When Meta's AI Errs, Its Terms Say You're Liable

Remove the stray internal note at the top (“Fragment is accurate, no issues found.”) and output the corrected body.

One Sentence Meta Wrote Into Its Terms

Meta’s revised Meta Commercial Terms¹ now include a new clause titled “AI Features.” It comes down to two sentences. One says AI features can make mistakes and won’t always produce results that are accurate, complete, or appropriate. The other says the business using those AI features bears sole responsibility for every risk arising from the feature and its output.

Put plainly: if Meta’s AI botches an ad image, the legal fault falls on the customer who ran the ad.

This already happened. Last June, the American outdoor retailer REI pulled an Instagram ad. Meta’s AI had edited a bike photo and, in the process, added an extra set of handlebars and deleted a crank arm. REI said it had been enrolled in that AI tool without its knowledge. Under the new terms, the risk of an accident like this one falls on REI.

The revised terms were originally set to take effect on October 30. But after media outlets and advertisers raised concerns with competition authorities in the EU and the UK, Meta told Reuters on October 23 that it would push the effective date back three weeks, to November 20. Meta’s stated reason wasn’t that it planned to rewrite the language — it was that it wanted time to explain to partners what was changing and what wasn’t.

I read this episode as a snapshot of how much terms of service and law now matter for AI. However smart a model is, a single line in a contract decides who pays when it gets things wrong. Today I want to walk through three clauses in Meta’s terms centered on that one line, then read the same sentences again through the lens of Korean law.

---

Running a Page Alone Makes You a Party to These Terms

The name “Commercial Terms” makes it sound like it only applies to companies spending on ads, but the revision defines business use far more broadly. It covers writing ads or marketing messages, selling or listing products, building product catalogs, managing Pages, running business groups, and using Ads Manager or Meta Business Suite. Even a neighborhood café that has never run a single ad but simply operates an Instagram business account is a party to these terms.

It’s worth understanding how consent works here, too. Article 7 of the revision states that continuing to use Meta products for business purposes after the update notice counts as agreeing to the new terms.² If you don’t agree, you’re told to stop all business-purpose use entirely. There’s no partial opt-in — like keeping your Page but rejecting just the AI clause.

The AI features clause didn’t exist in the current terms that have been in effect since March 2024. Beyond the two sentences already mentioned, two more provisions are included: the business bears responsibility for judging whether an output fits its business purpose, and businesses must not alter or remove any watermark or content-authentication metadata that AI attaches to its outputs. There’s no separate list specifying which AI features this clause refers to. As of the evening of October 23, this clause remained word-for-word identical to the pre-revision text.

The One Who Flips the Switch Isn’t the One Who Pays

To be fair, Meta isn’t the only company that puts the responsibility for AI output on the user. Terms of service for AI-equipped software commonly include lines like “provided as-is” or “it’s up to the user to judge whether the output meets their needs.” Everyone knows generative AI can get things wrong, and it’s understandable that a provider can’t guarantee every single result.

What makes Meta’s case different is who actually controls the switch. According to trade publication Adweek’s reading of the new AI feature terms, the clause covers not just features advertisers choose for themselves, but also features Meta turns on unilaterally. The creative-enhancement function in Meta’s ad-automation tool Advantage+³ adjusts an ad image’s background, brightness, contrast, and format, and can generate up to 150 variations from material an advertiser has uploaded. Agencies report experiencing Meta attaching site links to their ads, translating ad copy, and inserting AI-generated images and carousels built from brand photos.

The US agency Exverus pulled ads that had already cleared multiple rounds of quality review and had been running live, stripping out the site links, translations, and AI images Meta had added. Online bookseller Bookshop.org said it was automatically enrolled in the new feature during an account update, which caused AI illustrations to appear in its ads — and that it has since turned off every one of Meta’s AI ad features.

