Connecticut Wants Written Acceptance Before Your AI Subscription Bills

Since 1 October 2026, a covered AI subscription in Connecticut cannot charge a card until the consumer has accepted its key terms in writing. After a companion act rewrote the rule in June, it reaches generative AI systems with more than 1 million monthly users. Connecticut's AI subscription law is really a checkout rule, and several widely read summaries still describe the version that was replaced.

The short answer

Under Public Act 26-15 as amended by Public Act 26-100, a covered generative AI provider may not enter, renew or collect payment for a Connecticut consumer subscription until two things happen. It must send written notice of key terms, including usage limits and any right to reduce functionality. The consumer must then accept in writing. It applies from 1 October 2026.

What Connecticut's AI subscription law requires at checkout

The rule started life in Substitute Senate Bill 5, which Governor Ned Lamont signed on 27 May 2026 as Public Act 26-15. Faegre Drinker places the provision in Section 1. Davis Polk calls it the first state law aimed specifically at subscription-based AI offerings.

A provider may not enter into a subscription, renew one, or collect a fee until it has given written notice of the key terms. The consumer must also have accepted those terms in writing. Faegre Drinker puts it bluntly: "Consumers must provide written acceptance before the provider may collect fees". The sequence is fixed. Notice, then acceptance, then charge.

That ordering makes this a checkout problem. What many AI products lack is a moment before the first charge where the limits are shown and the buyer's agreement is recorded. Teams that file this under "update the terms" will miss it, because the statute regulates the order of events.

Enforcement sits with the Attorney General alone. A violation is an unfair or deceptive trade practice under the Connecticut Unfair Trade Practices Act. Covington reports penalties of up to $5,000 per willful violation. Faegre Drinker confirms there is no private right of action for this section.

PA 26-15 was rewritten within a week: who the AI subscription rule now reaches

As signed in May, the provision reached anyone doing business in Connecticut that offered "AI technology" to a consumer on subscription. Davis Wright Tremaine read that as any product using or incorporating AI.

Then came House Bill 5222. Covington reports that Governor Lamont signed it on 2 June 2026 as a companion bill that "modified some of the provisions of SB 5, including the provision on AI subscriptions". Clearon AI places the revised rule in Section 46 of Public Act 26-100, effective 1 October 2026.

Connecticut AI subscription rule: the May 2026 text against the June 2026 amendment

Element	PA 26-15 as signed, 27 May	As amended by PA 26-100, 2 June
Product covered	Any AI technology offered on subscription	A generative AI system offered on subscription
Provider threshold	None reported	More than 1 million users per month
Access condition	None reported	Publicly accessible to consumers for personal use
Who is protected	Connecticut consumers	Connecticut residents acting outside a commercial or employment role
Notice and written acceptance before entry, renewal or payment	Yes	Yes
Disclosure of limits and discretion to degrade	Yes	Yes, with tokens, images and transcription named as examples
Where the narrowing concedes ground	Broad text is what most firm alerts still describe	A smaller provider that sells AI to consumers may fall outside it entirely

The four scope tests for a self-serve AI product

On the amended text, a product is in scope only if it passes four tests. It is a generative AI system, which Clearon paraphrases as machine learning that generates content such as images, audio or video. The provider does business in Connecticut. The system has more than 1 million users per month. It is publicly accessible for personal use and sold on subscription.

Covington adds that the definition of subscription "does not appear to be limited to autorenewals". A one-year prepaid plan counts too.

What falls outside the amended rule

Clearon reads the protected consumer through Connecticut's data privacy act, which excludes a person acting in a commercial or employment context. Clearon also says the section does not reach enterprise deployments, internal models, one-off API transactions or AI features inside other software.

That is a vendor's reading. A prosumer on a company card looks like a personal buyer to your checkout, so treat every Connecticut buyer of a personal plan as covered.

What the AI subscription disclosure must say before payment

The notice is not a link to your terms. Covington reports it must include "material information that is sufficient to enable a reasonable consumer to decide whether to purchase or maintain the subscription". It must cover two items at minimum.

The first is any quantitative or qualitative limitation, including, in Sidley's reading, "any restrictions triggered by the consumer's conduct". Covington and Clearon report that the amended text names tokens, generated or modified images, and transcription as examples.

The second is whether the provider has discretion to limit access to, or reduce the quantity or quality of, any functionality. If you can route a paying user to a cheaper model at peak load, say so before you take the money.

Why the limits clause targets AI pricing specifically

AI caps reset on rolling windows and plan pages describe capacity as multiples of another tier. Our analysis of quiet cuts to AI usage limits shows how often those caps change without a price change. When credits convert to tokens at a rate the provider sets, the real limit is two numbers deep, as our piece on credit and token billing confusion shows.

The bigger risk is a vague notice, not a missing one. "Usage limits may apply" is not material information. State the cap, the window, what happens on breach and whether the model can be swapped.

Written acceptance in a one-click checkout flow

The statute asks for acceptance "in writing", and Davis Wright Tremaine describes it as the consumer giving written notice of acceptance. No alert we read says what counts as writing here.

What a court is likely to treat as written

Under the federal E-SIGN Act and the Uniform Electronic Transactions Act, which Connecticut has adopted, an electronic record can generally satisfy a writing requirement. An affirmative checkbox, stored with the exact notice text shown, is the most plausible way to meet the rule.

One catch for counsel: E-SIGN sets an extra consent procedure at 15 U.S.C. 7001(c). Ask a lawyer whether it applies before building a separate consent screen.

Checkout patterns that fail the written acceptance test

The patterns most likely to fail are the ones built to remove friction. Browse-wrap terms in a footer show nothing before payment. A pre-ticked box records a default, not a decision. A one-tap upgrade from inside the product charges a saved card before any notice appears.

