Limits of AI in web design: the decisions that stay human
Partner · Web development, SEO and GEO
AI can write the code, the copy and the variants of a website in a few hours. That does not mean it decides what your site must achieve, which pages should exist, what your business promises or what the law requires of you. Those decisions stay human because someone has to answer for them: to your customers, to the Office québécois de la langue française, to Quebec's privacy regulator (the Commission d'accès à l'information) or to a court. Google also asks you to check everything AI produces by hand before publishing it. The simplest split: management decides, AI executes, a named person reviews and signs off.
- AI predicts the most likely sequence of words: it does not check facts. Google says so in its guidance, updated on October 1, 2026.
- What your site states commits you, even when AI wrote it: a statement or an advertisement binds the merchant.
- French-language rules, personal information and accessibility call for judgment on facts AI cannot see: your establishments, your forms, your customers.
- AI writes and designs what is most common. Your brand's voice and look are decided in writing, before the first instruction.
- Every publication needs a named person who approves it. Automated checks help that person; they do not replace them.
On this page
Human oversight
Human oversight is the division of roles that keeps a person accountable for every important decision when AI does the work. AI prepares, writes or codes. A person sets the framework beforehand, checks the result afterwards and approves publication. It is also called "human in the loop."
Who decides, who executes, who reviews
Agents like Claude Code (Anthropic's coding tool) and ChatGPT Codex (OpenAI's coding agent) now write a site's code from an instruction in plain language. Our article on what AI changes in website development describes the effects for a business and names the decisions that stay human. Here we go through them one by one, with what can happen when they are left to AI. The table sums up a division of roles that works for a redesign as well as for weekly changes.
| Decision | Who decides | What AI can execute | Who reviews |
|---|---|---|---|
| Strategy and goals | Management | Summarize your data, compare competitors' sites, draft a brief | Management, with the strategist |
| Pages to create | The strategist, with management | Group searches by intent, suggest a site structure | The SEO lead, then the sales team |
| Message and voice | Marketing | Write versions that follow your brand guide | The person responsible for the brand |
| Facts and promises | The person who knows the offer | Flag the prices, warranties and lead times to check | The offer owner, before each publication |
| Legal obligations | Management, advised by a lawyer when needed | List the points to check: French version, forms, cookies, alt text | The person in charge of personal information protection, a lawyer when needed |
| Design | Management and the designer | Produce variants, apply the visual identity | The designer, on mobile and on desktop |
| Publication | A named person | Run the automated checks, prepare the version to publish | The person who signs off |
When the stakes are high, the person who reviews is not the one who did the work: the author of a page, human or AI, is poorly placed to see its own mistakes.
The site's strategy and goals
A website serves a business goal: qualified leads, online sales, job applications, fewer calls to customer service. AI can summarize your data and draft a well-structured brief. It cannot choose for you between attracting more requests and attracting better ones. The answer depends on your margins, your capacity and what you want to sell next year.
Left to AI, strategy takes the most common shape: a site that presents the company, its services and a contact form. Nothing in it is wrong, but nothing in it steers visitors toward what you really want to sell.
Google's guidance on helpful content frames the question differently. Does your site have a primary purpose? Does your content show first-hand expertise, the kind that comes from actually using a product or serving a customer? AI has never served any of your customers: it shapes your expertise if someone gives it that expertise.
Our article on website redesign details the goals to set before you start. With AI, this step matters more: code gets written faster, so a poorly framed goal could reach your live site faster too.
The pages that should exist
The list of pages is the second decision. It is built around what your customers look for: the services they type into Google, the questions they ask ChatGPT, the objections they raise on calls. AI is useful here: it groups hundreds of searches by intent (what the person wants to accomplish) and suggests a site structure in minutes.
It does not know which searches lead to sales for you. Two phrases with the same search volume can have very different value for your business: one attracts buyers, the other students working on an assignment. Nor does it know which services you want to put forward.
The opposite risk exists too. Google warns that producing many pages without added value for readers may violate its spam policies (scaled content abuse). With AI, a hundred city pages can be produced in an afternoon: deciding not to produce them remains a human call. Our article on designing a site for visitors, SEO and AI explains how to organize these pages.
The brand's message and voice
A language model does not look for the best sentence: it predicts the most likely sequence of words based on its training data, as Google writes in its guidance. For brand copy, AI therefore naturally writes like the average of your industry. The copy is correct. It could also appear on your competitor's site.