Meta’s position is that advertisers can switch off each Advantage+ feature individually, at any time, from Ads Manager. But Exverus points out that many of those toggles sit hidden inside collapsed menus, and stay active until an advertiser unchecks them one by one. The features are on by default, the off-switch is hard to find, and the advertiser alone bears the risk of whatever comes out the other end.

Security expert Mike Wilkes of Aikido Security, in comments to CIO.com, framed it this way: if an animal-welfare organization asks Meta to generate a dog image and gets back something inappropriate, Meta is the only party that controls the model that produced it — yet the organization is the one left holding responsibility. Analyst Frank Dickson put it more bluntly: Meta builds the AI feature, admits it can make mistakes, and then hands that risk off to the customer — an inversion of where accountability should actually sit.

Crawling and Licensing: Only the Date Got Pushed Back

Beside the AI-features clause sit two more provisions. Most of the overseas coverage zeroed in on these two, since this is where the clash between publishers and Meta actually played out.

The first is website crawling. The revised terms state that Meta, or anyone acting on Meta’s behalf, may use automated software such as scrapers, bots, and crawlers, create test credentials⁴, and access and regularly pull updates from domains and webpages “related to” the business’s products or services. The language first made public in mid-October also said that content gathered this way could be used to “train” Meta’s products. When the backlash grew, Meta deleted the word “train,” arguing that the crawling clause only ever applied to ads and business products.

Legal experts still see the revised wording as sweeping. Matt Topic, a media and intellectual-property lawyer, points out that the word “develop” alone is broad enough to cover training, and that “business products” isn’t even a term defined anywhere in the terms. The test-credentials language remains untouched, and robots.txt⁵ — which Meta kept invoking in its statements — appears nowhere in the actual text of the terms. As I covered in Issue 215, robots.txt is a gentleman’s agreement with no legal force. When a signed contract runs up against a gentleman’s agreement, you can guess which one wins.

The second provision concerns content licensing. Businesses grant Meta a license to content they post on or “in connection with” Meta’s products — and this extends to content posted on a business’s behalf by agencies or outside contractors. According to CIO.com, Meta revised a different set of terms back in July to make this license perpetual, adding language stating that the license survives termination regardless of which party ends the relationship or why. Noah Kenney of the consultancy Digital520 interprets this to mean the license doesn’t end when the contract ends — it ends only when the content is deleted from Meta’s systems.

Also unchanged: the clause requiring businesses to indemnify and hold Meta harmless against “any and all” claims related to using Meta’s products, and the clause letting Meta change or discontinue products without prior notice or liability. Thomas Randall of Info-Tech Research Group argues that taken together — continued-use-as-consent, sweeping indemnification, and as-is provision of service — this amounts to something close to a one-way street.

Reading the Same Clause Under Korean Law

This is where the matter turns into a legal problem. Meta’s clause was written with American and European publishers and advertisers in mind, but Korea has a law that speaks directly to this kind of boilerplate.

Article 7 of the Act on the Regulation of Terms and Conditions (the “Terms and Conditions Act”)⁶ voids several types of clauses concerning the liability of contracting parties. Clause 2 of that article overlaps head-on with the provision at issue here. A clause that, without reasonable grounds, limits a business’s scope of liability for damages or shifts onto the customer a risk the business should bear is void. Article 6 presumes that any clause unfairly disadvantageous to the customer, or one the customer could not reasonably anticipate, has lost its fairness. Under Article 15 of the Act and its enforcement decree, the only industries exempted from Article 7 are international transport, international finance and insurance, and trade insurance — advertising platforms are not on that list.

Nor is this Korea’s first encounter with the issue. In March 2019, the Korea Fair Trade Commission reviewed the terms of Google, Facebook, Naver, and Kakao and demanded corrections. Facebook’s terms at the time contained a clause under which the license to a user’s content survived even after the user deleted it; the Commission held that deleting the content should terminate the license as well. As for Naver’s sweeping liability-waiver clause, the Commission found it inconsistent with the Civil Act, which imposes liability for damages where there is intent or negligence. Issues litigated 7 years ago are now returning to the fine print, this time wearing the label “AI.”

