# The Hidden Compliance Risks of AI Chatbots on Business Websites in 2026

Source: https://complymynt.com/blog/ai-chatbot-compliance-risks-2026
Published: 2026-10-20 · Topic: AI compliance · 21 min read
Author: ComplyMynt (https://complymynt.com)
Note: ComplyMynt is not a law firm and this article is not legal advice.

> AI chatbot compliance in 2026: 10 hidden risks covering bot disclosure, chat recording, privacy, wrong answers, prompt injection, accessibility, and children's data, with fixes.

## Key takeaways
- A website chatbot is not just a support widget. It collects personal data, makes statements on your behalf, and often sends conversations to third-party AI vendors.
- The biggest hidden risks are undisclosed bots, chat recording without notice, answers your business is held to, over-collection, prompt injection, and inaccessible chat windows.
- Requirements depend on where your users are and what the bot does. California, Utah, the EU AI Act, GDPR, the FTC Act, COPPA, and accessibility law can all apply to the same widget.
- Fix the cause: clear disclosure, a data map of every vendor, scoped answers with escalation to a human, logging rules, security testing, and verification after every model or prompt change.

AI chatbots are now on a large share of SaaS, ecommerce, and service websites. They answer pricing questions, qualify leads, book demos, and handle support. They are also one of the least-reviewed parts of a modern website. A chatbot is usually added by a marketing or support team through a script tag, connected to a third-party model, and changed through a prompt that never goes through legal, security, or accessibility review.

AI chatbot compliance is the work of identifying which disclosure, privacy, consumer-protection, security, accessibility, and AI rules apply to your chatbot, then implementing and testing controls that meet them. The core question is simple: what does the bot say, what does it collect, where does that data go, and who is accountable when it is wrong?

This guide covers ten hidden chatbot risks for AI and SaaS businesses serving U.S., UK, and EU users. Each section explains why the risk matters, how to check for it, and what a sound fix looks like. Numbers in brackets refer to the official sources listed at the end. Information reviewed October 20, 2026; laws and guidance change. ComplyMynt is not a law firm and this article is educational information, not legal advice.

## What counts as an AI chatbot for compliance purposes?

For this guide, an AI chatbot is any interactive text or voice interface on your website or app that generates responses using a language model or automated decision logic. That includes support assistants, sales and lead-qualification bots, onboarding guides, AI search boxes, and live-chat tools that use AI to draft or send replies.

Rule-based bots with fixed answers carry fewer risks, but disclosure, privacy, recording, and accessibility obligations can still apply. Generative bots add the risk of inaccurate, unpredictable, or manipulated output.


## Risk 1: The bot does not clearly say it is a bot

Several laws require businesses to tell people when they are talking to an automated system. California's bot disclosure law makes it unlawful to use a bot to knowingly deceive a person in California about its artificial identity to sell goods or services or influence an election, unless the bot is clearly disclosed [1]. Utah's Artificial Intelligence Policy Act requires disclosure of generative AI use in certain consumer interactions, including when a person asks [2]. Under the EU AI Act, providers must design AI systems that interact with people so users are informed they are dealing with an AI system unless that is obvious; these transparency duties apply from August 2, 2026 [3].

How to check: open the chatbot as a new visitor on desktop and mobile. Is the AI nature clear before the first message, in the widget header, and in the answer to the question 'Am I talking to a person?' Check whether a human-sounding name or photo suggests a person.

What a fix looks like: a persistent label such as 'AI assistant', a short first-message disclosure, an honest answer when asked, and a clear way to reach a human. Avoid human personas for AI agents.


## Risk 2: Chat recording and third-party vendors without notice

Many chat widgets send every keystroke and message to the chat vendor and sometimes to an AI provider and analytics tools. In the U.S., plaintiffs have brought many lawsuits under the California Invasion of Privacy Act (CIPA) section 631 and similar wiretap laws, arguing that chat vendors 'eavesdrop' on conversations without consent [4]. Outcomes vary by court and facts, but the litigation exposure is real and defense costs can be significant even when claims fail.

