Leon Furze

Just over a month ago, for reasons I can’t quite remember, I decided to try out a little experiment. I know it stemmed from that vague sense of frustration that Cal Newport claims most (tech) writers suffer at some point: the feeling of being thoroughly, pointlessly addicted to technology. Like pretty much everyone I know, I spend way too much time using my phone.

Between attention-seeking social media notifications, emails, and the near-constant urge to Google totally useless information, my phone is usually within arm’s reach. It’s a continuous pull on my often limited attention, and an unwelcome distraction from much more important things, like family time, eating, or doing pretty much anything that doesn’t involve staring mindlessly at a screen.

To make things worse, Generative Artificial Intelligence apps like ChatGPT, Microsoft Copilot, and Google Gemini are becoming more and more ubiquitous, and over the last 18 months, they have been developing interesting, engaging capabilities from simple text-to-text models to multimodal applications that are connected to the internet and capable of handling increasingly complex tasks. For someone who already uses their phone far too often, and now works with technology daily, it seems like a recipe for thumb-swiping disaster.

And yet, towards the end of February, I decided to test out whether GenAI could actually solve some of these problems, and if AI apps could comfortably replace one of the most frequently used apps on my phone: the browser. Even as I’m writing this, it seems like a weird experiment. Who in their right mind would willingly give up their main point of access to the internet for a whole month?

But, at the moment, GenAI apps don’t really use push notifications, infinite scroll, or pull-down-to-refresh; all of the classic tricks that social media uses to capture and keep our attention. So perhaps, I thought, GenAI could replace some of the more addictive apps on my phone and leave me with a device that was still functional, but much less alluring.

Out with the old…

LinkedIn: gone. Facebook: gone. Reddit, Instagram, Snapchat, and TikTok already gone a long time ago. Twitter… Well I never really used that anyway. One by one I deleted all of the main culprits for hogging my fickle attention – the social media apps with their gambling industry inspired logic and pervasive notifications, which seem to somehow. manifest even if you go to every effort to turn them off.

Next up, the browsers.

Mobile browsers don’t typically have notifications like social media, but they are designed to be addictive. Algorithms feed apps like Chrome, Edge, and Safari with your history and preferred content, generating news feeds that are just as dopamine-inducing as a Facebook feed or TikTok’s infinite scroll. The temptation to just pull out a phone and Google something whenever a question pops into my head (roughly every six and a half seconds, while I’m awake) is also strong, and often leads to a cascade of other apps. Better check my emails while I’m here… They had to go.

Removing the browser – Google Chrome, in my case – was actually a lot more difficult than the social apps because of the design of the Google ecosystem. It’s very hard to remove the Google Chrome browser and still have fully functional use of other applications, particularly when it comes to two-factor authentication and logins.

I also wanted to include Gemini in this experiment, but it doesn’t have a standalone app on iOS yet. So as a compromise, I tucked Chrome away in the “no one ever uses these apps” iOS App Library folder, and created a shortcut directly to Gemini on my homepage.

The Rules of the Game

This wasn’t really an experiment focused on GenAI: It was just a way to get me to stop using my phone so much, and in large part, I achieved that just by deleting the social media apps. But I didn’t just want to replace one addiction with another: I could see the risk of compulsively “chatting” with GenAI every time I needed a dopamine hit. So, alongside this experiment, I also set aside some time each day when I specifically would have access to either my desktop or my laptop. I limited (or tried to limit) social media, emails, and vague browsing to a finite window each day, and reserved AI for the incidental uses for which I would usually use the mobile browser.

Using GenAI to replace a mobile browser as a search engine is an obvious use-case for GenAI. Now that ChatGPT, Copilot, and Gemini all have access to the internet, there’s no reason to assume they can’t be used as alternatives to a traditional browser.

Of course, they work differently. There’s the problem of hallucination, confabulation, or bullshit to contend with, as well as things which just can’t be found, like specific information tucked away behind paywalls or institutional logins.

One thing that became clear over the course of the 30 days was that we use mobile browsers for much more than browsing. Even though we have separate apps for things like weather, maps, calculators, and other utilities, I find I just use Google Chrome for most of those features. That’s obviously deliberate on behalf of the developers. Once I got rid of Google Chrome, I found that I couldn’t trust generative AI to fulfil those functions. I didn’t want ChatGPT giving me directions or making up a weather forecast. Some of those apps which had been collecting dust suddenly sprang back to life.

Here’s a blow-by-blow account of each of those apps’ strengths and weaknesses in replacing a mobile browser:

Microsoft Copilot

This is by far the most browser-like application. Every interaction with Copilot includes footnotes and links to other websites. Clicking on those links takes the user to the websites without leaving the app, just like opening a new tab in any normal browser. Because it’s not a browser app, it misses out on some of the functionality which would otherwise allow you to keep track of multiple tabs, like bookmarks and so on. But just based on these early days, that’s definitely the way the Copilot app is heading. The ultimate goal seems to be for the app to represent the entire operating system and the whole ecosystem of Microsoft applications within it.

This is nothing new. In fact, Facebook has tried in many ways to become the internet, trapping users in its ecosystem by using features of the app that encourage liking, sharing, and saving webpages all in one place, and even trying to supply entire countries with internet connections through their platforms. This is just simple market capture: get a user and keep the user, whatever it takes.

ChatGPT

When you search for something with GenAI, it’s passing that search query to the language model and also using the language model to synthesise and summarise the results. Because of guardrails and safety features specific to the language models, this can also throw up some problems, censoring search results in unexpected ways or with the language model refusing to carry out particular tasks.

