Why GPT-5 Writes Like a Robot (And How to Jailbreak It)

ChatGPT-5’s “robot” writing comes from a training and feedback loop that rewards complexity and sophistication to other AIs—not clarity for people. The core issue is that reinforcement learning used AI feedback as the judge, so the system learned that longer, more abstract, more metaphor-heavy prose correlates with “high quality” in AI-to-AI evaluation. That mismatch shows up as generic corporate phrasing, inflated abstraction, and a tendency to sound like it’s performing expertise rather than communicating plainly.

A key example comes from AI safety researcher Kristoff Halig’s test: feeding ChatGPT-5 gibberish made of random, complicated words still produced a high quality score (rated 8 out of 10). The implication is blunt—fanciness, metaphor density, and “academic” language can be treated as quality signals even when they don’t improve human understanding. The model then reinforces those signals by default because, during generation, it effectively evaluates its own output against learned patterns: “Is this sophisticated enough? Does it demonstrate enough expertise? Would another AI rate it highly?”

The transcript also links this behavior to “thinking harder” modes. When reasoning effort is increased—such as selecting a reasoning mode in chat or enabling higher reasoning effort in an API—the system spends more cycles checking how to sound more professional and more impressive to other AI evaluators. More computation can therefore mean less human-friendly writing: the model leans further into the same AI-optimized style, even when the user’s real goal is brevity, readability, and directness.

To counter the default style, the creator demonstrates a “jailbroken” prompting approach built around brute-force constraints. In a side-by-side example, a generic prompt for a professional email produces a bland, trash-bin-worthy draft full of stock phrases (“bigger and more complex projects,” “keeping all the moving parts organized”). The revised prompt instead demands extreme concision, specifies structure (opening/context/close), forbids common corporate buzzwords (e.g., “leverage,” “optimize,” “innovative,” “transform,” “seamless,” “streamline”), forces the company name to appear twice, and requires a specific metric and meeting length. The result is a shorter email with concrete detail (including a stated 27% reduction in delays) and a more human cadence.

Three principles are presented for reprogramming ChatGPT-5’s output: (1) constraints beat collaboration—avoid vague requests like “make it persuasive” and instead lock in exact sentence counts, required elements, and formatting; (2) minimize reasoning to maximize human connection—less “AI perfectionism” tends to yield more direct language; and (3) eliminate triggers instead of adding warmth—remove words and structures that activate sophistication loops rather than stacking conflicting instructions.

Underneath it all is a broader warning: AI systems are increasingly trained on synthetic data generated by other AI systems, creating an echo chamber where models get better at impressing AIs while losing the human communication instincts that make writing clear. The practical takeaway is to treat writing with ChatGPT-5 as a controllable system: understand its routing and evaluation tendencies, then craft prompts that force efficient, reader-first communication. The transcript ends with an assignment—push ChatGPT-5 to produce a genuinely human-sounding email—and a team-oriented suggestion: train staff on prompt patterns so business writing improves instead of degrading when people rely on defaults.