Meta Just Made Its AI's Mistakes Your Fault. Your Other AI Vendors Got There First.

What Exactly Did Meta Change?

On 19 October 2026, Meta published revised Commercial Terms and Self-Serve Ad Terms that bind every business using its products, from Pages and ad accounts to WhatsApp Business numbers and AI-powered business agents. As CIO reported, the AI section says Meta's AI features "may make mistakes and may not always produce accurate, complete, or suitable results", and that customers "are solely responsible for and assume any and all risks associated with such AI features and any resulting content." Frank Dickson of Dickson Research said the clause "turns accountability upside down": the customer carries the risk for output from a model that Meta alone builds and controls.

The rest of the package points the same way. According to a client alert published in the National Law Review, the Business AI Terms require customers to indemnify Meta, including attorneys' fees, for claims arising from AI-powered business agents or the content they generate. The content license reaches material shared "on or in connection with" Meta products, including content posted by agencies acting for the business, and it survives termination "for any reason". Responsibility also extends to partners and contractors acting on the customer's behalf.

The authors give a concrete example: a WhatsApp business agent that misstates a refund policy. The business may face consumer claims and may also owe Meta a defence. On the other side of the ledger, PPC Land reports that Meta's own liability is capped at the greater of $100 or what the business paid in the previous twelve months.

After complaints from publishers and advertisers to regulators in the EU and the UK, Meta moved the effective date from 30 October to 20 November. It said the extra time was "to explain to our partners and stakeholders what's changing and what's not." The AI liability language did not change. Continued use after notice still counts as acceptance, and there is no partial consent.

Is Meta an Outlier, or Just the Most Honest Version of the Market?

Meta drew attention because of its scale and because the wording is blunt. The underlying allocation of risk is common. In an analysis for Stanford's CodeX centre, Olga Mack used TermScout contract data to compare AI vendor agreements with traditional SaaS contracts. The gaps are large.

88% of AI vendors cap their own liability, against 81% in broader SaaS, but only 38% cap the customer's liability, against 44%. 92% claim data usage rights beyond what is needed to deliver the service, against a market average of 63%. Just 33% indemnify customers against third-party IP claims. Only 17% explicitly commit to complying with all applicable laws, against 36% of SaaS agreements, and only 17% warrant that the product will conform to its documentation, against 42%.

Put plainly, the typical AI contract limits what the vendor owes you, leaves what you owe the vendor open, takes more of your data than it needs, and promises less about whether the product does what the documentation says. That is the default position most companies have already accepted, usually through click-through terms or a self-serve plan nobody sent to legal.

The large model providers do offer more. Microsoft's Customer Copyright Commitment, Google's generative AI indemnities, OpenAI's Copyright Shield and Anthropic's commercial terms all promise to defend paying customers against certain copyright claims over outputs. But, as Addleshaw Goddard and Proskauer pointed out when they were introduced, these promises are about intellectual property, not accuracy. They usually require you to keep the vendor's guardrails and content filters on. They often exclude free tiers. And they carve out modified outputs, combinations with other technology and uses that breach the vendor's policies, which describes most real enterprise integrations.

Why Can't You Just Blame the Model When It Gets Something Wrong?

Because courts and tribunals have already said no. In Moffatt v. Air Canada, decided in February 2024, the airline argued that its website chatbot was a separate entity responsible for its own statements. The British Columbia Civil Resolution Tribunal called this "a remarkable submission" and held that the chatbot "is still just a part of Air Canada's website." The airline had to honour the bereavement refund its bot had invented.

The sums in that case were small. The principle is not. The business that puts an AI system in front of customers answers for what that system says, regardless of who built the model. Meta's terms simply write that principle into the contract and add an indemnity on top, so the business also covers Meta's legal costs.

Liability can also travel in the other direction, toward the vendor, which is one reason vendors are tightening their paper. In Mobley v. Workday, the federal court in Northern California allowed discrimination claims to proceed on the theory that an AI hiring vendor could be acting as an agent of the employers using it. Vendors watching that case have every incentive to push risk back to customers by contract. Customers who accept that without changes end up exposed to claims from users and to indemnity claims from the vendor.

Regulation adds a third layer. Under Article 26 of the EU AI Act, deployers of high-risk systems must assign human oversight to people with the competence and authority to exercise it, monitor operation, report risks and serious incidents, and keep automatically generated logs for at least six months. Those obligations sit with the deployer whatever the vendor's terms say. European companies also face the revised Product Liability Directive, which treats software as a product, as we covered in our article on strict liability for software.

Where Is This Risk Hiding Inside Your Organisation?

The hardest part of the Meta change is not the wording. It is that most companies do not know how much of their business it touches. Aman Mahapatra of Tribeca Softech told CIO that many CIOs do not know how many Meta Pages, ad accounts and WhatsApp numbers carry their brand, because marketing teams and agencies manage them outside IT and legal oversight.

The same pattern holds for AI features generally. AI now arrives through four doors: model APIs your engineers call directly, AI features switched on inside SaaS products you already pay for, AI agents that vendors run on your behalf, and AI tools your contractors and agencies use while working for you. Only the first usually goes through procurement. The other three often arrive with a product update, a new self-serve plan or a vendor's own internal tooling.

