Aria and I Built a Website!
Disclaimer: This content is NOT sponsored by WIX. All opinions and experiences are my own and should not be seen as a definitive description of Aria’s capabilities or flaws.
TLDR: I accidently did some H/M Teaming when making my new website, and it was an… interesting experience. Some benefits, lots of limitations, and many things for us to keep in mind as we move forward when integrating “smarter” AI into our everyday lives.
I made a new website, and built it with the help of Aria, Wix's AI software. Despite writing about AI for years, this has been my first experience working in a Human/Machine Team! So, let me tell you about it!
In my PhD I spent a lot of time deciding how I wanted to define the term "human/machine (H/M) team". Despite it seeming like a straight-forward concept, H/M Teaming was incredibly difficult to define, especially when writing speculatively about the future rather than relying on past and contemporary examples to frame such a term. I settled on a definition that prioritized the quality of machine teaming, rather than any specific capability. Lakhmani et al quote McNeese et al (2018) in their article “Cohesion in Human-Autonomy Teams”, and argue that H/M teaming is:
… at least one human and at least one autonomy-enabled system or intelligent agents, who work interdependently towards shared goals.1
For my work, H/M teaming refers to distinct2 humans and machines working together, each with some element of agency, to complete a task interdependently. For Instance, writing this blog post does not involve H/M teaming, because it is only me who is working on it. AI software is not editing the text, telling me about the text I am typing, or formatting the layout. I am writing about AI, and so it does have more agency in the creation of this content compared to when I write blogs about other topics. However, that agency is distant to my current activities, diminishing its role and thus any teaming that might take place. The AI is not itself working with me to complete this goal.
Creating a new website was a very different experience.
*
I did not intend to engage in H/M teaming when building my website. Rather, I expected to find pre-determined templates and widgets, to edit some colours, and to get frustrated at a lack of comprehensive customisation. Instead, I was presented with a website editor that gathered my information (with permission) from when I set up a domain and generated a professional looking website, with pre-written titles, headings and descriptions. I’m sure I could then interact with the interface directly, to move elements, personalise visuals and correct the words the AI came up with (as it was typically incorrect). Yet I decided to engage with Aria, Wix’s AI chat software, to navigate the editor.
I did this primarily because I was tired, a little brain-foggy, and didn’t know how to use the editor software. I could have learned, but I decided I would try out the AI first in case it took away the need for me to figure out how all the features available to me worked. Generally, I asked Aria to move widgets around, change colours and images, add functionalities, and to clarify what elements did. The latter task was particularly valuable as it allowed me to ensure that the elements in question did what I wanted them to do, and when they did not, Aria explained why that was and what alternatives might work instead. This helped me to feel more in control of what was going on as the website was being built and allowed me to have more agency in what was happening than I would do using a more restrictive website building apparatus.
What struck me at this point wasn’t the intelligence of the AI system, but the human-like interactions we were having. Aria answered my questions in ways which seemed – at least linguistically – thoughtful and reflective. It would respond by saying things like, “good idea”, “this will really help your engagement”, and “if you would prefer a different solution, I can suggest…”. The dialogue was not mechanical, or purely technical (that is, heavily focused on specific tools and abilities). This is somewhat deceptive, I think. It presents Aria as if it can weigh up options and think for itself. Of course, it cannot, Aria is a Large Language Model (LLM) that uses set data to generate likely sequences of words in response to inputs. But the ease with which I was drawn into this perception – because of the kind of language Aria used (or rather, was designed to include when generating sequences) – created a sense of genuine interaction and teaming. I will talk more about the implications of this in a moment, but I should note that this is not a wholly negative situation to have found myself in.
This H/M teaming (however artificial it is in reality) was one which made building a website more accessible, and – at least from my perspective as the user – cheaper. Not needing to receive a lengthy tutorial to use a comprehensive piece of software or to constantly refer to user guides, meant that I could prioritise creation over learning, and that I could unlock more advanced features and tools without needing to interact with gatekeeping like pay walls or more advanced handbooks. As a result, I was able to build a more advanced website, with more informed choices about what I could include and how. Of course, my choices were based on what Aria told me, and I do not know what tools and features I was not shown, or whether they would have been helpful for the site that I built. However, given the level of technical skill I have alone, the website Aria and I built is far more advanced than what I could create myself even if I was given access to tutorials and information about the breadth of features, functionalities and tools available within the free WIX website builder. Indeed, regardless of the reality of choice and autonomy I was afforded by liaising with Aria, I perceived that I was able to make choices and create something sophisticated and comprehensive.
