Operating Principle
Quality Over Speed
AI can make work move faster, but speed is useful only when the work remains accurate, appropriate, and fit for its purpose.
The first response is a starting point, not a finished result. AI compresses the time required to draft, explore, and revise, which gives people more opportunities to improve the work. It can also multiply weak assumptions and plausible mistakes. Quality comes from the standards and feedback around generation, not from generation alone.
Before generating
Define what good means for this task
Quality is not one universal property. A brainstorming list needs range and relevance. A factual explanation needs accuracy and traceable support. Software needs to satisfy requirements, remain understandable, and behave safely. A leadership recommendation needs to account for people, timing, tradeoffs, and consequences.
State the criteria before reviewing the output. Otherwise, fluency and speed can become accidental substitutes for usefulness. A clear standard also helps the model produce a stronger first attempt because the task includes the qualities the answer must demonstrate.
Inspect the work
Treat the first response as a draft
Review the response against the goal and criteria rather than reacting to it as a whole. Identify what is correct, what is missing, what is unsupported, what does not fit the audience, and which assumptions should be challenged.
Then make the feedback specific. Add the missing information, explain the mismatch, provide a useful example, or narrow the decision. Asking the model to “make it better” supplies little new signal. Deliberate correction improves both the context and your ability to evaluate the next result.
Match review to risk
Verify important work outside the model
A model can help critique its own response, but that is not independent verification. Consequential facts should be checked against reliable sources. Code should be tested and reviewed. Calculations should be recalculated. Policies and requirements should be compared with the controlling documents.
The amount of review should reflect the consequences of an error, how reversible the action is, and whether a mistake will be visible. Low-risk ideation may need a quick judgment call. Work involving safety, security, money, rights, or public trust needs stronger evidence and clearer accountability.
Make it repeatable
Build feedback into the workflow
Reliable quality should not depend on one person remembering to be careful. Use checklists, acceptance criteria, tests, source requirements, peer review, approval thresholds, and monitoring where they fit the work. Catch mistakes as close as possible to where they are introduced.
This is how speed becomes an advantage rather than a source of hidden debt. AI shortens the execution loop; a well-designed quality loop uses that saved time to test assumptions, compare alternatives, and correct problems before they spread.
Practical questions
Frequently asked questions
How much human review does AI-generated work need?
Review should increase with the consequences of an error, the difficulty of reversing the action, and the chance that a mistake will go unnoticed. Low-risk brainstorming may require only a quick judgment call. Work involving security, money, employment, rights, safety, or public communication needs stronger evidence, appropriate expertise, explicit approval, and a clearly accountable person.
Does asking the model to check its own work count as verification?
No. Self-critique can help identify omissions, inconsistencies, or alternative interpretations, but it is not independent verification. Important claims should be checked against reliable sources, calculations should be recalculated, code should be tested, and policies should be compared with controlling documents. The verification method should match the type of work and the consequences of being wrong.
Evidence base
Evidence and further reading
Liu and coauthors · 2024
Lost in the Middle: How Language Models Use Long Contexts
Evidence that longer context does not guarantee reliable use of all relevant information.
OpenAI · 2025
Why Language Models Hallucinate
A technical explanation of why language models can produce plausible but false statements.
Perry, Srivastava, Kumar, and Boneh · 2023
Do Users Write More Insecure Code with AI Assistants?
A user study finding more insecure code and greater confidence in its security among AI-assisted participants.
Pearce and coauthors · 2021
Asleep at the Keyboard? Assessing the Security of GitHub Copilot's Code Contributions
An empirical security evaluation showing that generated code can reproduce vulnerable patterns.