I’d Happily Pay Double If ChatGPT 5 Actually Worked Like It Used To

The Decline of ChatGPT: User Frustrations and the Challenges of Evolving AI Products

As an early subscriber to ChatGPT, I have witnessed firsthand the rapid development and impressive capabilities of this AI language model over time. Like many users, I was excited about the potential of OpenAI’s technology and was willing to invest in its promising features. However, recent experiences with the latest iteration—referred to as ChatGPT 5—have led to growing frustration and disappointment.

Persistent Issues: Inaccuracy and Hallucinations

One of the most concerning problems with ChatGPT 5 is its propensity to hallucinate—generating plausible-sounding but false information. For instance, when uploading a PDF document and asking about specific policies, the model confidently cites rules that are not present in the document. Even after pointing out the inaccuracies, it tends to double down, providing alternative fabricated references. This behavior undermines trust and complicates its usefulness for research or informational tasks.

Deterioration in Handling Basic Tasks

Tasks that previously posed no challenge, such as editing text for grammar and clarity, now often result in overcomplication. Instead of providing concise and accurate revisions, ChatGPT 5 sometimes rewrites entire sections, introduces unintended content, or shifts the tone and intent of the original message. Notably, earlier versions like GPT-4 handled these tasks more effectively, preserving the integrity of the user’s intent.

Ineffectiveness of Prompt Optimization

Many users, including myself, have experimented with different prompting strategies—be it more specific instructions, breaking down complex tasks, or setting system prompts—in an attempt to improve response quality. Unfortunately, these efforts often fall short, as ChatGPT 5 appears unresponsive to careful guidance, seemingly filling gaps with assumptions rather than adhering to the user’s intent. It’s akin to speaking to someone only partly attentive, filling in the blanks with their own imagination.

Competitive Performance and User Frustration

In a personal comparison, I find myself running four chat interfaces side-by-side—ChatGPT, Claude, Google’s Gemini, and DeepSeek—pasting identical prompts into each. Shockingly, in most cases, ChatGPT produces the least accurate or useful responses. In fact, the free tier of DeepSeek often outperforms ChatGPT’s premium service, which is a disconcerting reality for a paid product.

The Future of AI as a Product

Despite my reservations, my desire to remain loyal to ChatGPT persists.