, and published this in 2019. Their finding is that a lot of what gets sold as automation depends on people doing small tasks out of sight, labeling, checking and stepping in when the software gets stuck.
How was today's edition? Tap one below. Pick "Could be better" and tell me what to fix; I read every answer.
How was today's edition?
Thank you for reading. We’re all beginners in something. With that in mind, your questions and feedback are always welcome and I read every single email!
-James
By the way, this is the link if you liked the content and want to share with a friend.
Some * designated product links may be affiliate or referral links. As an Amazon Associate, I earn from qualifying purchases. This helps support the newsletter at no extra cost to you and Amazon makes a tiny hair less.