How to check: use browser developer tools to watch network requests while you type in the chat. List every domain receiving chat content, whether data is sent before the user presses send, and whether the vendor may use conversations for its own purposes such as model training.

What a fix looks like: a clear notice at the start of the chat explaining that conversations are recorded and processed by named categories of service providers, contracts that restrict vendor use, model-training opt-outs where available, and an updated privacy notice. Get counsel's view on consent wording for your jurisdictions.


## Risk 3: Your business is held to what the chatbot says

A chatbot speaks for your company. In Moffatt v. Air Canada (2024), the British Columbia Civil Resolution Tribunal held Air Canada responsible for a refund policy its chatbot described incorrectly, rejecting the argument that the bot was a separate entity [5]. The FTC has also warned that companies are responsible for deceptive claims, including claims made through AI tools, and has taken action against misleading AI marketing [6][7].

How to check: build a test set of real customer questions about pricing, refunds, cancellation, contract terms, security, and legal or medical topics. Run it after every prompt, model, or knowledge-base change and record the answers.

What a fix looks like: limit the bot to approved source content, block or redirect high-risk topics (legal, medical, financial, refunds outside policy), link to the authoritative policy page in answers, show a short accuracy notice, and escalate to a human for commitments. Keep a log of tested answers as evidence.


## Risk 4: Collecting more personal data than you need

Users paste surprising things into chat: account numbers, health details, passwords, and other people's data. Under the GDPR you need a lawful basis, purpose limitation, data minimisation, and transparency for chatbot processing [8]. The UK ICO's guidance on AI and data protection explains how these principles apply to AI systems, including fairness and transparency [9]. U.S. state privacy laws add notice, sensitive-data, and consumer-rights obligations.

How to check: review stored transcripts (with appropriate access controls) for sensitive data, check retention settings in the chat and AI vendors, and confirm whether transcripts flow into your CRM or analytics.

What a fix looks like: tell users not to share sensitive information, redact common sensitive patterns before storage, set short retention periods, restrict who can read transcripts, and make sure access and deletion requests cover chat data.


## Risk 5: Chat data used for training or shared beyond the stated purpose

Some chat and AI vendors reserve the right to use customer conversations to improve their models. The FTC has warned that quietly changing terms to use consumer data for AI training can be unfair or deceptive [10]. If your privacy notice says chat data is used only to answer questions, vendor training use may contradict it.

How to check: read the data-processing terms of every chat and model provider. Confirm settings for training, retention, and sub-processors, and compare them with your privacy notice and customer contracts.

What a fix looks like: disable training on customer data where possible, sign data processing agreements, list the vendors or categories in your notice, and review international transfer safeguards for EU and UK users.


## Risk 6: Prompt injection, data leakage, and insecure integrations

Generative chatbots can be manipulated. The OWASP Top 10 for LLM Applications lists prompt injection, sensitive information disclosure, and excessive agency among the top risks [11]. A bot connected to order systems, CRMs, or internal documents can be tricked into revealing data or taking actions it should not. NIST's Generative AI Profile describes similar risks and controls [12].

How to check: try to make the bot reveal its system prompt, other users' data, internal documents, or API keys. Test whether it follows instructions hidden in pages or files it reads, and review what tools and data it can access.

What a fix looks like: least-privilege access for the bot, server-side authorization checks on every action, output filtering, rate limiting, secrets kept out of prompts, and regular red-team testing documented as evidence.


## Risk 7: The chat window is not accessible

A chatbot that cannot be used with a keyboard or screen reader can block people with disabilities from support or purchase journeys. U.S. courts have applied the ADA to websites, and the Department of Justice has published guidance on web accessibility [13]. WCAG 2.2 sets the widely used technical benchmark, including focus visibility, keyboard access, status messages, and target size [14]. The European Accessibility Act also applies to many consumer-facing digital services in the EU from June 28, 2025.