Each door has its own terms, and each can bring an indemnity you did not negotiate. A support platform's AI reply drafting, a CRM's automated email agent and an ad platform's generative creative tools all produce customer-facing content under your name. If the terms say you assume the risk, that is now your exposure, and your insurer may not cover it either, as we described in our piece on AI insurance exclusions.

This is why the first step is an inventory, not a legal memo. You cannot negotiate, mitigate or switch off what you cannot list.

What Should You Actually Negotiate?

Leverage varies. A mid-sized advertiser will not rewrite Meta's terms, but most enterprise AI and SaaS agreements are negotiable at renewal, especially when an AI feature is a reason for the purchase. Five points matter most.

First, make the vendor responsible for what the vendor controls. A reasonable position is that the customer owns its prompts, configuration and the decision to publish, while the vendor stands behind defects in the model, the product and any automatic changes it makes without instruction. Meta's ad tools are a useful test case: PPC Land notes that advertisers pulled ads after Meta's AI made edits they had not asked for, and the terms still put the risk on the advertiser.

Second, separate the liability caps. A general cap of twelve months' fees may be acceptable for downtime, but not for data misuse, confidentiality breaches or IP claims. Ask for higher or uncapped liability on those, and resist uncapped customer indemnities. If the vendor caps its own liability, yours should be capped too.

Third, limit data rights to delivering the service. Given that 92% of AI contracts claim more, this is the most common gap. Exclude training and "product development" on your data unless you opt in, require deletion on termination, and refuse licenses that survive termination indefinitely. Fourth, require notice before material changes to AI features or terms, with the right to switch features off or leave without penalty. "Continued use is acceptance" should not apply to new AI functionality. Fifth, read the conditions on any IP indemnity and confirm your real integration does not fall into the carve-outs. If you modify outputs, combine them with other tools or disable filters, find out whether you are still covered.

Which Engineering Controls Make Unfavourable Terms Survivable?

Some terms will not move. For those, the protection has to come from how the AI is integrated. The National Law Review alert recommends human review, logging, an off switch and an escalation path for AI output that touches pricing, refunds and policy statements. That is good legal advice, and it is also an engineering backlog.

Start with a policy layer between the model and the customer. Anything that makes a commitment, such as prices, refunds, eligibility, delivery dates or legal terms, should be checked against a source of truth before it is shown or sent. The Air Canada bot would have been harmless if its refund answer had come from the actual policy rather than from text it generated.

Next, log everything you may need to defend. Store the input, the model and version, the retrieved context, the output and whether a human approved it. The EU AI Act already expects deployers of high-risk systems to keep logs for at least six months. In practice, log retention should match the limitation period for the claims you are most likely to face.

Then make every customer-facing AI feature switchable per tenant, per channel and per feature, and test the switch. When a vendor ships an unwanted AI change or a new term you cannot accept, you need to be able to turn that capability off in minutes without a release. Finally, keep the integration portable. A thin abstraction over model and vendor APIs turns "we cannot accept these terms" from a threat into an option. We covered the architecture in our article on model portability.

None of this is exotic, and none of it requires slowing down AI adoption. It requires treating AI output as untrusted input to your business processes, the same way good engineering teams already treat data from any third-party API.

What About AI Used by Your Development Partner?

Meta's terms make customers responsible for agencies and contractors acting on their behalf. That clause applies beyond Meta. Your development partner's AI tools, coding agents and model subscriptions come with their own terms, and some of those terms may grant the tool vendor rights over the code, prompts and data your partner handles for you.

Ask your partner three questions. Which AI tools touch your code and data, under which plan and which terms? Do those terms allow training on your inputs, and has that been switched off? And who holds the account: your organisation, the partner, or individual developers on personal plans? A coding assistant on a consumer plan, used by a contractor on your codebase, can carry exactly the data rights and liability allocation your legal team would have refused in a direct contract.

Coding agents add a quality dimension too. Trail of Bits reviewed more than 130 of its own upstream patches and confirmed 21 regressions, which is why it now runs dedicated regression checks on agent-assisted fixes. If the vendor terms say you own all risk from AI output, your partner's review process is your real protection, so make it contractual: disclosed tools, enterprise accounts held or approved by you, training opt-outs confirmed, and recorded human review of AI-assisted changes.

How Does a Nearshore Dedicated Team Help?

Most of the work in this article is not legal drafting. It is inventory, integration and controls: finding every place AI touches your customers, putting a policy layer and logging in front of it, building tested kill switches and keeping vendor integrations portable. That work competes with your roadmap for the same engineers, which is why it tends to wait until a term change or an incident forces it.

Stepto builds dedicated development teams in Serbia that work inside your repositories, your review rules and your vendor policies. The same engineers stay on your account sprint after sprint, so AI tool usage is disclosed, enterprise accounts are under your control, and AI-assisted changes are reviewed by people who know your system. That is the documented evidence a lawyer, auditor or insurer will ask for when a vendor's terms leave the risk with you.

Being in Belgrade helps in practice. Stepto works on CET, the same business hours as Berlin and Paris and one hour ahead of London, with a useful overlap with the US East Coast. Engagements typically start within 2-4 weeks of the first call, with a two-week paid trial, at rates 36-51% below Western European and US equivalents.

If your legal team is reviewing AI vendor terms this quarter and needs to know what can be controlled in code, talk to us. We can help you map your AI surface and build the controls that keep unfavourable terms from becoming uncovered losses.