Access to this kind of output is important, as it allows people without specialist training and financial and/or time resources to create something that would otherwise be reserved for someone who was or could hire web design professionals to develop this quality of site. Indeed, as AI advances and becomes integrated into more user-facing interfaces, “ordinary” users will be able to complete tasks that would otherwise be inaccessible to them. This democratises this work and empowers people to build more advanced products with fewer barriers, and lowering the threshold for sophisticated outputs and marketing.
Inevitably, this will probably just raise the standards for entry level outputs, and expectations will increase as the quality of entry-level, specialist and professional outputs recalibrate. Regardless, the possibilities for accessibility stand. Concerningly, however, so do the problems of outsourcing professional skills to AI. Here, accessibility runs parallel to the technologizing of the work of humans, providing previously specialist tools to everybody, whilst removing opportunities for paid work for these professionals. Of course, humans are involved in designing the AI, and the editing interfaces that users use. However, this does not patch the gap that AI will create.
*
Added to this is the ease with which users are encouraged to use AI tools in place of professionals. For instance, when I started populating my blog with material, there were buttons on the user interface which would navigate users to AI tools, next to identical buttons referring to (presumably human) professionals to help / write a blog. This surprised me, as it placed AI tools as valid and equal alternatives to human input, making it easy (if not easier) to engage AI tools to generate blog content. The equality of these tools in the design of WIX’s interfaces demonstrates the ease with which WIX wishes to provide AI tools and assets to support a person’s business and marketing, not just as a replacement for human specialist skills (by providing default AI assistants, like Aria), but as a valid and equal alternative.
What was concerning is the ease with which users could interact with AI tools, compared to the difficulty I had in getting Aria to comprehensively claim its own work. Here, there was an imbalance between the level of AI involved in (and that potentially used) in creating my website, compared to the level of disclosure that AI was used. Whilst I built the website, there was no inbuilt statement to tell visitors that it was built using AI, nor was there labels on AI created content (which I will talk about in a moment) to tell readers that this was created by AI, rather than by me. I had to ask Aria to add a declaration of its involvement, including details of their software. In response Aria wrote “This website was constructed using AI”. After several more prompts, and some words added of my own, Aria eventually generated the disclaimer that you can currently see on the website:
This digital space emerged through collaboration with Wix AI, specifically Aria. We harmonized automated structural logic with Ken's specialized perspective on divinity and defense to build this interface. Learn about the process in the blog entry titled "Aria and I Designed a Website".3
Even this disclaimer lacks substance, telling the viewer only that the website was designed using AI, and that I have written about it. It is also typically cringy, using buzzwords like “harmonized” to make the sentence sound more corporate. This does not tell the viewer how the AI was used, what kind of AI it is, or whether it is designed by WIX or a third party. The time it took me to prompt this disclaimer was concerning, as it demonstrated that the guidance given to Aria about declaring its own presence on the websites it makes is basic and uncomprehensive, even when given substantial instructions. Of course, I could have written something better myself, which met my needs, but part of this task was to see what Aria would generate itself, so that I could write about it here.
This is coupled with the wealth of automatically generated AI content created when the website was set up. I assume this was to allow users to see what blogs would look like, and to edit the features around them, however, I was not explicitly asked for permission4 for the AI to generate content, nor would it be clear to viewers that these blogs were made by AI if I were to accidently (or on purpose) publish the website whilst they were still present on Wix’s blog app. Of course, it was clear from the content that they were AI-generated, but they still ought to have been labelled as such. This abundance of AI-generated content by default, and the lack of clarity about what this content was for and how it would work, is concerning. For instance, I did not know that this content would go live when the website was published unless I manually deleted it, until I explicitly asked Aria about it. This left open the risk that content about me (as generated by the AI) would go live on the internet without my full knowledge or consent, and could potentially be picked up by other LLM’s combing the web, and used to generate content about my work which would be fundamentally incorrect. Thankfully I realised this was a possibility and deleted the content before publishing my website.