How to check: open, use, and close the chat with only a keyboard. Test with a screen reader to confirm new messages are announced, the launcher has an accessible name, focus does not get trapped, and the widget does not cover key buttons on mobile.

What a fix looks like: choose a vendor that documents WCAG 2.2 AA conformance, fix custom launcher and overlay issues, provide an alternative contact method, and retest after vendor updates.


## Risk 8: Children use the chatbot

If your service is directed to children under 13, or you know a user is a child, COPPA requires verifiable parental consent before collecting personal information, and chat messages can contain it [15]. The FTC has also opened inquiries into AI chatbots acting as companions, with a focus on effects on children and teens [16]. The EU and UK have additional expectations for services likely to be used by children.

How to check: decide whether children are likely to use your site, test how the bot responds to a user who says they are young, and review what is stored.

What a fix looks like: age-appropriate design, limits on data collection, safe-response rules, and a documented decision about whether the bot is available to minors.


## Risk 9: Automated decisions made through the chat

Some bots do more than answer questions: they screen job applicants, qualify credit or insurance leads, or decide eligibility. Under GDPR Article 22, people have rights regarding decisions based solely on automated processing that significantly affect them [8]. Several U.S. states regulate automated decision-making and profiling, and the EU AI Act treats certain uses, such as employment screening, as high-risk [3].

How to check: map every decision the chatbot makes or influences, and whether a human reviews outcomes that affect access to jobs, credit, housing, insurance, or essential services.

What a fix looks like: human review for significant decisions, notices explaining the logic and the right to contest, bias testing, and documentation of the system's purpose and limits.


## Risk 10: Nobody owns the chatbot after launch

The most common chatbot failure is organizational. Prompts, knowledge bases, and models change weekly, but nobody re-tests disclosures, answers, data flows, or accessibility. The NIST AI Risk Management Framework emphasizes governance, ongoing measurement, and clear accountability across the AI lifecycle [12].

What a fix looks like: name an owner, keep an inventory of chatbot vendors, prompts, and data flows, require review before prompt or model changes go live, and monitor transcripts and network behavior for drift.


## AI chatbot compliance checklist

- Clear AI disclosure in the widget, first message, and when asked; easy route to a human.
- Recording and vendor notice at chat start; privacy notice updated to cover chat data.
- Network map of every domain receiving chat content; no data sent before the user submits.
- Vendor terms reviewed for training use, retention, sub-processors, and transfers.
- Answers limited to approved content; high-risk topics blocked or escalated; test set run after every change.
- Sensitive-data warning, redaction, short retention, restricted transcript access.
- Prompt-injection and data-leakage testing; least-privilege tool access; server-side authorization.
- Keyboard and screen-reader testing against WCAG 2.2 AA; mobile overlay check.
- Documented decision on children's use and automated decisions with human review where needed.
- Named owner, change review, and ongoing monitoring with evidence kept.

## A 30-day chatbot remediation plan

- Days 1–5: inventory every chatbot, vendor, model, prompt, and integration; capture network traffic and screenshots as a baseline.
- Days 6–12: add disclosure and recording notices, update the privacy notice, and change vendor training and retention settings.
- Days 13–20: build and run the answer test set, restrict risky topics, and add human escalation.
- Days 21–26: run security and accessibility testing; fix the highest-severity issues first.
- Days 27–30: verify fixes, record evidence, assign an owner, and schedule monitoring.

## What an automated scan can and cannot tell you

An automated scan can detect chat scripts, third-party domains, cookies, missing labels, and some accessibility issues. It cannot judge whether your disclosures are adequate under the laws that apply to you, whether the bot's answers are accurate, or whether vendor contracts restrict data use. Those need human testing, document review, and, for legal determinations, qualified counsel.


