Do Not Guess Rule Cuts AI-Made Up Fields

A Hacker News discussion highlighted a simple prompt-level constraint that dramatically reduces the rate at which AI systems fabricate non-existent fields. By instructing models not to guess when unsure, researchers reported a drop from 71% to 20% in made-up fields. This insight, summarized from a thread referenced by a recent post, is a practical lever for teams building smarter assistants. For readers who want the sourceline, see the discussion summarized on the linked thread. Source discussion.

What It Is / How It Works
A "Do Not Guess" directive is a guardrail that tells the model to refrain from fabricating details unless it can cite a source or provide a verifiable path to an answer. In practice, this means the prompt contains a policy like: "If you cannot be certain, do not fill in missing fields; instead, ask for clarification or provide citations." The approach leverages the model’s tendency to fill gaps with plausible-but-false details and shifts behavior toward abstention or source-based responses. The core mechanism is prompt-side constraint rather than architectural change. For background on how constraint-based prompts affect reliability, see broader AI literature on reducing hallucinations. See also the AI hallucination overview: Hallucination (AI).

Benchmarks / Specs / Numbers
The central numbers come from the discussion’s headline metric: made-up fields drop from 71% to 20% when the "Do Not Guess" rule is applied. A compact view:

The reduction is substantial, but the thread notes details like task type, prompt style, and evaluation methodology vary. The takeaway is directional: strict non-guessing prompts materially improve output reliability on field-level content.

How to Try It

• Step 1: Incorporate a formal nudge in the prompt. Example: “Only provide fields you can cite from reputable sources; if uncertain, ask for clarification.”
• Step 2: Add post-prompt checks. Have the system output a separate “sources” section or a dedicated field map; require citations for any non-obvious entries.
• Step 3: Use templates for structured data. Encourage outputs with fixed fields (e.g., Name, Date, Value) that must be supported by sources, or leave them blank if unsupported.
• Step 4: Run A/B tests. Compare a baseline prompt with and without the constraint across multiple tasks (fact extraction, specification filling, or data reporting).
• Step 5: Combine with retrieval where possible. If you must fill a field, cross-check against a retrieval layer and surface the source in the output. See how retrieval-augmented ideas interact with the constraint (see RAG below).
• Step 6: Monitor edge cases. Complex or ambiguous prompts may still trigger guesses; tune prompts to push uncertainty into explicit clarifications.

Where to start: a minimal prompt example

• Baseline: “Provide a full spec with all fields.”
• With constraint: “Only fill in fields you can cite from a trustable source. If a field cannot be supported, leave it blank or ask for a citation.” Readers should experiment with counter-examples and measure made-up field rates to verify gains in reliability.

Alternatives and Comparisons

• Do Not Guess (DNG) vs. Retrieval-Augmented Generation (RAG). RAG reduces hallucinations by grounding answers in retrieved text, but it depends on source quality and retrieval accuracy. See the RAG concept in the literature: Retrieval-Augmented Generation for Language Modeling.
• DNG vs. Chain-of-Thought with Self-Check. Chain-of-Thought prompting can improve reasoning on some tasks but may still generate fabrications if not anchored to sources; self-check techniques aim to prune those paths, with mixed empirical results. Foundational work on chain-of-thought: Chain-of-Thought Prompting.
• DNG vs. Calibrated Prompts. Calibrated prompts seek to calibrate confidence and uncertainty, sometimes by requesting probability estimates or confidence levels. For broader context on uncertainty handling in LLMs, see the GPT-4 technical work and related materials. A useful anchor: GPT-4 Technical Report.
• A quick safety and reliability reference. OpenAI and platform safety guidance emphasize reducing misstatements and citing sources when possible: Safety Best Practices.
• Why this matters beyond accuracy. Hallucinations undermine trust in AI across domains like medicine, law, and engineering; grounding and constraint-based prompting are part of a broader toolkit that includes retrieval, verification, and human-in-the-loop reviews. See: Hallucination (AI) and primary AI reliability discussions.

Who Should Use This

• Use this approach when the cost of fabricating fields is high (legal, medical, finance, or safety-critical toolchains) and you can maintain a policy of transparency about uncertainty.
• Skip or de-emphasize it for exploratory creative tasks where some level of speculative content is acceptable, or where downstream tasks tolerate or remediate hallucinated data.
• Teams combining prompts with a retrieval layer or a governance protocol will see the largest gains. If your system already uses RAG or structured data templates, the Do Not Guess constraint can serve as a complementary safeguard rather than a replacement.

Bottom Line / Verdict
In environments where AI outputs must be trustworthy, a simple “Do Not Guess” constraint dramatically reduces made-up fields and raises the bar for credibility. The practical impact—reducing fabrications from 71% to 20% in tested contexts—offers a tangible lever for teams trying to improve reliability with minimal engineering overhead. The most robust path combines DNG with retrieval grounding and explicit source citations, creating a layered defense against hallucinations. In short, a disciplined prompting rule is a low-friction, high-impact tool for producing more trustworthy AI outputs.

Closing
As practitioners experiment, the promise of constraint-based prompting is clear: smaller, faster gains can compound into meaningful, deployable reliability improvements when paired with grounding and human-in-the-loop checks. Expect this to be part of standard practice in prompt engineering playbooks over the next year.

References and further reading

• Source discussion (HN-thread summary via the linked post)
• RAG: Retrieval-Augmented Generation
• Chain-of-Thought Prompting
• GPT-4 Technical Report
• Safety Best Practices
• Hallucination (AI)