From the podcast
Rules of engagement.
36 concrete rules for working with AI, collected from twelve conversations. Most of them began as an attempt to answer one of the open questions at the end of an essay – a guest saying: here is what I would do on Monday.
Learning and expertise
AI takes over the work through which beginners used to learn. These rules put that work back on purpose.
- 01
Protect the boring years
Let beginners do things the slow way – deliberately, even when the machine would be faster. That is where judgement grows.
Krzysztof in Speed as a promise · Time & AI
- 02
One day a week without AI
Juniors sketch and write without any AI one day a week. They will complain. Keep it anyway.
Noa in The machine that always agrees · Ideation & AI
- 03
Shadow a stranger
Once a month, every junior spends a day following someone from a completely different role – an engineer, a nurse, a shop owner. No AI that day.
Noa in Centaurs, cyborgs and the human in the loop · Ideation & AI
- 04
Bring back the oral exam
Ask people to explain their own work without a screen – and then ask "why?" three times. Text can be generated; a second follow-up question in a live conversation cannot.
Krzysztof and Yannis in Snow on demand · Time & AI
- 05
Rehearse the outage
Once a quarter, switch the system off for one night, like a fire drill, and see who can still run the work by hand.
Claire in Who holds the stopwatch? · Organizations & AI
- 06
Promote the people who make others good
When delivering becomes cheap, the scarce person is the one who grows other people. Make teaching count in careers.
Niels in Who holds the stopwatch? · Organizations & AI
Decisions and responsibility
The machine absorbs the uncertainty; somebody still has to carry the decision. These rules say who, and how.
- 07
Someone must sign
Every decision has a named person behind it – and whoever signs must be able to see and change the threshold. If they cannot move the number, they should not have to sign.
Claire in Who holds the stopwatch? · Organizations & AI
- 08
Write down the reason
Let the algorithm propose, and let the human write down why they follow it or not – every time. Then judgement stays exercised, and the answer to "why?" is a reason, not "the system".
Francesca in Snow on demand · Time & AI
- 09
A face for decisions about people
Wherever the person affected has the right to look someone in the eye – a sentence, a school report, an asylum decision – a human decides. Not because humans are more accurate, but because a judgment you cannot argue with is not a judgment.
Krzysztof in Snow on demand · Time & AI
- 10
Check a sample by hand
A responsible person checks a sample of the machine's output by hand, like an auditor – not everything, but enough to know.
Niels in The org chart in the drawer · Organizations & AI
- 11
Version the premises
Treat the rules the AI works by like a rulebook with version numbers: every version has an owner, a date and a reason. If nobody can say who changed a premise and why, nobody decided it.
Niels in Who signs? · Organizations & AI
- 12
Keep an agent register
Every AI agent gets an entry: what it may do, what it may spend, which data it can see, who is responsible for it, and when it was last checked.
Niels in Who signs? · Organizations & AI
- 13
No silent vendor updates
No model update that touches business logic without release notes in plain language – and a signature from someone inside the organization.
Claire in Who signs? · Organizations & AI
- 14
Reasons and time for the human in the loop
The person approving sees why the system proposed something, not just the result – and has the time to judge. Approving three hundred routes in an hour is not deciding.
Claire in Who signs? · Organizations & AI
- 15
Publish the thresholds
In public administration, thresholds and variables are policy: publish them, let parliament or an ombudsman ask for them, and give citizens a real appeal to a human who can override.
Niels in Who holds the stopwatch? · Organizations & AI
Working with the machine
Small habits for the moment you sit in front of the model – so that it widens your thinking instead of closing it.
- 16
Ask for five reasons, never for the reason
A machine makes things fit; a surprise is something that does not. Ask for five explanations, including unlikely ones – and then send someone to have coffee with the forklift driver.
Claire and Niels in The AI nobody is allowed to use · Organizations & AI
- 17
Give it roles, not the verdict
Never ask the machine "is this good?". Let it play the sceptical investor, the frustrated user, the competitor who wants you to fail – all three at once.
