When the wording is the liability.

Tools that write for you raise the stakes, because they write quickly and in many places at once: the website, the sales follow-up, the support chat, the social post.

Each can phrase the same claim slightly differently. In some industries, a slightly different phrasing is a different claim.

1. The rules still apply when a machine writes it

Regulators have been direct about this. In June 2024 FINRA reminded its member firms that it intends its rules to be technology neutral, and that its standards for communications with the public apply whether a person or a technology tool wrote them.

The FTC has taken the same view in enforcement. In September 2024 it announced Operation AI Comply, five actions over deceptive claims involving AI.

One of them was DoNotPay, which had marketed itself as “the world’s first robot lawyer” and, according to the FTC, never tested whether its output matched a human lawyer’s. The final order, announced in February 2025, required it to pay $193,000, to tell subscribers from 2021 to 2023 about the settlement, and to stop claiming it performs like a lawyer without evidence to back that up.

In March 2024 the SEC settled what it called “AI washing” charges against two investment advisers, Delphia and Global Predictions, among its first cases of the kind. They paid $400,000 in combined penalties for statements about their use of AI.

Global Predictions had called itself the first regulated AI financial adviser and could not produce documents to support the claim. Both firms were charged under the Marketing Rule and under the rule that requires written policies to prevent violations, which in Delphia’s case included having none to keep its advertising accurate.

2. Inconsistency can be the problem in itself

The clearest example is a small claim. In February 2024 a British Columbia tribunal ruled against Air Canada after its website chatbot told a grieving customer he could apply for a bereavement fare after travelling, when the airline’s own bereavement page said the opposite.

Air Canada argued, in the tribunal’s words, that the chatbot was “a separate legal entity that is responsible for its own actions”. The tribunal called that “a remarkable submission” and held the airline responsible for all the information on its website.

The award came to CA$812.02. The amount is small, and the principle is the part worth noticing: when two of your own sources disagree, you may be held to the one the customer relied on.

3. The money is the small part

None of these amounts would sink a large company. The orders attached to them are what last: limits on what the company may say from then on, notices to past customers, and marketing under closer watch.

Correcting the wording afterwards almost always costs more than getting it consistent beforehand.

4. Where the risk sits, industry by industry

The sensitive wording differs by sector, and the pattern does not: a specific claim, stated with conditions, that a tool can easily loosen.

• Financial services. Rates, fees, performance, eligibility, and any suggestion that returns are guaranteed or that a product uses capabilities it does not have.
• Health and health tech. What a product is approved or cleared for, who it is for, and any outcome it implies. A broader phrasing of an approved claim can be as serious as an invented one.
• Insurance. What a policy covers, the exclusions and conditions attached, and how claims are handled.
• Legal and professional services. Credentials, what a service can replace, and anything that reads as a promised result.
• Security and compliance vendors. Certifications, their scope and their dates, which assistants tend to flatten into a simple yes.

5. What consistent wording takes

Every case above is a claim that went out without the evidence or the control around it. Consistent wording comes from putting that control in one place that every tool reads.

• Approved wording, written once in its exact form, with the conditions and disclosures that have to travel with it.
• Evidence for each claim, since several of these cases turned on a company being unable to back up what it said.
• A named approver for each claim, so it is clear who agreed to it and when.
• Retired wording marked as retired, with a date, so an old version does not resurface in a new draft.

The tools that read the record also need rules for behaviour. A customer-facing agent answers policy questions only from the approved text, refuses claims that are not on the list, and hands anything outside it to a named person.

Because the record is versioned, you can show what the approved wording was on any given date. That is the question a regulator or a court is likely to ask.

6. What this does not replace

We are not lawyers, and none of this is legal advice. A record does not replace your compliance team or any review your regulator requires.

What it changes is where their effort goes. Approved claims are checked once and reused as written, so review can concentrate on what is new.

Start with the audit. It is free, it is one pass, and you keep what it finds whether or not anything follows it.