The Chatbot Promise Checklist

01 Why your bot's description gets read

On October 14, 2026, eight states (Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia) each sued OpenAI over ChatGPT, after Florida sued in June. None of them used an AI law. They used consumer-protection law: the kind that asks whether what a business said about its product is true. These are allegations; no court has ruled, and OpenAI disputes them.

The idea isn't new. In 2024, Air Canada's website chatbot told a grieving passenger he could claim a bereavement fare after his flight. He couldn't. The airline argued it wasn't responsible for what the chatbot said. The tribunal disagreed: the chatbot "is still just a part of Air Canada's website." About $650, and a principle.

Your bot's words are your words.

Nobody has tested these laws on a support bot yet, and this isn't legal advice. The checks below follow the four words a typical chat widget uses: "Hi, I'm your safe, accurate assistant, here for everyone, always."

02 Check 1: "safe" and "accurate"

Find every sentence about your bot. They live in more places than the widget:

Your website. The help page, the product page, the widget's greeting.

Your app store listing. Including its age rating: the states quote ChatGPT's 13+ rating.

Your ads. The states quote a LinkedIn ad that told parents ChatGPT "helps keep teens safe by default."

Your privacy policy. Under Nebraska's law, a false statement there is a deceptive trade practice on its own.

What the bot itself promises customers. Refunds, discounts, delivery dates. That's where Air Canada got caught.

My promise list

Sentence: [THE EXACT WORDS]
Where it lives: [PAGE / LISTING / AD / POLICY / BOT REPLY]
Proof it's true: [THE TRANSCRIPT OR TEST THAT SHOWS IT]
If no proof, rewrite as: [WHAT THE BOT ACTUALLY DOES]

03 Check 2: "for everyone"

If your bot isn't meant for kids, say so, and then do something about it. The worst place to be is a policy that says you don't knowingly talk to children while your own logs say you do.

Say who it's for on the page where the bot lives, not only in the policy.

Ask, and route. Someone who says they're fourteen goes away from open chat.

Read your logs against your policy. If the logs show minors, the policy sentence is the one to fix first.

A protection is only as good as the gate in front of it. The complaint says OpenAI's teen rules switch on when a user says they're under 18, or when its age prediction believes they are.

04 Check 3: "always"

Decide in advance when the bot stops. This is the one to get right.

Handoff rule

The bot says it's a bot and hands off to [PERSON OR TEAM, WHO ACTUALLY PICKS UP],
with the conversation attached, when someone:
- asks for a person
- repeats themselves
- brings up money, health or legal trouble

Crisis rule

If someone sounds like they're in crisis, the bot does not carry on as usual
and does not just drop them. Right away it gives [YOUR LOCAL CRISIS LINE]
and brings in [A REAL PERSON].

Test it in long conversations, not just the first message. California's companion chatbot law leaves out bots used only for customer service. Do it anyway.

05 Check 4: the record

What gets read is dates: a dated policy, your test transcripts, a change log, a record of every handoff. Your record should show the safeguard came first, not the week after a complaint. Once a month, have someone on your team play three people:

A fourteen-year-old. An angry customer. Someone who's clearly not okay.

Save those transcripts. When something does go wrong, write down what happened and what you changed, with the date.

Monthly log

Date: [DATE]
Tested by: [NAME]
Fourteen-year-old: [ROUTED AWAY? Y/N]
Angry customer: [HANDED TO A PERSON? Y/N]
Not okay: [CRISIS LINE + PERSON, RIGHT AWAY? Y/N]
What we changed: [ONE CHANGE, WITH THE DATE]

Start here. Which word is coming off your chatbot's description first: safe, accurate, or always?

Sources

State of Nebraska v. OpenAI: the complaint (October 14, 2026) Nebraska Attorney General: lawsuit announcement Florida Attorney General: the first state lawsuit (June 2026) Times Colonist: Air Canada responsible for errors by website chatbot Future of Privacy Forum: California SB 243 and chatbot legislation

Mathew's LibraryLibrary card

Free access

Keep reading free with your library card

One email opens this checklist and all 3 resources in Mathew's library. No account, no password.