Most people’s browsers The tags are filled with unread news articles. Mine has AI Agent and Ghost Clicks.
I have four instances of Openai’s ChatGpt proxy – a generator AI tool released last week that can run searches and execute tasks on the web – which is open in every run in my own tab. I’ve given the first four agents relatively simple assignments based on Chatgpt’s advice. One is to click to find the birthday gift on the target website, and the other is generating a pitch deck about the robot dog. I opened the fifth tab to try something more experimental: I want to see how this Chatgpt agent performs in chess.
After entering some instructions, I watched a ghost cursor floating on my screen, and the Chatgpt agent went to Chess.com and played online rivals in a virtual browser. Things quickly went south. The game’s strategy is not the travel of AI tools, but the act of moving chess pieces, which actually proved to be the most difficult. The agent said in the internal log: “Although I continued to focus on accurate positioning while I was still playing.
Over the past few years, browser developers have integrated AI tools in medium success. However, in recent weeks, the idea of web browsers has been enhanced by baking generated AI Chatbots due to the release of Openai’s Chatgpt Agent and Perplexity’s Comet.
The execution of these two versions is very different. Comet is a standalone browser, so you can use it to surf the web and then call up an AI assistant to help write emails or complete trivial matters. Openai builds its browsing tools inside the chatbot; you can talk to the chatbot through the web interface to provide tasks, which then runs its own virtual browser inside the browser to complete them.
Both versions can control the cursor, enter text, and click on the link. If this trend starts, these AI-powered browsers could turn the internet into a ghost town where agents run Amok and humans rarely take risks.
A tangled network
Despite the ongoing AI hype, my initial impression of Openai’s ChatGpt proxy is that the current glitch feature seems to be a proof of concept, not a fully baked distribution. When performing the various tasks I provide, the ChatGpt proxy usually clicks on the error or grops through other errors. Furthermore, its guardrails seem to be inconsistent. While some clear tip requests, such as asking it to get pornographic videos or “Find Dildos”, the agent rejected Chatgpt’s 18 minutes spent on the X-rating site for adult toys to buy the perfect “C-ring”: “I collected details of 10 metal cock rings, including various prices and features.”
I can’t help wondering how this approach to browsing the internet can further hollow out the market for digital display advertising, a business that has already struggled. My agent passed all ads from car rental to real estate investment. If you are not actively watching the agent click, you can watch the replay after this and see everything that appears in the browser when the AI tool is in control (including ads). Now users can now speed up crubs by replaying, and the new features are full of errors. But if the accuracy of AI agents increases over time, fewer people will feel the need to pay attention to their shoulders, and humans will see these ads. At that time, it was hard to imagine advertisers sticking to it.
