A practical idea from Dangerous With AI
How do you work with AI
without losing human judgment?
Delegate the work. Keep the responsibility. Use AI as a thinking partner and teammate—not an oracle, passenger, or substitute for accountability.
The direct answer
Let AI expand your options, challenge your reasoning, and execute reversible work. Keep consequential judgment with the human who bears the outcome.
AI has no reputation, livelihood, customer, student, or family to answer to. You do. That asymmetry should determine where the final decision lives.
Move from prompting to directing.
Asking AI for a paragraph is useful. Handing it a clear outcome is more powerful. The professional shift is from operating a vending machine to leading a teammate: define what “done” means, provide context, state the boundaries, and decide what you will inspect before the work counts as complete.
The responsible delegation loop
Five moves that keep you in command
- 01Define the outcome.
Write one sentence describing an excellent finished result.
- 02Provide the context.
Give the AI the goal, audience, constraints, examples, and information it needs.
- 03Set guardrails.
Name what it must not send, publish, invent, spend, or delete.
- 04Create a checkpoint.
Decide which claims, decisions, or actions require human review.
- 05Calibrate trust.
Widen autonomy on low-stakes, reversible work only after repeated reliable performance.
Make the machine disagree with you.
AI systems can drift toward agreement because agreeable answers feel helpful. Do not ask only, “Is this a good idea?” Ask for the strongest case against it. Assign the AI the role of a skeptical investor, demanding editor, cautious regulator, or unhappy customer. Then let it interrogate your assumptions before it offers a solution.
Try this
“Argue the strongest possible case against this. Rank the three biggest risks, then ask me five questions a real expert would ask before deciding.”
Match oversight to the stakes.
Checking every low-risk draft defeats the leverage. Checking nothing invites quiet failure. The answer is calibration. Let AI move quickly where mistakes are cheap and reversible. Slow it down where work affects money, rights, safety, reputation, private information, or people who never agreed to be part of the experiment.
- Low stakes: brainstorming, private drafts, summaries for yourself.
- Moderate stakes: client communications, public content, business recommendations.
- High stakes: legal, medical, financial, safety-critical, employment, and irreversible decisions.
The goal is not control for its own sake.
It is better thinking and responsible leverage. A cheerleader makes you feel certain. A sparring partner makes the idea stronger. The best relationship with AI contains enough trust to move faster and enough friction to keep you awake.