## How ComplyMynt helps

ComplyMynt assesses AI and SaaS products across six domains: AI governance, privacy, consent, accessibility, public security posture, and legal and policy surfaces. For chatbots, that means testing disclosures, mapping chat data flows and vendors, running answer and prompt-injection test sets, checking keyboard and screen-reader use, and delivering an evidence-backed remediation backlog. Our six-stage method moves from Discover and Collect to Assess, Remediate, Verify, and Monitor, and continuous monitoring is recommended after an audit to catch drift when prompts, models, or vendors change.


## FAQ

### Do I have to disclose that my website chatbot is AI?
Often yes. California's bot law, Utah's AI Policy Act, and the EU AI Act's transparency rules require disclosure in many situations, and a clear disclosure also reduces deception risk under consumer-protection law. The safest practice is a persistent 'AI assistant' label and an honest answer when someone asks.

### Can my business be liable for wrong chatbot answers?
Yes. In Moffatt v. Air Canada, a tribunal held the airline responsible for incorrect information its chatbot gave. Limit the bot to approved content, link to official policies, and escalate commitments to a human.

### Why are companies being sued over website chat?
Many U.S. lawsuits claim that chat vendors record or intercept conversations without consent under wiretap laws such as California's CIPA section 631. Clear recording notices, vendor contracts, and counsel review reduce the risk.

### Is chatbot data covered by GDPR?
Yes, when it relates to identifiable people in scope of the GDPR. You need a lawful basis, transparency, minimisation, retention limits, vendor contracts, transfer safeguards, and a way to handle access and deletion requests for chat data.

### How often should a chatbot be re-tested?
After every change to the prompt, model, knowledge base, vendor, or integrations, plus on a regular schedule. Chatbots change faster than most website features, so ongoing monitoring matters.

## Authoritative references

- [[1] California Business and Professions Code §17940–17943 (bot disclosure)](https://leginfo.legislature.ca.gov/faces/codes_displayText.xhtml?lawCode=BPC&division=7.&title=&part=3.&chapter=6.&article=)
- [[2] Utah Code Title 13, Chapter 72 — Artificial Intelligence Policy Act](https://le.utah.gov/xcode/Title13/Chapter72/13-72.html)
- [[3] EU AI Act — Regulation (EU) 2024/1689 (Article 50 transparency)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj)
- [[4] California Penal Code §631 (California Invasion of Privacy Act)](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=PEN&sectionNum=631)
- [[5] Moffatt v. Air Canada, 2024 BCCRT 149](https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html)
- [[6] FTC — Keep your AI claims in check](https://www.ftc.gov/business-guidance/blog/2023/02/keep-your-ai-claims-check)
- [[7] FTC — Operation AI Comply (2024)](https://www.ftc.gov/news-events/news/press-releases/2024/09/ftc-announces-crackdown-deceptive-ai-claims-schemes)
- [[8] General Data Protection Regulation (EU) 2016/679](https://eur-lex.europa.eu/eli/reg/2016/679/oj)
- [[9] UK ICO — Guidance on AI and data protection](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/)
- [[10] FTC — AI companies: quietly changing your terms could be unfair or deceptive](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive)
- [[11] OWASP Top 10 for Large Language Model Applications](https://owasp.org/www-project-top-10-for-large-language-model-applications/)
- [[12] NIST AI Risk Management Framework and Generative AI Profile (AI 600-1)](https://www.nist.gov/itl/ai-risk-management-framework)
- [[13] U.S. DOJ — Guidance on Web Accessibility and the ADA](https://www.ada.gov/resources/web-guidance/)
- [[14] W3C — Web Content Accessibility Guidelines (WCAG) 2.2](https://www.w3.org/TR/WCAG22/)
- [[15] FTC — Children's Online Privacy Protection Rule (COPPA)](https://www.ftc.gov/legal-library/browse/rules/childrens-online-privacy-protection-rule-coppa)
- [[16] FTC — Inquiry into AI chatbots acting as companions (2025)](https://www.ftc.gov/news-events/news/press-releases/2025/09/ftc-launches-inquiry-ai-chatbots-acting-companions)