A Better Internet: AI and Responsible Context

Copy to clipboard

Web 2.0 has been a battleground for your context

It’s no secret that the tech giants that underpin the internet have grown so powerful because of the data they have collected on us all as we skip our way through our online lives. Google tracks what we look for, where we shop, what we email and what we watch. Meta, TikTok, Snapchat, Reddit etc. have intimate knowledge of the vast amount of content that we consume, comment on and share with our friends and foes. Even the current crop of AI giants are figuring out how to monetise your attention. By the time you read this, OpenAI (ChatGPT) and Anthropic (Claude) may already be presenting you with some content that advertisers have paid for you to see. This could be in the form of sponsored results (like Google Search) or right in the body of an AI’s chat response.

This isn’t a groundbreaking insight from me - this is so well known at this point that you’re already losing interest in this article. Here’s a phrase that I coined to explain this succinctly: “If you're not paying for the product, you are the product”. I didn’t coin it but it would be clever if I had.

Copy to clipboard

Where this has led us to now

The outcome of this battle for your context and engagement has had a number of negative effects which other researchers and commentators are far better placed to discuss than I. However insofar as we’re discussing our online worlds in the context of, well, context, it’s useful to raise a couple of issues, aka ”externalities” to borrow a dry economics term. Firstly, in the race to secure our valuable data the platforms have engaged in the most ingenious hacking of human psychology and neuroscience ever. They’ve built digital experiences which are so engaging on the fundamental brain-hardware level that our attention is hijacked for hours a day, more often than not in the full awareness that it’s terrible for our own wellbeing.

All of this attention has been converted into a huge trove of data on each and every one of us - almost every bit of content we engage with online is tracked; what we did to get to that content, how we interacted with the content and what we did next. All of this information is refined by tech companies and turned into their best understanding of who we are. If you think this sounds like a pretty poor way to get a true picture of an individual person then I would agree - what’s missing is real human understanding, aka “context” as we shall talk more about below.

Below are some tools that have been developed for tracking what you do online. Sometimes these tools (like cookies or session replays) can be used more innocently to enable a smooth web experience or to improve user experiences. However for the most part these tools are like private detectives building a picture of what you are doing in your digital life to get a better understanding of you as a potential asset to be exploited.

Despite the plethora of tracking tools, we see the shortcomings in this approach every time we’re served irrelevant content (paid and organic), trapped in our echo chambers, or consistently presented with the same content because once upon a time we showed a vague passing interest in it. How often are you followed around the internet by that item that you left in your shopping cart somewhere or mentioned to a friend in an email?

The exploitation of our context is janky at best and a nefarious tool for manipulation at its worst. We are constantly being bombarded by nudges to go in directions that ultimately lead to nowhere or, worse, lead to a waste of energy and time and adverse impacts on our happiness.

Copy to clipboard

Let’s talk about context in context

I realise that I’ve just criticised the tech giants for hijacking your attention so I am going to try to be brief. For those of you who haven’t played with AI tools a great deal, you might think of context as yet another simple word for a complicated technical concept. Engineers often use such words to make AI appear abstruse (I learnt this word recently) and impossible to understand (this is what abstruse means). It really isn’t: context means exactly what you think it does - it is the relevant information that surrounds a particular domain and improves a person’s or machine’s understanding within that domain.

Think about it this way - an AI vision model reads an image and processes the pixels (a guy with a fancy whiteboard explains) and interprets it as a dog catching a frisbee in a garden. A human tells AI that this is in fact a photo of their childhood dog playing in their parents’ garden 20 years ago. The AI is now able to combine its semantic understanding of what it’s looking at with this context. It knows that 20 years ago you had a dog (AI has identified this as a German Shepherd) and you used to play frisbee with it in your parents’ grassy garden. With its knowledge of German Shepherd lifespans, AI might infer that your dog is now dead.

Moving on from dead dogs: consider an AI-powered robot that you’ve asked to go food shopping; task context for it would be you telling it:

• what you want to eat for dinner
• when you want to eat

Seems obvious, right? However, bad outcomes could occur if the ShopBot didn’t already know relevant personal context like:

• you recently discovered that you’re intolerant to dairy (check ingredients for milk - which seems to be in everything)
• your little kitchen only has a hob and you don’t own an oven or microwave

ShopBot would need to know certain additional personal context:

• you only have £25 left in your bank account
• your supermarket loyalty card has active vouchers for half-priced orange-flavoured, chocolate-covered jelly discs (and you love Jaffa Cakes)

This is a stupid example but you get my point. You understand that the AI-robot has a legitimate need for context to best achieve a specific outcome to your satisfaction. You understand why you need to share your personal context and you also understand why the ShopBot would need to do additional work on your behalf to check other environmental contexts that exist out there (e.g. opening hours, stock availability). Side note: AI models have fantastic environmental context after chewing up the entire internet, for example, you don’t need to explain that a supermarket is a place where you buy food, or what dairy-free means. AI models understand the concepts of cheaper, fresher, spicier etc. etc. This is why they are good at deriving intent and delivering outcomes.