Voice is decided in writing, before the first instruction: the words you use, the ones you avoid, how a price is displayed, what your site never promises. AI needs precise rules, with examples of accepted and rejected sentences.
We apply this method to our own site. As of October 20, 2026, falia.co ran 28 automated checks before every publication, several of which verify writing rules (anglicisms in French, punctuation, the exact names of our offers). Those rules did not come from AI. In September 2026, Claude used an anglicism in a French email we were preparing. A person caught it. Since then, the phrase to avoid has been on a list the site checks at every publication. AI applies the brand voice; it does not invent it.
The truth of facts and promises
This is the decision that can cost the most when it is delegated. AI sometimes confidently invents a figure, a warranty or a lead time: this is called a hallucination (false information presented as true). Two 2025 studies show the scale.
- On the news. In October 2025, the European Broadcasting Union (EBU) and the BBC published an evaluation of more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity by journalists from 22 public service media organizations. 45% had at least one significant issue and 20% contained major accuracy issues, including made-up details.
- On the law. Published in 2025 in the Journal of Empirical Legal Studies, a Stanford study tested legal research tools from LexisNexis and Thomson Reuters designed to limit these errors. Their answers still made up information 17 to 33% of the time. The versions tested dated from 2024.
On a website, a hallucination can look mundane: a two-year warranty instead of one, a delivery time that is too short, a certification you do not hold, a price that no longer exists. The law makes no exception for an AI error. Section 41 of Quebec's Consumer Protection Act provides that a statement or an advertisement binds the merchant. Section 219 prohibits any false or misleading representation "by any means whatever." Section 228 also prohibits failing to mention an important fact.
A British Columbia tribunal has already ruled on the question for a chatbot. In February 2024, Air Canada was held liable for the inaccurate information the chatbot on its site had given a customer about a bereavement fare. The airline argued that the chatbot was responsible for its own actions; the tribunal answered that the airline is responsible for all the information on its website.
Promises are checked one by one: prices, warranties, lead times, certifications, results, comparisons with a competitor. The person who checks them is the one who knows the offer, not the one who wrote the page.
French, personal information and accessibility
Three frameworks apply to almost every Quebec business website. AI can list the points to check. It cannot decide whether your site complies, because the answer depends on facts it cannot see: where your establishments are, what happens to the data from your forms, who you sell to. This section does not replace legal advice.
French: Bill 96
The Office québécois de la langue française points out that section 52 of the Charter of the French Language applies to commercial publications on a website. In January 2025, a Montreal company whose site existed only in another language was convicted a second time and fined $1,500. AI translates quickly. It does not check that every page, every form and every confirmation email exists in French under conditions at least as favourable. Our article on Bill 96 and your website details this test.
Personal information: Law 25
Law 25 requires a privacy impact assessment for any project to develop or redesign an information system or an electronic service delivery system that involves personal information. A site whose forms collect names and email addresses could be covered. The same assessment is required before communicating personal information outside Quebec, for example to an AI tool hosted elsewhere. If your site lets AI make a decision about a person on its own (turning down a quote request, for example), you must inform that person and give them the chance to submit observations to a member of your staff. Your site must also publish the title and contact information of the person in charge of personal information protection.
Accessibility
In Quebec, the accessibility standard applies to public bodies. A private business can still be required to make its site accessible through other routes, as our article on web accessibility in Quebec explains. In February 2026, WebAIM analyzed one million home pages: 95.9% had detectable WCAG failures, up from 94.8% in 2025. According to WebAIM, this increase likely reflects, among other things, AI-assisted coding. Automated tools only detect part of the problems: a person also tests the site with a keyboard and a screen reader.
The design choices that carry your image
AI produces ten variants of a home page in a few minutes. The choice between them tells your customers who you are, what price range you sit in and who you are speaking to. As with copy, AI leans toward what is common. Without precise direction, the site it designs could look like many others.
Two points call for an explicit decision. The first is ownership. In the United States, the Copyright Office concluded in January 2025 that prompts alone do not provide enough human control to make the user the author of what AI produces. Before making a generated image the core of your identity (a logo, a brand illustration), ask a lawyer what you will be able to protect.
The second is transparency in online retail. Google Merchant Center policies, referenced in Google's guidance, require AI-generated images to carry metadata identifying them as such. AI-generated product titles and descriptions must also be labelled. The rest (readability, contrast, placement of the call-to-action button) is judged on the site itself, on mobile as well as desktop.