Noa in The machine that always agrees · Ideation & AI
- 18
Rate first, reveal later
Everyone rates the ideas alone first; only then is the AI ranking shown. Shown first, the machine's number becomes an anchor and people nod.
Noa and Yannis in A hundred ideas a minute · Ideation & AI
- 19
Provocations, not answers
Show machine ideas as something to attack: "Here is what the machine thinks. Who wants to destroy it?"
Noa in A hundred ideas a minute · Ideation & AI
- 20
Ask it to irritate you
Do not ask for the best idea. Ask for the strongest argument against your plan, or how a beekeeper would solve the problem.
Francesca in Snow on demand · Time & AI
- 21
A fire drill for trust
Every month, give people tasks where the AI is wrong on purpose and see whether they catch it. In a workshop: the facilitator secretly plants one wrong number.
Yannis and Noa in Centaurs, cyborgs and the human in the loop · Ideation & AI
- 22
Summaries for choosing, not for knowing
Use summaries to decide what to read – not instead of reading it.
Krzysztof in Snow on demand · Time & AI
- 23
Let it sleep on it
A "sleep on it" mode: describe your problem in the evening, and the machine answers with two questions instead of a solution.
Francesca in Snow on demand · Time & AI
Organization and culture
How a team or a company draws its lines – what the machine may do, what stays human, and what is never measured.
- 24
The first page is written by hand
The model never writes the first page of anything that represents you: the opening of a tender, the letter to a customer after a mistake, the first paragraph of a job ad.
Claire in The AI nobody is allowed to use · Organizations & AI
- 25
One hard line: the data
Tolerate shadow AI and learn from it – people use their own tools for good reasons. Draw the hard line where data leaves the building.
Elisabeth, Niels and Claire in The AI nobody is allowed to use · Organizations & AI
- 26
Break rules in the open
A leader who breaks an AI rule and says so sets a precedent that can change the rule. A leader who breaks it secretly creates a double standard and destroys it.
Niels and Claire in The AI nobody is allowed to use · Organizations & AI
- 27
Keep a decision log
In every project: who framed the problem, which AI outputs were picked and why, how they were changed. It sounds bureaucratic; it is the most useful document in the project.
Noa in Whose idea is it? · Ideation & AI
- 28
No client data in public tools
Briefs, source code and meeting notes from clients never go into public AI tools.
Noa in Whose idea is it? · Ideation & AI
- 29
Faster drafts, same review time
When AI makes a draft in ten minutes instead of two days, the time for review does not shrink with it.
Francesca in At the edge of control · Time & AI
- 30
A pause nobody measures
Protected time that is not reported, not evaluated, not even recorded. The moment a pause is tracked, it becomes a target.
Krzysztof in At the edge of control · Time & AI
- 31
Team numbers, no red lights
No real-time performance lights per person for managers. People see their own numbers; the team sees the team's.
Claire in Who holds the stopwatch? · Organizations & AI
Society and regulation
Technology moves faster than parliaments. These proposals try to keep deliberation without losing the pace.
- 32
Rules that expire
Parliament decides quickly on a provisional rule – with a sunset clause. After two or three years it ends automatically unless renewed after a proper debate.
Krzysztof in Out of step · Time & AI
- 33
Test zones with the regulator inside
Regulatory sandboxes – supervised test zones, which the AI Act requires every EU country to set up – where companies try a system while the regulator watches and learns.
Francesca in Out of step · Time & AI
- 34
A priority list for scarce power
When electricity is short, compute is rationed like water in a drought: weather forecasts and medical research before entertainment.
Krzysztof in Out of step · Time & AI
- 35
A collecting society for training data
Like radio and music: developers pay into a collective fund, and the money is distributed to the people whose work trained the models.
Noa in Whose idea is it? · Ideation & AI
- 36
A short right for machine-made works
Seventy years of copyright for a logo made in twenty minutes with a machine is absurd. A short right – five or ten years – would fit better.
Yannis in Whose idea is it? · Ideation & AI
The speakers of the podcast are fictional characters with AI-generated voices, and the conversations were written with AI. These rules are proposals from those conversations – starting points for your own, not advice. The essays and exposés behind each episode, with all their sources, are linked from the episode pages.