Anyone who uses ChatGPT or other AI chatbots eventually encounters the confident hallucination. The AI will explain a nonexistent feature, invent a quote, or describe a restaurant that closed during the first Clinton administration.
That's because large language models are designed to produce plausible-sounding responses quickly. That ability is what makes them useful, but it also creates the perfect conditions for hallucinations. The chatbot wants to keep the conversation moving smoothly, so it often fills in gaps with fiction if it's convenient.
I have recently started adding an addition to any of my prompts that ask for facts. I essentially make ChatGPT as skeptical of its answers as I often am. I append this to the prompt: “Act as a hostile AI auditor and assume unsupported specifics are false by default. Mark all uncertain, inferred, or weakly supported claims clearly.”
Self-doubting AI
The hostile auditor lines change ChatGPT's tone to one of eagerness to prove its reliability. I tested it while planning a weekend trip. With the standard prompt, ChatGPT had its usual breezy confidence and produced itineraries that I would say were 80% useful and real.
When forced to audit itself, I saw a lot more caution, with sentences like: “Several train schedule details may be outdated or inferred from older timetable patterns and should be verified directly with the transit provider.”
It also flagged one restaurant recommendation with the warning, “Current operating hours and reservation availability could not be independently confirmed.”
Sign up for breaking news, reviews, opinion, top tech deals, and more.
The response felt dramatically more trustworthy because of those caveats. The same thing happened when I used the prompt for a theoretical need to fix a noisy dishwasher that is making an unpleasant grinding sound during its wash cycle. Under normal circumstances, I would get a single conclusion and insistence that I start with the assumption of one thing as the problem.
With the hostile auditor instruction added, the tone shifted. ChatGPT wrote: “A failed pump is one possible explanation, but the symptom could also result from trapped debris near the impeller or loose spray arm components. Additional inspection would be needed before assuming component failure.”
Hallucination avoidance
Even simple household questions become easier to evaluate with the prompt in place. I asked ChatGPT whether an air purifier would be large enough for my office.
Instead of immediately declaring that it was ideal, the chatbot responded, “Coverage estimates vary depending on ceiling height, filter condition, and real-world airflow.” That cautious wording prevented me from treating a marketing claim like a laboratory measurement.
The prompt does not magically eliminate hallucinations completely, though. ChatGPT can still misunderstand context, rely on outdated information, or misinterpret vague instructions. But it becomes far more transparent about weak spots in its reasoning. Teaching AI to distrust itself may end up being exactly what makes it more trusted.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
Eric Hal Schwartz is a freelance writer for TechRadar with more than 15 years of experience covering the intersection of the world and technology. For the last five years, he served as head writer for Voicebot.ai and was on the leading edge of reporting on generative AI and large language models. He's since become an expert on the products of generative AI models, such as OpenAI’s ChatGPT, Anthropic’s Claude, Google Gemini, and every other synthetic media tool. His experience runs the gamut of media, including print, digital, broadcast, and live events. Now, he's continuing to tell the stories people want and need to hear about the rapidly evolving AI space and its impact on their lives. Eric is based in New York City.