The Framework Act on Artificial Intelligence is also in play. Article 31 of the Framework Act on Artificial Intelligence, which took effect in January this year, obligates an “AI business operator”⁷ to disclose that generative AI output was in fact AI-generated. Article 2 of the same Act splits AI business operators into “development operators,” who build and supply the AI, and “user operators,” who use that AI to provide products or services. If the AI Meta switched on by default generated the ad image, who bears the disclosure obligation and liability for that output under the law’s categories — and who bears it under the contract’s categories? Meta’s instruction not to remove the watermark aligns with the direction of this law, but whether that alignment matches the way the contract shifts all the risk onto the customer is a separate question that needs scrutiny.

Still, one cannot simply assume these provisions apply to Meta’s terms as written. Which country’s law Meta’s commercial terms designate as governing law, which courts they designate as having jurisdiction, and how far Korean law reaches into a contract signed by a Korean company — these are all questions that need to be checked independently. Whether “reasonable grounds” existed is likewise something that has to be contested case by case. These are precisely the questions a lawyer should answer — not a marketing team.

This piece organizes publicly available contract language, news coverage, and statutory text into a commentary; it is not legal advice. If a contractual dispute is at stake, consult an expert first.

What to Check Before November 20

The 3-week delay gives Meta time to explain itself, but it gives businesses that run on Meta time to check their own standing.

1. Count how many Meta assets carry your brand. Aman Mahapatra of the tech consultancy Tribeca Softek says most CIOs can’t say how many Meta pages, ad accounts, or WhatsApp numbers bear their company’s name — because marketing teams and agencies created them, and IT never kept a list.
2. Open the Advantage+ creative settings in Ads Manager, campaign by campaign. Check which enhancement features are turned on, and whether any were newly activated by a recent account update.
3. Reread the liability clauses in your agency contracts. Meta’s terms hold the business responsible even for content an agency uploads. If your agency contract is silent on who reviews AI-altered creative and who’s liable when something goes wrong, that blank gets filled in by your company — by default.
4. Audit robots.txt and anything behind a login wall. Meta runs separate crawlers for AI training/indexing, advertising, and AI search, so rules can be set per use case. If a members-only area is tied to Meta assets, it’s worth having someone review exactly what the test-credential clause permits.

The terms preview page has been showing different effective dates depending on the visitor’s region. As of the evening of October 23, the UK and French versions were still listing October 30 as the effective date. Saving a copy of the current wording now will let you compare it against whatever changes later.

Oswarld’s Lens

Meta’s terms of service state that if its AI makes a mistake, legally, that’s the customer’s problem. I think this single sentence shows just how important AI-related terms of service and legislation are about to become. This is the time for lawyers and the law.

Until now, when people talked about adopting AI, the questions were mostly about which model to use, how accurate it is, and how much it costs. Meta’s terms of service tack on one more question behind all of that: who pays when it’s wrong? And the answer to that question isn’t found in a model’s performance sheet — it’s found in contracts and statutes.

This question won’t end with Meta’s ads, either. Most AI features that businesses use are arriving as defaults that platforms simply switch on, and those platforms’ terms of service could carry the same kind of language. At that point, whether someone inside your organization can read the liability clauses in a contract, identify where the law draws the line, and find room to negotiate becomes just as important as how well you use AI itself.

Closing

Meta’s commercial terms now take effect on November 20th—pushed back three weeks—but as far as I can tell, nothing substantive has changed: the clause that puts liability for AI features on the business, the clause allowing crawling of “related” sites, and the license that survives termination are all still there. The Federation of European Publishers isn’t asking for a delay; it’s asking for the policy to be scrapped entirely, and voices in the US Senate have called for withdrawal too. In Korea, coverage has been limited to CIO Korea and ITWorld reporting the news, and I haven’t yet found any report of a domestic company or institution formally raising objections.

---

💬 At Reader‘s company, who actually reads the liability clauses in terms of service or contracts when you adopt an AI tool? Legal? Procurement? Or does no one read them? Let me know in the comments.