The final sign-off
The last decision: who agrees to publish. Google asks you to manually check the accuracy of all AI-generated content before publishing it, including title elements, meta descriptions, structured data and image alt text. These elements are barely visible on the page, but they appear in search results.
Automated checks can block a broken link, a badly written price, a page missing its English version. They do not know whether the warranty stated is the right one. On our own site, each of the 28 checks translates a decision we made. The sign-off remains a person's. Claude Code executes, a senior expert decides and signs off.
A named person for each type of content
Prices and promises do not necessarily have the same approver as blog articles.
A separate working copy
AI changes a copy of the site, never what your customers see. Publishing is a separate step and versioning lets you return to the previous version.
A review proportional to the risk
A typo in an article can be fixed after publication. A price, a warranty or a legal notice is checked before.
Seven questions to settle before the first instruction
- Which business goal should the site serve first?
- Who chooses the pages to create? Based on what data?
- Is our voice written down somewhere, with examples?
- Who checks prices, warranties and lead times before each publication?
- Who is in charge of personal information protection? Is their title on the site?
- Who approves the design on mobile and on desktop?
- Who signs off on publication?
AI does the work faster. It does not answer for what your site promises.
Our AI-managed websites work on this split: Claude or ChatGPT executes within your brand guide, automated checks block anything that breaks a rule, then a senior expert approves publication. You can also approve it yourself if you keep control.
Want to sort out what AI can take on and what should stay with you? In a marketing consultation, a partner spends 90 minutes with you and sends you a written summary.
Frequently asked questions about the limits of AI in web design
Can AI build my entire website?
It can write the code, the copy and the images. It cannot decide what the site must achieve, which pages are useful to your customers, what you promise or what the law requires of you. A person sets those decisions and checks the result.
Who is responsible if AI writes a mistake on my site?
Your business. Quebec's Consumer Protection Act provides that a statement or an advertisement binds the merchant. In 2024, a British Columbia tribunal held Air Canada liable for inaccurate information given by the chatbot on its website.
Does Google penalize a site written with AI?
Not in itself. Google asks you to manually fact-check all AI-generated content, including titles, meta descriptions and alt text. Producing many pages without value for readers may, however, violate its spam policies.
Can I have my site translated by AI to comply with Bill 96?
Yes, as a starting point. The French version must be complete and offered under conditions at least as favourable as the other version, forms and confirmation emails included. A person fluent in French reviews it before publication.
Can I put my form data into ChatGPT?
With caution. Under Law 25, communicating personal information outside Quebec first requires a privacy impact assessment. The simplest approach is to keep personal information out of your instructions: AI does not need it to change a page.
Do I own a logo created with AI?
The question is not settled the same way everywhere. In the United States, the Copyright Office concluded in January 2025 that prompts alone are not enough to be the author of what AI produces. Before making it your identity, ask a lawyer for advice.
- Google Search Central, Google Search's guidance on using generative AI content on your website, updated October 1, 2026, accessed 2026.
- Google Search Central, Creating helpful, reliable, people-first content, accessed 2026.
- European Broadcasting Union (EBU), News Integrity in AI Assistants, international study with the BBC, October 22, 2025.
- Magesh, V. et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Journal of Empirical Legal Studies, vol. 22, 2025.
- LégisQuébec, Consumer Protection Act (CQLR, c. P-40.1), sections 41, 219 and 228, accessed 2026.
- Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149, February 14, 2024.
- Office québécois de la langue française, Infraction à la Charte de la langue française : le site Web de Fournitures Hermes inc. (in French), May 9, 2025.
- LégisQuébec, Act respecting the protection of personal information in the private sector (CQLR, c. P-39.1), sections 3.3, 12.1 and 17, accessed 2026.
- Commission d'accès à l'information du Québec, Principaux changements apportés par la Loi 25 (in French), accessed 2026.
- WebAIM, The WebAIM Million 2026, data collected in February 2026.
- U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability, January 29, 2025.
Geneviève puts the strategy for your engagement into action. She leads all our web development projects: Shopify, WordPress and the new ways of building a site with AI. She manages our team of developers and translates your business needs into technical language. She runs your organic search (SEO), your visibility in AI answers (GEO) and your site's conversion rate optimization (CRO). Her work is at the heart of three goals: Attract customers with SEO and AI, Improve your site's conversion, and Strengthen your visibility in AI answers. With Gabriel, she also builds the landing pages for your advertising campaigns. She writes mainly about SEO, AI visibility and web design.
About Falia →
Strategy
SEO
Paid advertising
Conversion
AI visibility
Web design