AI subscription checkout patterns: the weak version against the version that works

Moment	Weak version	Version that works
New signup	"By subscribing you agree to our Terms" under the pay button	Key terms panel listing caps, windows and model-swap rights, plus an unticked checkbox that gates the pay button
In-app upgrade with a saved card	One tap, charge, then a toast with a terms link	Interstitial with the new plan's limits and a confirm action before the charge call
Free trial converting to paid	Terms accepted at trial start, card charged silently at day 7	Terms and limits accepted at trial start, recorded against the plan that will bill
Mobile app store purchase	Store sheet only, no provider notice	Provider notice and acceptance in the app before the store sheet opens
Evidence kept	A boolean in the users table	Timestamp, user, plan, notice version hash and the rendered text
Where the weak version wins	Higher conversion from fewer steps	One added step, which you can test against your own drop-off data

An extra checkbox costs some conversion, and I have no measured figure for how much. What it buys is a record that survives an Attorney General inquiry.

AI subscription renewals, plan changes and mid-term limit cuts

Davis Polk and Sidley both report that a renewal notice must disclose limits that are new or modified from the preceding term. Davis Wright Tremaine adds discretion the provider will exercise for the first time, or has modified. Clearon reads every renewal as needing the same notice and acceptance.

The monthly renewal question nobody has answered

Read literally, each monthly renewal needs written acceptance before the fee is collected. Read purposively, the renewal duty exists to surface changes, so an unchanged plan needs nothing new. The reported text leans literal, because the acceptance duty is not limited to changed terms.

I would not ask every monthly subscriber to re-click every month. Capture acceptance at signup that expressly covers renewal on the same terms. Then force fresh acceptance before the first charge after any change to limits or discretion, and before each annual charge.

That makes a mid-term cap cut on a covered plan a product release with a checkout step. Our pieces on grandfathered pricing at renewal and the ChatGPT plan pricing timeline track exactly that kind of change.

The SaaS checkout compliance checklist for PA 26-15

Each row below maps a reported element to a build and the evidence to keep. The evidence column matters most, because an inquiry asks what the consumer saw.

Connecticut AI subscription checkout checklist: element, build and evidence

Statutory element, as reported	What to build	Evidence to keep
Scope: generative AI, over 1 million monthly users, personal use	A written scope memo with your user count method and the plans it covers	Monthly user counts and the memo, dated
Connecticut consumer	Flag on billing address or declared state at checkout	The location signal used per transaction
Material information for a reasonable consumer	Key terms panel shown before payment, in plain language	Versioned notice text with a content hash
Quantitative and qualitative limits, including conduct-based ones	Caps in units, reset windows, fair-use triggers, what happens on breach	Same versioned notice
Discretion to limit access or reduce quality	A plain statement of model swaps, throttling or feature withdrawal rights	Same versioned notice
Written acceptance before entry or payment	Unticked checkbox or confirm action that gates the charge call in code	Timestamp, user, plan, notice hash
Renewal notice of new or modified limits	A change-detection job that diffs plan limits between terms	The diff, the notice sent and the acceptance before the next charge
No acceptance, no charge	A billing guard that blocks the payment call without a valid record	Blocked-charge logs

The billing guard makes the rule a property of the system, so a future upgrade flow cannot bypass it. The change-detection job matters because limits often live in config files that other teams edit without telling the checkout owner. Our rate limit transparency standard lists fields that make a reasonable first draft of the key terms panel.

How the AI subscription rule sits beside Connecticut's auto-renewal law

The AI rule does not replace Connecticut's automatic renewal law. That law was amended in 2023, and further amendments took effect on 1 July 2026. Attorney General William Tong announced on 16 July 2026 that the state would enforce them.

The new duties include an annual reminder, wider cancellation options, and acting on a voicemail cancellation within one business day.

The auto-renewal law governs the reminder and the exit. The AI rule governs entry and each renewal charge. The cleanest build treats both as one subscription-state system with one evidence log.

Clearon also reports a separate Section 47, effective 1 October 2027, creating a Department of Consumer Protection independent-verification pilot. Our analysis of Connecticut's deployer obligations covers the rest of Public Act 26-15. California's rules are in our breakdown for software vendors.

Where this reading is weakest

This translates reported statutory text into product requirements, and that text has gaps.

Two careful readings describe two different laws

The narrowing rests on Covington's July 2026 post and Clearon's analysis of Section 46 of Public Act 26-100. The legislative research summary heading points the same way. Yet Davis Polk's 3 September 2026 update and Sidley's August post describe the broad May text without mentioning the amendment.

Neither firm says the amendment failed, so I read their silence as scope, not dissent. But I could not open either enacted act, because the Connecticut General Assembly's site refused every connection. A smaller provider relying on the threshold is relying on two secondary sources, one of them a vendor.

So build anyway. The notice, gate and log cost a few sprints. Being wrong about scope costs an enforcement inquiry with no record to show.

The rule may matter less in practice than on paper

There is a fair counter-case. Few providers clear 1 million monthly users, and those that do already show plan limits somewhere. No enforcement action under this section has been reported in the three weeks since it took effect. The penalty is modest next to such a provider's revenue.

Nobody has defined how monthly users are counted, what "in writing" requires, or whether each monthly renewal needs fresh acceptance. I still think the pre-payment gate is worth building, because every reading of every version agrees on it.

Frequently asked questions

Where to start this week

Run the scope test first. Pull monthly user counts by product, list which plans are sold for personal use, and write down how you counted. If you are clearly below the threshold, record the reasoning and set a review date.

Then find every path that authorises a payment, including upgrades, trial conversions and renewals. Each needs a guard that refuses to charge a Connecticut consumer without an acceptance record tied to the current notice version.

Related reading