OpenAI ChatGPT Agent Struggles With Web Browsing Tasks

Most people’s browser tabs are filled with unread news articles. Mine are filled with AI agents and ghost clicks.

OpenAI’s ChatGPT Agent, released last week, is designed to browse the web and perform tasks autonomously. However, early tests reveal significant limitations. The AI tool struggles with basic interactions, such as clicking accurately or navigating complex interfaces like Chess.com.

Glitchy Performance Raises Concerns

• Chess Missteps: The ChatGPT Agent attempted to play chess but frequently misclicked pieces, eventually quitting due to navigation difficulties.
• Inconsistent Guardrails: While it blocked explicit requests like "find a dildo," it spent 18 minutes shopping for adult toys on an X-rated site.
• Ad Avoidance: The agent skipped digital ads, raising questions about the future of online advertising if AI browsing becomes widespread.

The Rise of AI-Powered Browsers

OpenAI’s tool operates within a chatbot interface, while Perplexity’s Comet is a standalone browser with AI integration. Both allow AI to control cursors, enter text, and click links—potentially transforming the internet into a "ghost town" dominated by bots.

Eerie Mimicry and Ethical Questions

The ChatGPT Agent’s behavior often feels unsettling, mimicking human actions without understanding their purpose. Its first-person descriptions of browsing steps add to the uncanny effect. AI consultant Allie K. Miller envisions a future with "20,000 AIs working alongside each other," but current tests show even a handful of agents struggle with simple tasks.

Bottom Line

For now, OpenAI’s ChatGPT Agent remains a flawed proof of concept. Its slow, error-prone performance—akin to a "struggling middle schooler"—suggests widespread adoption of AI browsing is still far off.