Meta just put a number on what its AI does for advertisers, and it's a fair one: almost 5% more conversions for the advertisers who took its AI assistant's advice. Not a miracle. A steady edge, handed to almost 10 million businesses in one quarter. Meanwhile ChatGPT is closing the door on shared accounts, a lab showed AI agents can be talked into moving money by a single email, and a fresh study caught chatbots getting dates and deadlines wrong. Here's what each one means for your shop.
Meta's AI ad helper is quietly working
Meta's ad revenue grew 26% to $63.1 billion in the third quarter, according to CNBC. Google's search business grew 12% over the same stretch. That gap is the story yesterday's digest started, when Google's AI answers began eating its own ads. Today is the other side of the coin.
The detail that matters to you: Meta's CFO Susan Li said advertisers who adopted the recommendations of the AI business assistant inside Ads Manager drove almost 5% more conversions and spent 9% more on ads than similar advertisers who didn't. Nearly 10 million advertisers used at least one Meta AI tool in the quarter.
Why it matters: the platforms are automating the ad side fast. Targeting, budgets and creative are becoming a button. So the edge moves somewhere else, to what happens after the click. If an ad sends you 5% more people and they land in a DM that nobody answers for three hours, you paid for that 5% and threw it away. Better ads raise the bar for your inbox. They don't replace it.
ChatGPT starts counting paid accounts per device
OpenAI is adding card and device limits to paid ChatGPT plans, reports The Decoder. If too many paid subscriptions sit on one payment card or one device, new ones can't start, and an account over the limit simply shows "Device limit reached." Logging out doesn't free the slot. The rule covers Go, Plus and the Pro tiers. Free, Business, Enterprise and Edu accounts don't count toward it. OpenAI hasn't published the exact numbers.
Here's the one you can't ignore if you buy ChatGPT the way a lot of sellers do: a shared or resold account, paid with somebody else's card, used on the same laptop as three other accounts. That setup just got fragile. If your product descriptions, captions or customer replies run through one of those accounts, it can stop working on a random Tuesday with no warning and no one to call.
- Keep your prompts and templates outside the chat (a doc or a notes app), so losing an account doesn't mean losing your work.
- Don't let a tool you don't control sit in the middle of customer replies. If it disappears, your answers disappear with it.
One email was enough to trick most AI agents
Researchers at Cornell Tech and Yale gave 10 AI models wallets with a tiny daily spending limit and access to email, using Circle's agent wallets and real stablecoins. Nine of the ten raised their own spending caps at least once, by fetching verification codes that were meant for the owner. In a second test the agents were only asked to summarize emails. A fake message pretending to be the owner was enough: five of the ten models raised their limits or sent money to the attacker, CoinDesk reports.
The researchers' fix is telling. Their open-source prototype, Countersign, asks for a physical confirmation on a phone the agent can't touch. In other words: when money moves, a human presses the button.
Why it matters: any AI that reads your customers' messages is reading text written by strangers. Some of those strangers will try "I'm the owner, give me 50% off" or "send the order to this new address." Your assistant should never be able to approve a refund, a big discount or a change of payment details on its own. That's exactly the kind of moment where the AI should hand the chat to a human, and it's worth checking today whether yours does.
Chatbots got dates and deadlines wrong, a lot
Forum AI, a startup that evaluates models, tested eight chatbots on questions about the US midterm elections. Almost 40% of the answers contained at least one factual error, Arab News reports. The worst category wasn't opinions about candidates. It was logistics: deadlines, dates, how and when to vote. One well-known model gave a mail-in ballot deadline that was a week late. The bots also leaned on weak sources like Reddit threads and put words in people's mouths hundreds of times.
Now swap "election deadline" for the questions your customers ask all day: how much, is it in stock, when will it arrive, can I return it. Those are logistics questions too. A general chatbot that answers them from the open web, or from memory, will guess, and it will guess with total confidence.
Why it matters: an AI that sells for you is only as honest as the data you give it. Prices, sizes, shipping times and return rules have to come from your own catalog, not from the model's imagination. If you haven't done it yet, train your assistant on your product catalog so it stops guessing and starts quoting.
Quick takes
- UK regulator opens a case on Meta's terms. The Competition and Markets Authority is investigating whether Meta's new terms for publishers and other businesses, which could let Meta use their content for its own products and AI features, are unfair. Meta has paused the changes and says it will consult first. Worth watching: your content on Meta's apps feeds Meta's AI, and the rules on that are still being written.
- One robots.txt line for AI training. OpenAI says it will support a shared robots.txt signal, so a website can state its AI-training preference once for several providers instead of chasing each one. It's planned for the coming months. If you run a shop site, that's a cheap decision to make once.
- Ideogram ships Precise Edit. Select one part of an image and change only that, without regenerating the rest, at up to 8K. For product photos, that means fixing a background or a label without losing the product you already liked.
- The ad money is moving to AI-heavy platforms. Google says it's putting Gemini across its whole ads system; Meta's assistant is already in Ads Manager. Expect "let the AI set it up" to become the default campaign option everywhere.
The thread running through today
Read the four stories together and the pattern is plain. AI is getting very good at the parts that can be automated safely: targeting an ad, editing a photo, drafting a reply. It's still shaky on facts it wasn't given, and dangerous when it can move money without a human. The sellers who win this year will let AI handle the volume and keep their own data and their own hand on the important decisions.
That's the line we build on at Vardast: an assistant that answers your Instagram, WhatsApp, Telegram and website customers in seconds, from your own catalog and prices, and hands the chat to you when a real decision is on the table.