Are SaaS Terms of Service Changing for AI? What Vendors Are Doing

Are SaaS terms of service changing for AI?

Yes. SaaS terms of service are changing for AI because vendors need a legally defensible way to train models on customer data, and the old "we may use data to improve our services" language was never written to cover that. Vendors are rewriting decades-old boilerplate into explicit AI-training clauses, pushed by competitive pressure to build proprietary models and by regulators warning that quiet, retroactive changes to data terms can be illegal.

The result is a two-part price for renting software. You pay the subscription fee, and you pay in data. Stanford Law's CodeX center, drawing on contract data from TermScout, found that 92% of AI vendor contracts claim data-usage rights beyond what is strictly necessary to deliver the service, compared with a 63% market average across SaaS contracts generally. That gap is the hidden tax. It never shows up on an invoice, and by 2027 it's close to standard practice, one more entry in the true cost of what rented software actually buys you.

How did "improve the service" quietly become "train our model"?

Most SaaS contracts have carried some version of a data-use clause for years: language letting the vendor use customer data, usually aggregated or anonymized, to fix bugs, tune performance, or benchmark uptime. That language was written for deterministic software that didn't learn from what you fed it.

Generative AI changed what "improve" can mean without changing the words on the page. A clause that once authorized routine service maintenance can now be read to authorize feeding your files, code, or designs into a foundation model. Legal commentary describes this as a structural pattern: vendors integrate AI into an existing platform through a routine product update and a revised terms page, without ever renegotiating the underlying contract with the customer.

The New York Times documented this shift happening in real time. Google made an eight-word addition to its privacy policy in 2023 permitting public information to train its AI chatbot, and Meta told European users their public Facebook and Instagram posts would train its large language model. The pattern has since moved from consumer platforms into core business software:

• Zoom (2023) drew backlash after a terms update appeared to grant "perpetual, worldwide" rights to use customer data for machine learning; Bellingcat's research director called it worrying enough to cancel the subscription outright, and Zoom walked the language back.
• Adobe (2024) triggered a similar reaction over a clause suggesting the company could "access" customer content, and issued a clarification after the backlash went public.
• GitHub Copilot (2026) began using interaction data from Free, Pro, and Pro+ users to train AI models by default, including code from private repositories while Copilot was active in them, with an opt-out deadline rather than an opt-in choice.
• Vercel (2026) opted its Hobby and Trial Pro tiers into AI training by default while paid Pro accounts were opted out and Enterprise was never opted in at all, effectively pricing training rights by subscription tier.
• Salesforce (2026) enabled use of customer data for "global predictive AI models" by default, a setting many customers discovered only by digging into a Setup toggle.

Each case follows the same script: a term written for one era of software gets stretched to cover another, and customers find out after the fact.

How widespread is this, really?

This isn't a handful of vendors making the same mistake. It's a market-wide realignment. Common Paper's 2026 SaaS Contract Benchmark Report, built from 16,140 signed cloud service agreements across 2,223 companies, found that language mentioning AI now appears in 53% of cloud service agreements, up from 38% in 2024, and that 20% of contracts now carry a negotiated AI-specific edit, up from under 4% two years earlier.

Customers are pushing back, but slowly and unevenly. Explicit prohibitions on AI training rose from under 1% of agreements in 2024 to 14% in 2026, while language explicitly granting vendors training rights has stayed flat at roughly 3%. Read together, those numbers mean most contracts are neither clearly protective nor clearly permissive. They're silent, and silence tends to favor whoever wrote the contract.

The Stanford/TermScout data adds detail to why that silence is dangerous. Only 17% of AI contracts explicitly commit to complying with all applicable laws, versus 36% for standard SaaS agreements, and only 33% of AI vendors offer indemnification against third-party IP claims, against a 58% market average. AI vendors are, on average, asking for more and promising less.

Why are SaaS vendors rewriting terms right now?

Three forces are pushing this rewrite onto this exact timeline.

1. A regulatory warning shot. In February 2024, the FTC stated plainly that a company which quietly changes its terms to permit AI training on previously collected data, without prominent notice, risks an unfair-or-deceptive practices finding under Section 5 of the FTC Act. Vendors responded not by abandoning the practice but by making it explicit and contractual, which is legally safer even if it isn't friendlier to the customer.
2. The EU Data Act's switching deadlines. The regulation has required cloud and SaaS providers to remove contractual and technical switching barriers since September 2025, and from January 12, 2027, it bans all vendor-imposed switching charges outright. Vendors that want to lock in data rights are racing to do it before customers gain a cheaper way to leave.
3. Competitive pressure to build models. Every SaaS company with an AI roadmap needs training data, and its own customer base is the cheapest, most relevant source available. Rewriting a compliance clause is far faster than acquiring a comparable dataset elsewhere.

What does renting SaaS actually cost you now?

The clearest evidence that data rights function as payment is what happens when customers realize the arrangement after the fact. Figma was hit with a proposed class action in November 2025 alleging it automatically opted customers into having their designs and intellectual property used to train its generative AI tools, despite prior promises that it would not do so.

Both cases share a structure worth naming plainly. The customer paid a subscription fee believing that was the whole price. The vendor also collected a second, unpriced asset: the customer's proprietary content, used to build a product the vendor will eventually sell back to the market, possibly including that customer's competitors. That second payment never appears on an invoice, which is exactly what makes it a tax rather than a fee. The true cost of renting your software stack has always included more than the line item on the bill, and AI training rights are the newest entry on that ledger.

Own vs. rent: who controls the terms?

This is where ownership changes the calculus. A vendor can amend its terms of service unilaterally, often with nothing more than a notification email and a 30-day window before the new terms take effect. A customer running self-hosted or internally built software has no such counterparty. There's no third party with standing to decide, six months from now, that your workflows are a training set.

That asymmetry is the core argument for treating software as an asset you own rather than rent. It doesn't mean every tool should be built in-house. It means the calculation for which tools to rent should now include a line for data exposure, not just seats and uptime.

How do you audit your existing SaaS contracts?

If you're already several SaaS vendors deep, the fix isn't switching everything at once. It's auditing what you signed.

1. Find the "improve the service" clause. Search every active SaaS contract for language about service improvement, aggregated data, or product development, and read it against what the vendor's AI roadmap actually is.
2. Check the default, not the policy. Legal guidance for 2026 procurement treats training-off-by-default and opt-in consent, rather than opt-out, as the baseline standard, not a bonus concession.
3. Reject de-identification carve-outs. A clause that only restricts use of "identifiable" data usually still permits training on de-identified content, which is not the same as no use at all.
4. Ask for sub-processor disclosure. Know which downstream vendors and model providers touch your data before you sign, not after an incident.
5. Benchmark against opt-in leaders. Anthropic's commercial terms do not use customer inputs or outputs for training by default, and Google Cloud states it will not train or fine-tune models on customer data without authorization. Use those as the standard you negotiate toward, not the exception.
6. Watch the EU Data Act calendar. Switching charges shrink now and disappear entirely on January 12, 2027, which changes your leverage in any renewal conversation before that date.

None of this eliminates the exposure entirely. A 2026 Gartner survey found 57% of employees already use personal GenAI accounts for work, and a third admit uploading sensitive company data to tools their employer never approved. Contract language can only govern what a company officially signs up for. It can't govern what an employee pastes into a chat window on a Tuesday afternoon. That's a separate problem, but it's the same underlying lesson: data has become the currency, whether or not anyone priced it into the invoice.