Copy to clipboard

Web 2.0 took your context and used it badly

In the Web 2.0 era, prior to going out shopping you might have searched for “Dairy-free recipes for under £10 in hob-only kitchen”. The sneaky little internet detectives shown in the table above might infer that you’re allergic to dairy, don’t own an oven and are broke. The detectives (e.g. Meta’s pixel in the table above) would then start blabbing to pharmacies all about your intolerance, then they would head over to an appliance retailer and start shouting about your lack of oven and they might even decide to cross-reference what they infer about your financial situation with the credit data they have access to. The Buy Now Pay Later companies have heard that you’re not flush with cash right now and are cleverly telling you that you can buy a compact air fryer (worth it btw) and don’t need to worry about something as trivial as paying for it right now.

Not only does this feel sort of shady, it is also a fairly poor attempt at trying to serve your needs as a real human. You specifically asked for a recipe to inform your supermarket shopping and instead your attention is being diverted to a bunch of stuff you haven’t asked for. These detectives don’t know your intolerances are well under control (and something you don’t want reminding of every 2 minutes), you don’t want an air fryer and you really don’t want to go into debt.

Copy to clipboard

AI and consensual, relevant and responsible context

You might reasonably ask: in the era of widespread AI, isn’t this problem only going to get worse? Companies that track you online now have this incredible power in their hands to crunch through the vast reams of data they’ve collected on you and target you even more aggressively. You would be right to ask this and, yes, this is certainly going to be a risk.

I would argue that there are glimmers of hope for a better internet and relationship with technology. This hope stems from the ownership and management of context. Let’s go back to our ShopBot example above. Let’s imagine that ShopBot is powered by two things: a food shopping-focused AI model and detailed personal food and kitchen context. Let’s also imagine that ShopBot is driven by a single mission: shopping for your food. ShopBot won’t suggest news articles to you, it has no opinions on music, it doesn’t care that you wear shorts in winter and that you want the new foldable iPhone. It only cares about your financial situation insofar as it affects your food budget and it knows that any meal that requires an oven is a non-starter.

In this world, instead of your context existing as an inaccurate patchwork in the recommendation engines of tech companies, it resides within ShopBot’s single-minded and efficient AI brain. ShopBot is so effective at extracting semantic meaning and human intent from context that it knows after a week of just eating instant noodles (you told it this) that you might need some vegetables and fibre. You’ve also had a chat with ShopBot about your dietary goals and it might remind you that you’ve committed to eating more lentils and want to reduce your meat intake.

ShopBot is powered by an AI model (maybe a much smaller, shopping-optimised model) but as your agent ShopBot uses your context in ways that you’ve consented to and with only those third parties that are required to achieve your outcome. In the same vein, ShopBot will never attempt to use your context for anything besides buying food. In this world you don’t mind sharing personal, possibly sensitive context because you know it is going to be used responsibly in delivering the outcomes you desire. ShopBot is your trusted agent going out into the world (or the internet) with the right level of education about you to perform. Your context is not going to be sold to the highest bidder who has paid for an advert which targets “dairy allergic medicine”, “small kitchen” and “on a budget” in their paid content advertising campaign.

In this new world there is a catch - ShopBot isn’t free. ShopBot needs oil and battery charging and its AI brain costs money to run. The laws of economics still apply and somebody needs to pay for ShopBot - and as we know if you aren’t paying for the product then you are the product. That said, we would hope that AI models would become cheap enough to run (and robots cheap enough to make) that the cost of ShopBot would be well worth how effectively and privately your shopping gets done. Think of it this way: since Googling still requires effort on your part to assess the results, you might be reluctant to pay per search — so Google is free to use, but serves adverts and tracks you. If an AI-powered tool took care of most of the effort you might be willing to pay.

An artist's impression of ShopBot

Copy to clipboard

Leveraging context responsibly across domains

Maybe friends want to book a restaurant for your birthday surprise and maybe those friends care about what you might like to eat. With your permission ShopBot might be allowed to share certain specific context with your friend’s RestaurantOMatic to help them decide where to book for your birthday. You haven’t had a steak in weeks (because you’re broke) so you’re delighted to learn that your friends have booked Flat Iron (great, but you would have preferred Hawksmoor if they’re paying) - nonetheless they pick up the bill and you walk home licking your vanilla ice cream, extremely content with the world - you’ve temporarily forgotten you’re dairy-challenged (you really don’t have it under control). That’s a tomorrow problem.

Copy to clipboard

Wrapping up

Thank you for your attention and context (this blog has tracking that indicates that you’ve read this far in the article 🕵️‍♂️), here is a wrap-up:

• A better internet future could be: moving context ownership from a poor whole-person patchwork owned by tech companies to a collection of highly detailed domain-specific contexts owned by the human
• Protecting context and ensuring that it is used consensually, responsibly and to achieve outcomes that are relevant to humans and the task at hand
• Creating rich context through the voluntary sharing of information and powerful extraction of semantic meaning and intent

Copy to clipboard

Plugging Hawkker

I’m currently building a version of RestaurantOMatic that we heard about above. It’s called Hawkker and it really does follow the principles in this article. Users provide rich context about what they like to eat, what they think of the places they’ve eaten at and why they are interested in eating at places they are yet to visit. This context is shared amongst food-loving friends in the hope that we can all find better food to eat. You can join the waitlist at the link above.