This highlights a key aspect of H/M teaming, which is that work which AI does independently / as default without the human being aware. In most cases, these tasks are not transparent or fully communicated to the user, meaning that humans who are unfamiliar with a system (like I am with Aria) may not know what tasks have been carried out on their behalf. In the context of my website, this left a wealth of AI-generated (and ultimately fake) content open to being published without my knowledge. Or, more concerningly, published on purpose as a way of making the website look more populated than it was. This frequency with which WIX asks it’s AI to produce realistic looking content without labelling it as AI generated, or communicating to users what it has created, presents an image of H/M teaming where the human is always left ill-informed about the full work that the AI is doing, leaving them vulnerable to unknowingly publishing AI content, including incorrect information about themselves and others. Thus, creating a teaming that is ultimately opaque, and imbalanced in terms of consent and knowledge, where the human is significantly vulnerable to the quiet working of AI which may go against their aims and intentions.
*
This problem of opaque-ness is compounded by the reality that AI is not that good at doing the tasks we ask it to do. This leaves users in a context where they are potentially unable to create the things they want to, unable to understand the things they have created, or – worse – create real-looking fake content that they do not know how to identify and navigate. For instance, repeatedly Aria would tell me that something I wanted to do was not something that it could facilitate for me, and I had to do it manually using a different tool. This wasn’t a problem, except that the instructions Aria gave me to do this (click this button, select this option, etc) were often incorrect, either because the options I would choose wouldn’t do what I wanted (and what Aria said they would do) or because the buttons and options didn’t exist in the first place.
Likewise, Aria would occasionally state that it had completed a task without actually doing so. For instance, I asked Aria to reshape a box to make it oval rather than rectangular, and when Aria said that it had been done, the box was still rectangular. After refreshing the page, in case the page was taking time to load the new asset, the box was still rectangular. In both of these examples, I would say that the instructions were incorrect or that the task had not been completed, and Aria would apologise and try again. With regard to the instruction, I was told that the problem was that I was using an older or newer software than the one Aria had access to, and so the instructions would not be applicable on this version. Whether this was actually true, I do not know. Indeed, several times I was told that the functionality Aria had been instructing me to access in fact did not exist at all, and that directing me to it was a mistake. These were frustrating interactions, and drew into sharp focus the problem with using AI in place of other access tools and interactions with professionals – AI is always confident about its answers, even when it is wrong. For, it does not “know” whether it is right or wrong, it just responds with what it generates to be a likely sequence of words to respond to the question asked.
*
This, together with the false instructions Aria gave, automatically generated content, and the potential gatekeeping of certain features (by failing to offer them as options to answer a user’s needs) creates the potential for poor results, without the human involved knowing that the results are, indeed, false, misleading or of poor quality. This offsets the accessibility and democratising potential of this technology, leaving the human at the mercy of AI’s capabilities and weaknesses, and the (inevitably good, but poorly communicated) intentions of its designers. For, if it was not for my pedantic checking that features were correctly added, the properties of certain content, and my willingness to tell Aria that it was wrong, I may not have known that the features included were incorrectly applied, or that information was incorrect. This places a significant responsibility on the humans in H/M teams like this one to check, question and challenge the work of their machine teammate.
In a world where AI (especially LLM’s) is being increasingly integrated into products and services, in ways which are not transparent (or disable-able) to the user, we will need to be more aware and literate about these tools, so that we can effectively interrogate the work it produces, the tasks it undertakes, and its limitations. For, this technology is probably not going anywhere, and we will increasingly be confronted with opportunities for H/M teaming with AI agents that we know little about, and which vary in their strengths and willingness to disclose themselves. Of course, this is not all bad, the ability to team with powerful AI’s does make certain products more accessible, and allows for people to create advanced and sophisticated outputs cheaply and quickly. Yet, if we are not more literate about this teaming, we risk accidently producing content that is riddled with unexpected elements, mistakes and false information, limiting its usefulness and impact as content.
References
1 S. G. Lakhmani et al, “Cohesion in Human-Autonomy Teams: An Approach for Future Research”. Theoretical Issues in Ergonomics Science 23, no.6 (2022): 687-724 (688). DOI:10.1080/1463922X.2022.2033876.
2 By “distinct” I mean separately identifiable agents, these agents don’t need to be distinct in other ways, and can include systems made up of multiple human, and machine, components, and machines imbedded in human physiology.
3“Vision and Collaboration” from drkennethwilkinson-roberts.com.
4 I am sure I did in fact give consent for this, but in a way which was unclear to the user without reading the full terms and conditions. Thus making it easy for users to consent to something they were not fully aware of due to a preference for speed over being fully informed (a preference which I – embarrassingly – have when it comes to interacting with websites like this).
Comments