There are so many problems with this idea (privacy, security, accessibility, deceptive advertising, design) that I'm impressed it actually got funded. If you're just going to have a LLM generate most of the content, why even sell this as a personalization thing and just have a LLM generate the entire website on the fly per request?
Fair criticism! We actually got funded for a different idea (this was a pivot) so perhaps you're onto something.... Jokes aside -- and to address your points.... Privacy: Visitors currently opt-in and provide only the company they are from (no other PII). Security: Our current version handles only text (we have images + dynamic content in the works), but we do our best to build a sanitisation pipeline around this, it's the same kind of problem that any arbitrary-code widget builder needs to account for. Accessibility: I'm keen to hear what concerns or ideas you have here since we are interpolating content and copying existing structure (so that a11y is kept as intact as possible). Advertising: Companies choose to integrate this themselves on their own website, so if anything we see it as forming part of their marketing strategy.
Also re: ‘deceptive advertising’ - if what you mean is that the output of the model will misrepresent what the host site actually sells/provides, that does happen on occasion, and it’s a top priority for us. We’re constantly working to reliably improve what Kenobi outputs in any given situation. It’s partially a research problem (i.e. deeply understanding the host website, and the nuances of their offerings, as well as research on the visitor, where the time window is seriously constrained), and it’s a model + prompting problem.
One of the ways we want to enhance Kenobi is to allow the host site to plug in more data sources (e.g. an FAQ, KB, etc). This helps solve the research problem, and it also increases the utility of the tool because more the very generic website can be be customized in more nuanced, specific, and esoteric ways, depending on the visitor.