📨 If you have a colleague who runs ads on Meta or reviews AI tool contracts, send this their way.

References & Further Reading

Primary sources

• Evan Schuman, “Meta’s new commercial terms say if its AI makes a mistake, it will legally be the customer’s fault,” CIO, October 19, 2026. Read article ··· This is where today’s piece started. Keep in mind it was published before Meta revised the crawling clause and pushed back the effective date.
• Reuters, “Meta to Delay New Commercial Terms Criticised by Publishers, Advertisers,” October 23, 2026. Read article ··· Contains Meta’s official explanation for pushing the effective date to November 20th.
• Luis Rijo, “Meta delays website-crawling terms to Nov 20 after publisher complaints,” PPC Land, October 23, 2026. Read article ··· A clause-by-clause comparison, as of the evening the delay was announced, of what changed in the terms and what didn’t.
• PPC Land, “Meta makes advertisers solely responsible for its AI ad edits from Oct 30,” October 22, 2026. Read article ··· Lays out the REI, Xburrus, and Bookshop cases, along with the trouble advertisers had trying to turn off Advantage+ features.
• Korea Law Information Center, Act on the Regulation of Terms and Conditions, Articles 6, 7, and 15; Framework Act on the Development of Artificial Intelligence and Establishment of Trust, Articles 2 and 31. View statutes ··· The original text of the provisions cited today. Article 7, Item 2 is a single line — worth reading for yourself.

Background

• Matt G. Southern, “Meta Drops Training From Its Crawl Clause And Keeps The Crawl,” Search Engine Journal, October 19, 2026. Read article ··· Puts the before-and-after wording of the crawling clause side by side, along with legal experts’ interpretations.
• Newstomato, “‘Fix Google’s Unfair Terms’ — KFTC Issues World’s First Corrective Recommendation,” March 14, 2019. Read article ··· A rundown of the clauses South Korea’s Fair Trade Commission flagged in Google’s, Facebook’s, Naver’s, and Kakao’s terms back in 2019.
• Meta, Commercial Terms preview. View terms ··· The effective date shown and the wording itself may differ depending on your region.

Related issues worth reading

• Issue 81, “Becoming the AI 250 Million Small Business Owners Actually Need” ··· Covered Meta’s strategy to automate small-business advertising with AI.
• Issue 215, “The URL Berners-Lee Chose Died Within Four Months” ··· Looked at why robots.txt remains a gentlemen’s agreement rather than law.

---

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Meta Commercial Terms: The terms that apply when Meta products like Facebook, Instagram, and WhatsApp are used for business purposes — advertising, selling, or running a Page. They layer on top of the general terms of service that apply to individual users. ↩

2. Consent by continued use: A mechanism whereby, once a change to the terms has been announced, continuing to use the service is treated as agreement to the new terms. No separate “I agree” click is required — use itself constitutes consent. ↩

3. Advantage+: Meta’s suite of ad-automation tools. AI handles targeting, budget allocation, and creative optimization on the advertiser’s behalf. Its creative-enhancement feature in particular lets the AI directly edit the images and copy an advertiser has uploaded. ↩

4. Test credentials: Account information created for testing purposes. The controversy here is that this clause can be read as giving Meta grounds to create such accounts to access pages that would otherwise require a login. ↩

5. robots.txt: A text file placed at the root of a website specifying which crawlers may access which parts of the site. It carries no legal force and relies entirely on crawlers’ voluntary compliance. ↩

6. Act on the Regulation of Terms and Conditions: A Korean law preventing unfair clauses in the standardized terms businesses draft in advance to contract with multiple customers. It applies not only to consumer contracts but also to agreements with business clients, and is enforced by the Korea Fair Trade Commission. ↩

7. AI business operator: The entity bound by obligations under the Framework Act on AI. The category splits into “AI development operators,” who build and provide AI, and “AI-using operators,” who use that AI to provide products or services. ↩