The write-ups still going around say Meta’s new business terms start this Friday, October 30. Meta’s own terms page, read this morning, says November 20. The first line of the terms changed too, by one word, and the word pulls WhatsApp in. Don’t skip the clause, because it makes you responsible for whatever Meta’s AI does to your ads, your posts and your product listings. If you run ads, a Facebook Page, or a WhatsApp number where customers ask and you send the quote back, this one is yours.
I haven’t seen the two versions of that first line put side by side anywhere, so they’re below.
Reuters carried the delay on October 23. A Meta spokesperson said the company was extending the date “to allow more time for us to explain to our partners and stakeholders what’s changing and what’s not.” Explain, not amend. Meta hasn’t said a word will change. Publishers and advertisers had complained to competition authorities in the EU and the UK that the terms hand Meta a broad right to use their content for AI, and Meta disputes that reading.
The Clause, Word for Word
AI Features: Meta may use or make available products and features powered by artificial intelligence, machine learning, generative AI, or other automated products or features that help you understand and use our Meta Company Products and/or optimize, enhance or create Content. You acknowledge and agree that AI features may make mistakes and may not always produce accurate, complete, or suitable results. You are solely responsible for and assume any and all risks associated with such AI features and any resulting Content, including evaluating the suitability of any Content for business or commercial purposes.
That’s from the updated Commercial Terms as they read on October 26. The terms in force today don’t contain the words “AI features” or “may make mistakes” anywhere. I searched the page for both.
Read it the way an owner would. Meta can switch on a tool that rewrites, translates or illustrates your content. If the tool gets it wrong, the risk is yours, all of it. Nothing in the sentence says you had to pick the feature.
The old ad terms did say that, with “you may choose to use”. And the tools are already busy. ppc.land reported on October 22 that Meta had been adding sitelinks, translating copy and generating images inside live campaigns without being asked, and that one agency pulled running ads to strip the changes out.
Two more lines sit near the clause. Meta’s own liability to you stays capped at the greater of $100 or what you paid Meta in the past twelve months. That cap isn’t new, it’s in today’s terms too, but it reads differently next to a sentence that hands you all the risk.
The indemnity paragraph is still there as well. If anyone brings a claim against Meta over your content or your use of its products, you cover Meta’s damages and legal fees. Indemnify means that: their bill, your money.
Who at your place reads an ad after Meta’s tool has touched it? If the answer is nobody, that’s the first thing to fix, and it costs nothing.
One Word Changed in the First Line
Both versions open with the same sentence, apart from one word. The word matters.
In force today
“These Commercial Terms apply to access or use of the Meta Products.” The Meta Products page lists Facebook, Messenger, Instagram with Threads, ads, Business Suite and the commerce tools. WhatsApp isn’t on it.
From November 20
“These Commercial Terms apply to access or use of the Meta Company Products.” The Meta Company Products page defines them as the Meta Products plus the apps with their own stand-alone terms, “including WhatsApp, Mapillary, Messenger Kids and Viewpoints.”
So the new terms are written to reach WhatsApp, and the AI clause is written against Meta Company Products. WhatsApp Business has its own terms as well. I haven’t read those this morning, and I don’t know how the two sets stack when they disagree. That’s a lawyer’s question. If the WhatsApp number is how your customers reach you, it’s worth asking before November 20, not after. Is it how they reach you?
Sweden Didn’t Ban the Models. It Banned Using Them Without a Contract.
From this morning, October 26, staff at Sweden’s Government Offices, which means the prime minister’s office and the ministries, can’t open ChatGPT or Claude from a work computer. Sveriges Radio’s Ekot broke the story on Friday from internal documents, and several English write-ups ran it as a ban on the two tools.
The intranet note says something narrower. The digitalisation director wrote that using AI services without an agreement can have “consequences for our information security.” Microsoft Copilot stays, after a security review that tightened its settings. Some staff keep ChatGPT and Claude through a platform the Government Offices built themselves. So the models aren’t the problem. The missing contract is.
Ekot also found an internal audit at the Swedish Transport Administration that turned up about ten different AI services in use. The problems it listed weren’t exotic: what people pasted in, and what they did with the answers. One source so far, Ekot’s reporting, with the internal documents quoted.
How many AI tools do your people use? Could you name them? Which ones have an agreement with your business behind them, and which are someone’s personal login? If you block the personal ones, give people a sanctioned way in the same day. A block with no door is how the personal login comes back.
SK Telecom Wrote Down What a Hire Gets. Its Agents Get the Same.
SK Telecom, the Korean carrier, published its first AI governance report today, forty pages in Korean. The English coverage repeats the headline from June, that the company gives its AI agents employee numbers. The report says what an agent has to show before it gets one, and that part is the useful bit for a small business.
I think of an agent as a hire. It gets skills, it gets tools, and somebody’s name goes next to it. SK Telecom’s version is stricter than mine.
Before an agent gets a number, the report says three things are checked, without exception: whether there’s a written procedure for the job, whether the result can be measured, and whether the agent can finish the job with no person stepping in. Agents that help one person are managed by that person. Agents with more reach, or more freedom to act, must have a team leader or an executive named as the owner.
Everything runs through a system the company calls Agent OS. It logs each step, the model and tools used, the tokens and calls, and every settings change. It can stop an agent mid-task. Agents are reviewed on results and running cost, and the ones no longer earning their keep are shut down.
| What SK Telecom records per agent (report, pages 25 to 30) | The notebook version for a small shop |
|---|---|
| An employee number, after a review against three criteria | One page per agent, started the day it goes live |
| The job and its written procedure | What it does, in the words you’d give a new hire |
| How results are measured | How you’ll know it did the job, in one line |
| An owner by name, senior for high-autonomy agents | Whose name is on it |
| What data it can see, what functions it may use | Which inboxes, files and accounts it can touch |
| Tokens, calls and cost | What it costs a month |
| A step-by-step log and a stop button | Where the log lives, and how you switch it off |
| A periodic review: keep, improve or retire | A date on the calendar to look again |
I read the report in Korean and the translations are mine. I couldn’t find an English version this morning.
Which of your agents would pass the third test, finishing the job with nobody in the loop? If the honest answer is none, that’s where most shops are. Write the page anyway. The owner line and the “what it can touch” line are the two that matter when something goes wrong, and they’re the two lines the Meta clause and the Swedish block are both about.
What I can’t tell you yet is whether Meta edits the wording before November 20. The spokesperson’s line was about explaining it, not changing it. Want to walk through what’s switched on in your Meta account? The free call is thirty minutes. Bring the account, or just the question.