Can decide
Relevant information, time to consider it, and a way to verify the reason.
Change can quietly become difficult to reverse
Early in adoption, people can still do the work. They can examine an AI decision and return to the previous method if the system fails.
Assistance — AI supports human work
Reduced review — automation becomes the default and human checks decline
Capability loss — staffing, skills, manual procedures, and legacy systems diminish
Dependence — normal operations become hard to maintain without AI
A problem is found — formal authority remains, but stopping the AI now stops the function itself
Being able to stop at deployment is not the same as being able to stop five years later.
A human in the loop is not enough
A person may sit in front of the screen. There may be an approval button and a written power to stop. But if that person lacks time, cannot understand the basis of the AI output, and has no workable alternative after stopping, can they genuinely say “no”?
Human-in-the-loop is not the same as effective human refusal.
Can stop
Clear authority and the technical ability to pause, refuse, or redirect.
Can withstand stopping
Manual operations, alternative providers, legacy systems, and safe restart routes remain available.
The problem is not only AI error
The problem remains even when AI is highly accurate. In society, accuracy is not the only consideration. AI also brings speed, scale, dependence, distributed responsibility, and consequences that may be difficult to undo.
Hiring, lending, insurance, healthcare, welfare, public administration, education, contracts, transfers of funds, large-scale publication, infrastructure, and physical execution may all be difficult to repair after the result has spread.
Keeping people part of the future
Even as AI becomes faster and more capable, it does not follow that human beings must become unnecessary.
What matters is not only preserving outputs. It is preserving a future in which people can learn, grow, create, care, build relationships, make judgments, and take responsibility.
Each field below shows what may be lost, the boundary beyond which return becomes difficult, concrete friction that can preserve options, and a LUMINA-30 question for testing whether the protection is real.
Societal friction is not about preserving human hardship. It is about using AI while preserving room for people to grow, participate, contribute, and live lives they can still call their own.
What societal friction means here
Societal friction does not mean slowing AI indiscriminately or adding meetings and signatures to every use.
It means preserving room for people to check, stop, reconsider, and choose another path before a consequence becomes difficult to reverse.
The question is not whether to use AI. It is whether we can use AI while preserving a state in which a human “no” can still matter.
Bad friction
- A human signature with no realistic choice
- An approval that cannot actually overturn the outcome
- An appeal process designed to exhaust the user
- Review with no identifiable responsibility owner
- “Human review” without independent verification
Good friction
- Enough information and decision time
- Access to a person and genuine re-review
- Clear stop authority and technical stop capability
- Fallback, manual operations, and safe restart
- Evidence, logs, exit, and portability
Make friction proportionate to irreversibility
Summarizing an internal document is not the same as signing a contract. Drafting a message is not the same as publishing it to a million people. A recommendation is not the same as an irreversible transaction.
The more difficult the consequence is to undo, the more robust the remaining human options should be. That is proportionate friction.
Small friction beforehand can prevent enormous repair afterward.
Institutions can preserve room to return
Societal friction does not always require a new regulation. Existing institutional and operational mechanisms can preserve future choices.
Reconsider before finality
A short hold for high-impact decisions, human re-review, and a usable appeal route.
Start within limits
Staged rollout, bounded deployment, stop conditions, periodic review, and exit conditions.
Preserve what follows a stop
Manual operations, alternative providers, legacy systems, data portability, exit capacity, and safe restart.
Name the stopping actor
Decide in advance who can stop what, under which conditions, and on what evidence.
A good institution may not be one that stops AI. It may be one in which “stop” remains a real option when it becomes necessary.
Public procurement can preserve future choices
Public bodies are not only regulators; they are also buyers of AI. Procurement can examine stop capability, human override, logs and evidence, alternative operations, data portability, provider exit, and notification of serious events. This can preserve room to return at the contract stage, even without changing the law.
These are translation examples, not official policy or procurement requirements demanded by LUMINA-30.
The same question looks different from different positions
Members of the public
Being able to reach a person when an AI-assisted decision does not make sense.
Frontline workers
Having the time, information, and real capacity to stop when something appears wrong.
Leaders
Gaining efficiency without making the entire organization unable to function when AI stops.
Developers
Designing the last safely reversible point before publication, contracting, money transfer, or physical execution.
Audit and governance
Looking beyond a record that “a human approved” to evidence that the person could actually refuse.
Government and public bodies
Translating the issue into appeals, staged rollout, procurement, stop authority, continuity, and periodic review.
Research
Asking whether human factors, resilience, assurance, safety, governance, or other established work already covers the problem, or whether another observation unit adds value.
If the problem is expressed as one question
Before irreversible AI escalation, could humans still effectively refuse, stop, verify, or redirect?
LUMINA-30 does not determine whether AI is safe from this question alone. It asks whether safety, assurance, governance, audit, and risk management have overlooked a critical boundary: does a human “no” still have practical effect?
LUMINA-30 does not assume that this problem is novel. If established research or mature practice already addresses it sufficiently, that is also an important conclusion.
One question can lead to different implementations
In one organization it may become a stop procedure. In a company it may become a procurement term or contract clause. In public administration it may become an appeal or human re-review. In a development team it may become a check before irreversible execution. In a research group it may become a comparison with established work.
LUMINA-30 does not ask these implementations to become one system. It offers a common question that different societies and organizations can use to design the room they need to return.
Preserving a human role does not mean returning to the past. It means moving forward with AI while people remain participants in shaping the future.
One thing is enough to take away
The next time you make an important decision about AI, ask:
“After we move beyond this point, will a human still be able to say no?”
That question may reveal what your field needs to preserve.