I Won't Use AI Web Browsers Until They Fix These 3 Fatal Flaws
You can scarcely browse the internet these days without hearing about AI web browsers, be it Copilot Mode in Edge, Gemini in Google Chrome, OpenAI’s Atlas, Opera’s Neon, or Perplexity’s Comet, among others. I've tried most of them, but I'm not yet ready to switch to any of them due to their dubious value, privacy issues, and widespread performance problems. These browsers will undoubtedly improve over time, but here are the reasons you're better off avoiding them for now.
- AI Agents Continue to Underwhelm
All the AI browsers I've tried integrate AI assistants, which are essentially chatbots you can open at any time and on any web page with a click. Most can also automate tasks, generate media, and translate content, among other things. Of course, these features just save you the trouble of tabbing over to your favorite chatbot.
The standout feature of AI web browsers is supposed to be their agents, which come in two flavors: in-browser and virtual. In-browser agents take control of your browser to carry out whatever task you give them, while virtual agents do the same in dedicated browser instances separate from yours. But what can you ask an agent to do? A simple example would be telling it to find an intriguing roast chicken recipe and then add those ingredients to your Instacart. The premise sounds great, but I've yet to use an agent that delivers on it in practice.
Regardless of the AI browser you use, agents get lost in advertisements and CAPTCHAs, stall out, or just take an inconveniently long time to do something. Even when they work and don’t take forever, there’s still something missing. For example, I've seen agents Google a recipe, pick the first result, and add those ingredients to my cart. But what if that first recipe isn't the best one, or if you already have some of the ingredients on hand? Or, as I've also seen happen, what if an agent adds ingredients not called for in the recipe? To mitigate those issues, you must babysit the agent, which defeats the purpose of them doing tasks for you in the first place.
This sort of issue feels like it should be temporary. After all, AI has improved significantly in an extremely short time, so why shouldn't agents? The problem is that it’s near impossible to train an agent on every action you could take on every website in existence. But that’s not all, because AI agents aren’t just competing against other agents; they’re competing against native AI solutions that are popping up on pretty much every site and in every service.
For example, in my testing, AI agents struggle to handle prompts related to Google Sheets. The developers of an AI web browser could certainly optimize their agent to work better with Google Sheets (and other popular apps), but that seems like a waste of effort when Sheets’ Gemini integration can do what you wanted in the first place. Expecting a third-party browser to match the functionality of native AI features is unrealistic.
- Your Web Browsing Isn’t Private
Even if all the AI features of these browsers worked perfectly, you would still have to deal with their inherent privacy issues.
An AI web browser, fundamentally, needs access to your web browsing data. An assistant can’t answer questions about a particular tab you have open without access to said tab, and an AI agent can’t take control of your browser without access to the data within it. Moreover, if what you ask an agent to do requires you to sign into an account, it needs your password.
Aside from collecting an enormous amount of your data, some, including OpenAI’s Atlas, even use it to train their underlying models by default. Furthermore, many AI web browsers don’t bake in proper safeguards. For example, Gemini in Chrome can see and respond to what’s in any tabs you share with it, including a password manager tab. All of this isn’t to mention the unique vulnerabilities of AI web browsers, such as prompt injection attacks, in which a browser fails to distinguish between malicious code and regular content.
Even AI web browsers with relatively good privacy policies, such as Opera Neon’s, still demand you hand over an ample amount of your data. Data leaks and hacks happen all the time, so how comfortable are you with putting a trove of valuable personal information in the hands of a single company?
- They Aren’t Worth Any Amount of Money
In this modern age, it’s easy to simply ignore the privacy concerns of AI browsers. However, you may still care about how they affect your wallet.
Some AI web browsers are largely free, such as Comet, which monetizes itself with a $5-per-month journalism subscription. Other browsers are less generous. OpenAI’s Atlas, for example, requires a premium ChatGPT subscription to use agent functionality, which is arguably its most notable feature. Some browsers take things even further, such as Opera's Neon, which costs $20 per month and isn't part of a larger chatbot service.
Then, you need to weigh the price for an AI web browser against the limitations of its AI agents and the privacy concerns. I just can’t imagine spending $20 per month for what any of the AI web browsers on the market offer. Even if they get better, that’s a high price to pay for yet another subscription service. Would you rather have Netflix or an AI web browser? Probably Netflix.
Don't Go Out Of Your Way to Adopt an AI Web Browser
If you already pay for an AI service that offers a web browser or regularly use a service like Perplexity that offers most of its AI web browsing features for free, there's no harm in trying out this functionality. However, none of these browsers is worth paying for on its own. Agentic features, in particular, need more time in the oven, assuming they can eventually finish cooking at all.
Realistically, AI agents will never be perfect, and there’s little indication that AI web browsers will get significantly cheaper or more private anytime soon. For my money and time, I’d much rather use a traditional browser, such as Chrome, Edge, or Firefox, which continue to add AI features you can try out for free. And if you don’t want to use their AI features, you just don’t have to.
Disclosure: Ziff Davis, PCMag's parent company, filed a lawsuit against OpenAI in April 2025, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.
