You Reviewed the AI You Bought. The Software You Already Own Just Started Training on Your Members.

In July, HubSpot updated its terms so that every customer using its enrichment tools was opted in, by default, to a program that pooled contact data across accounts and fed it into the company's AI.¹ The opt-out was split across three separate switches, and turning one off left the other two running. The backlash was fast enough that HubSpot withdrew the whole change within days. Its product chief wrote, in a line worth keeping, "We made a mistake."

I want to be fair to HubSpot. They listened, and they reversed. The reason to keep talking about it is that HubSpot is not the interesting part. It is the same shape as the AI notetaker that auto-joins the board call, a default-on feature nobody in the room agreed to. And it is now the default motion of the software industry, most of which does not reverse.

Look at what shipped this year. Salesforce's Spring '26 release surfaced a setting called Opt Out of Customer Data Access, and the news was not that it existed but that access had been on by default all along, meaning customer data could be used to train Salesforce's global predictive AI models until an admin went looking for the switch.² Atlassian began collecting in-product data from Jira and Confluence to train its models on August 17, and on the lower-priced tiers you cannot fully turn it off.³ None of this is new, either. LinkedIn did it last November, when it began training content-generating models on member profiles and public posts in Europe, the UK, Canada and Hong Kong, with the opt-out under Settings & Privacy, then Data privacy, then Data for Generative AI Improvement.⁴ Zoom did a version of this back in 2023, walked it back under pressure, and the pattern has only accelerated since.⁵

None of these companies are villains for wanting the data. AI features get better with more of it, every platform knows it, and the commercial pull toward pooling and cross-account learning is the same everywhere.⁷ That is exactly why this will keep happening. HubSpot blinked because the backlash was loud and immediate. The quieter changes, the ones that ship in a release note and a settings toggle nobody reads, are the ones that stick.

For a business, this is a data-governance headache. For an association, it is closer to a breach of trust, because the data in these systems is not really yours. Your members handed it over. The certifications, the giving record, the board minutes sitting in a wiki, the volunteer's home address in the CRM. When a vendor flips a default and starts training on that, your members did not agree to it and, in most cases, will never know. You are the steward, and the decision that matters most is being made in an admin panel you do not control, by a product manager you will never meet. It is the same question I keep coming back to: is your data usable only on your terms, or on theirs.

Here is where associations get talked past. Ask a vendor whether they sell your data and they will say no, truthfully, and you will feel reassured. Selling is not the question. Training is. A model that learns from your data and then serves other customers is not a sale, and the contract language that permits it almost never uses the word train. It uses improve. Law firms spent this year warning clients that a routine clause letting a vendor use your data to improve, build, or enhance its services can, in an AI product, amount to a license to train on it.⁶ The word looks harmless. It is the whole ballgame.

So here is the afternoon's work, and none of it needs IT. This is the pixel-audit move from a few weeks ago, pointed at your settings instead of your public pages. Write down the handful of tools that actually hold member data: your AMS or CRM, your email platform, your community or membership site, your event system, the wiki where staff keep everything. For each one, open the account or admin settings and look for anything mentioning AI, data use, model training, or improvement. Note whether it is on. Then open the vendor's terms or data processing agreement and search the page for four words: improve, train, model, and develop. You are not reading the whole contract. You are finding out whether it already says yes.

If a setting is on by default, or the terms authorize using your data to improve or develop the vendor's services with no carve-out for model training, you have your answer, and it is the one you were afraid of. If you would rather ask a human, send your account rep three blunt questions and read the replies literally. Does our data train models that also serve your other customers. Is that on by default. Is there one switch that turns all of it off, or several. A good vendor answers plainly. An evasive one tells you they do not sell your data, or that it is covered under improving the service, or that it is really three different settings. Every one of those is a wrong answer, and now you can tell.

The reason to run this yourself, this week, is that it does not require anyone who writes SQL. It is clicking through settings and reading a page of terms with the search box open. An executive director can do it in an afternoon. And the thing you are protecting is not an abstraction on a roadmap. It is the one asset an association cannot rebuild or repurchase: the data its members trusted it to hold.

Quick takes

If you run Salesforce, this one is concrete. The Spring '26 release turned the opt-out into a self-service toggle in Setup, searchable under Opt Out of Customer Data Access. Before that release you had to file a support case to stop it, which is a polite way of saying almost nobody did. Worth a five-minute check even if you assume someone already handled it.

Atlassian is a lesson in reading the fine print by tier. On Enterprise you can turn both metadata and in-product collection off. On Standard and Premium you get partial control at best, and on the metadata side of the lower tiers there is no opt-out at all. The lesson generalizes: the answer to "can we turn this off" is sometimes no, and that is worth knowing before you renew, not after.

The most useful shift this year is not a product. Training rights have stopped being boilerplate. The law-firm guidance this year treats them as a term you negotiate: narrow what the vendor may use your data for, bar training outright, or buy the private version that many enterprise AI providers offer, where training serves only you, though often at a premium.⁶ A flat ban is not always the right call for every tool. For member data, it usually is. You have more leverage than you think, especially at renewal. The clause you want prohibits use of your data to train or improve any model that serves anyone but you.

Worth a read

HubSpot Backed Down on Data Sharing. The Next Vendor Might Not. (Sirocco Group). The clearest short read on why this pressure is structural, and three questions to ask inside your own organization: which data-sharing features are already switched on, where each setting lives, and who owns vendor terms.

Salesforce AI Data Sharing Opt Out: What Nonprofits and Associations Need to Know (Fionta). A rare walkthrough aimed at our sector rather than at enterprise sales teams, with the exact clicks to find the setting.

What Makes AI Vendor Deals Different (Ward and Smith). Why the SaaS review checklist you have used for a decade was not built for AI vendors, including the ones that will ask to train on your data.

My guess is that within a year, defaulting AI training to on will look the way selling browsing history looks now: technically disclosed, broadly resented, and slowly being regulated out of existence. Until then, the toggle is yours to find, and the only real question is whether you find it before your members do.

---

From the Mind of Ravi Rooprai is a weekly column on association tech, data, and AI. Read the perspectives for the longer arguments behind it.

Researched with AI assistance and fact-checked against primary sources. The analysis, judgment, and writing are mine. How this column is made →