It’s happened a few times with Copilot, evading search queries or being overly conservative in its responses. But it was much more of a problem with ChatGPT. In fact, at the time of writing, ChatGPT – both the free version and the Plus version – is almost useless.

I’ve become used to using GenAI for things like quick summaries, but because of the ongoing battles over copyright, intellectual property, and data protection laws, it’s impossible now to get Chat GPT to adequately browse the web for information to summarise what it finds. Its responses when it comes to writing code or longer output are also lazy, with OpenAI claiming to absolve the app of all responsibility. Its voice is so formulaic and cliched that it’s painfully robotic. It’s hopeless at editing, revisions, or making useful suggestions about writing.

In fact, by the end of this process, the only reason I haven’t deleted my ChatGPT subscription is FOMO over whatever OpenAI is going to release next. Whatever it is, it can’t be much worse than the incredibly disappointing “GPTs” or the minor efficiency gains in the model between GPT-4 and 4 turbo. Whatever comes next, if they don’t improve the actual guts of the ChatGPT application, I don’t think I can stay loyal to the OG of GenAI much longer. As a replacement browser, it was useless. It doesn’t have the capacity to load pages within the app, so following links was outside the bounds of my little experiment because every time I clicked something, it pulled Chrome or Safari out of the drawer where I’d hidden them away.

At this point in time, Chat GPT is about as useful as it was when it was released in November 2022.

Perplexity

Perplexity is really interesting. It is not a chatbot in the same sense as the other AI applications. It doesn’t pretend to care what you think of it. It’s not nearly as anthropomorphised as ChatGPT, Copilot, Gemini, or their Claude from Anthropic. And even though using the “copilot” (no relation) mode in Perplexity Pro makes the answers more verbose, but it still feels fundamentally like a search result. The ability to refine the target sources is more academic and also welcomed.

I found myself over the 30 days using Perplexity more and more for just general day-to-day trivia questions like, “Who’s that guy in that film with the thing from the 80s, or maybe the 90s? You know, the one with the music like…” Perplexity has also picked up a massive amount of funding from investors like Jeff Bezos, so we can anticipate it’s not going anywhere anytime soon. For simple searches, Perplexity was a perfectly acceptable replacement for the Chrome browser.

Google Gemini

Everyone I know who’s used GenAI has a love-hate relationship with Google’s products. It’s almost like we all want them to be good. We all expect them to be good, given Google’s near-monopoly status, but because of their more conservative approach to safety features and guardrails, even the ones which sometimes don’t work, Google’s models have just been a disappointment.

Unfortunately, Gemini is no exception. You would think a model released by the world’s biggest search engine could handle things like, I don’t know, search. Frequently over the course of the 30 days, I tried to Google something with Gemini, and it just pretended that it was thinking and then obfuscated, gave me messages like, “This is a complicated task, and I’ll do it in the background and return the results later,” which of course it never did, because that sentence is pure bullshit.

Once or twice, Gemini was useful for specific searches, such as when I asked for a chronological discography of every album that blink-182 and its individual members had made, including side projects and EPs (don’t ask). It was also quite good at coding tasks in python, though no better than other similarly powerful models.

Gemini will ultimately be built into Google Assistant, just like Copilot will be everywhere, but it needs to significantly improve before it’s even just as useful as Google Search.

Special Mention: Claude

Towards the end of my experiment, Anthropic released Claude 3 in three flavours. I was able to use the most powerful model, Claude 3 Opus, for a number of tasks, including building an entire application from scratch.

But for this experiment, it was not useful. Like Gemini, Claude doesn’t have a dedicated browser app. I tried the shortcut method, but frankly, using Claude on a mobile browser is a terrible experience. The user interface is clunky, the text box frequently disappears, and the most powerful model, Claude Opus, does not have an internet connection and has a data set cutoff point of August 2023. Within those limits, it’s not a useful search engine.

It’s very, very capable of writing code. And at least in the last couple of weeks, it’s been an excellent desktop GenAI. But as a mobile tool, it’s not there yet.

Practical AI Strategies is available now from Amba Press

The verdict

Honestly, I didn’t manage to go 30 days without using a mobile internet definitely cheated a few times when Chrome pinged out of the drawer that I’d put it in and I just carried on scrolling. It wasn’t a particularly fun experiment, either. At times, it was very frustrating for example when I just wanted to see something simple, like a local cafe’s opening times or menu. There were some web apps that I couldn’t use and some features of usual apps which required two-factor authentication that made entirely avoiding browsers impossible.

Overall, though, I was actually surprised by how much I could get done using generative AI applications. Between the four apps, there were strengths and weaknesses, but they got the job mostly done:

• Copilot was the most versatile, and the most browser-like, but its safety features made it frustrating.
• Perplexity was the most competent with search, as long as you understand that search is really all it’s designed for.
• Chat GPT was hopeless at everything.
• Gemini was only marginally better, mostly because it looks pretty.

Chrome and Safari have come back out of the desk drawer and resumed their normal positions on my phone. But with a month of no social media notifications or one-touch access to mobile clickbait, I’m not putting those apps back on my phone any time soon.

I anticipate there will come a day, probably very soon, when GenAI apps and social media merge into one grotesque hunchbacked addiction machine. For now, at least, we can get away with using some of these technologies as a replacement to the hopelessly addicting apps we’ve been chewing on for the past decade.

If you have questions or comments, or you’d like to get in touch to discuss GenAI consulting and professional development, use the form below: