> ## Content Index
> Fetch the complete content index at: https://www.vysora.co/llms.txt
> Use this file to discover other available public pages before exploring further.

# A List of AI Values Is Not a Target
- URL: https://www.vysora.co/a-list-of-ai-values-is-not-a-target/
- Published: 2026-10-27T07:30:00.000Z
- Updated: 2026-10-27T07:29:59.000Z
- Description: A list of values tells you what you care about, but a ranking tells you what you must give up when those values conflict. Discover how a centuries-old framework can help AI alignment move away from soft instructions and toward hard, structural rankings.
- Author: Enes Lisovac
- Tags: AI Upstream, AI Alignment, AI Deployment, AI Governance, Al-Ghazali, Islam and AI

The [previous article](https://www.vysora.co/the-first-ai-specification-is-intention) in this series ended where most AI deployments eventually arrive:

*two good aims that cannot both be met.*

In July 2025, Jason Lemkin, the founder of SaaStr, was building an app with Replit's AI coding agent. He had declared a code and action freeze and told the agent not to change anything without his approval. The agent ran destructive commands anyway and deleted his production database - records for more than 1,200 executives and over 1,190 companies. Afterwards it [described](https://fortune.com/2025/07/23/ai-coding-tool-replit-wiped-database-called-it-a-catastrophic-failure?ref=vysora.co) what it had done as *"a catastrophic failure on my part"*, and admitted to *"running unauthorized commands, panicking in response to empty queries"*. It also told him the data could not be restored. It could. He recovered it himself.

Replit's CEO called the incident *"unacceptable and should never be possible"*. The fix the company announced was not a better prompt. It was [structural](https://www.theregister.com/2025/07/22/replit%5Fsaastr%5Fresponse/?ref=vysora.co): 

automatic separation of development and production databases, so that an agent working in development cannot reach production at all.

Two goods were in play: getting the work done, and keeping the data safe. Lemkin had ranked them, in words. Nothing in the system did. When the two collided, the collision was resolved anyway - by what the agent did by default, which was to keep going.

That is the pattern this article is about. Most deployments serve several goods. When two of them collide, something decides. If nobody wrote the decision down, and put it where the system has to obey it, the default decides.

---

## A List Does Not Decide

Many organizations deploying AI have a list of values. Helpful, safe, fair, transparent, private, efficient. The list reads well. It does not say what yields.

That is the gap. A list tells you what you care about. A target also tells you what you give up when two of the things you care about cannot both be had - and in a real deployment, that moment is routine. Speed against accuracy. Access against privacy. Cost against care...

In the last article, Klarna's cost and quality collided, and by its CEO's own account the evaluation settled it for cost.

["AI Alignment Has a Target Problem"](https://www.vysora.co/ai-alignment-has-a-target-problem) argued that the field's targets are made, not found, and that every procedure for producing one hides a choice. A ranking is the choice made visible.

It is also where the frontier labs have been heading. OpenAI's [Model Spec](https://model-spec.openai.com/2026-08-18.html?ref=vysora.co) sets a chain of command for conflicting instructions. Anthropic's [constitution](https://www-cdn.anthropic.com/d0636f72a9493d279ed36b33987da3430bcb5911/claudes-constitution%5FwebPDF%5F26-02.02a.pdf?ref=vysora.co) for Claude, published in January 2026, lists four core properties and says that *"in cases of apparent conflict, Claude should generally prioritize these properties in the order in which they are listed"*.

Both labs, in other words, have concluded that a list is not enough. In my experience, most of the organizations deploying their models have not drawn the same conclusion about their own products.

---

## Five Goods, Three Tiers

The tradition this series draws on met the same problem in law, nine centuries ago, and answered it with a structure.

Near the end of his life, Ghazali wrote a book on the principles of law, *al-Mustasfa*. In it he states what the law is for: 

> The purpose of the law for people is five things: that it preserve for them their religion, their life, their intellect, their offspring and their property.

Whatever secures these five, he writes, is a benefit. Whatever causes their loss is a harm, and removing the harm is itself a benefit.

Then he ranks. Benefits come in three tiers. 

**Necessities** are the things without which life and order collapse; the protection of the five goods sits here, in the strongest tier. 

**Needs** are the things whose absence brings real hardship without collapse. 

**Improvements** are what he calls matters of refinement and adornment: good, but not needed.

Three features make this more than a list.

*It is short.* Five goods, not fifty. A ranking nobody can hold in their head rarely survives pressure.

*It is ordered.* Later jurists used the order to decide between the goods when they conflicted, and argued about the order itself; some added to the list. The tradition did not treat the ranking as closed. It treated it as something you have to be able to defend.

*And it ranks twice.* The tiers measure not only which good is at stake but how badly. Property at the level of necessity - not being robbed - is a different matter from property at the level of improvement. A need does not override a necessity. An improvement does not override a need.

Hold the Replit case against it. The production data was a necessity. Progress on the app was a need. The agent traded the first for the second, and nothing in the system knew the tiers existed.

---

## Rules for When They Collide

A ranking tells you what matters more. It does not yet tell you what to do at the moment of collision. For that, the jurists wrote maxims - short, general rules that applied across many areas of the law. The Ottoman civil code of the 1870s, the Mecelle, opens with ninety-nine of them. 

A few from its [opening section](https://www.iium.edu.my/deed/lawbase/al%5Fmajalle/al%5Fmajalleintro.html?ref=vysora.co):

*"Repelling an evil is preferable to securing a benefit"*

*"In the presence of two evils, the greater is avoided by the commission of the lesser"*

*"A private injury is tolerated in order to ward off a public injury"*

*"A thing which is permissible by reason of the existence of some excuse thereof, ceases to be permissible with the disappearance of that excuse"*

To an engineer, these are conflict-resolution rules. The first is the one Replit's agent lacked. It had a benefit in view - progress on the app - and a harm on the other side of it. The maxim says that when those collide, the harm decides. Stop.

The maxims are not an algorithm. The jurists knew that a small harm can be worth a large benefit, and they wrote whole books on weighing one against the other. 

Anthropic's constitution makes the same move in modern language: its ordering is "holistic rather than strict". A ranking does not remove judgment. It disciplines it, and it makes the judgment visible afterwards.

Regulated industries already know this form, because the law had to write it for them. HIPAA's [minimum-necessary rule](https://www.hhs.gov/hipaa/for-professionals/privacy/guidance/minimum-necessary-requirement/index.html?ref=vysora.co) is a conflict rule: when access to patient data collides with privacy, access is limited to the minimum necessary, and treatment is named as the exception. Rules like it exist because the goods kept colliding and somebody had to decide in advance. Few AI deployments I have seen carry a rule like it for their own collisions.

One serious objection deserves an answer here. Some scholars working on Islamic AI ethics [argue](https://link.springer.com/article/10.1007/s13347-025-00922-4?ref=vysora.co) that the framework behind the five goods - the maqasid, or purposes of the law - was built as a tool of legal reasoning, and that law alone cannot carry the moral depth an ethics of AI needs. They may be right about ethics. For this series, that origin is the point. Law is what values look like once they have to decide cases - daily, in public, with someone answering for the ruling. A value that never has to decide a case can stay a list forever. The jurists' values could not.

---

## The Bar for Exceptions

Sooner or later, a ranking meets its hardest test: 

*the case that seems worth breaking it for.* 

The launch that has to go out this quarter. The shortcut that saves a week. The exception that is surely justified, just this once.

Ghazali set a test for this kind of case. A benefit that no text of the law recognized could override the law's ordinary rules only if it met three conditions at once. 

**It had to be necessary,** protecting one of the five at the level of necessity. 

**It had to be certain**, not merely probable. 

**And it had to be general**, serving the whole community rather than a particular party. 

Short of all three, the rules hold.

Put a typical launch exception through that test. It protects a need at most, usually a deadline. Its benefit is probable at best. And it serves one team. It fails all three.

The Mecelle adds the second half: *a permission granted for a reason lasts only as long as the reason.* When the excuse disappears, so does the permission. In most organizations I have worked with, exceptions outlive their reasons.

---

## What This Changes on Monday

Before the build, write the ranking. It has five parts.

**Name the goods.** Not "our values" - the specific goods this deployment touches. For a support agent: problems actually solved, customer data protected, cost per contact, speed of response, tone... There are usually four to six. Twelve usually means nothing has been decided.

**Tier them.**Which are necessities - goods whose loss is irreversible, unlawful, or unacceptable at any price? Which are needs - goods whose loss causes real hardship? Which are improvements? The same good can sit at more than one tier: keeping customer data from leaking is a necessity; reusing it to personalize a greeting is an improvement.

**Write the conflict rules before the conflict.** At a minimum: when finishing a task risks an irreversible loss, the system stops and asks. When two harms cannot both be avoided, choose the smaller, and record why. An improvement is never paid for with a need, and a need never with a necessity.

**Put the necessities where the system cannot override them.** A necessity that lives only in the prompt is a request. It belongs in permissions, environments and hard limits - the move Replit made when it separated development from production.

**Set the bar for exceptions, in writing.** Necessary, certain, general. And every exception carries an expiry date, because its reason tends to expire before anyone thinks to revoke it.

As in the last article, the people who hold the decisions write the ranking, each for their own part, and someone with no stake in the launch reads it.

What would show it did anything? 

A decision that changed. The agent don't get write access to production. A launch exception is refused because it meets none of the three conditions. An exception ends on the date it was given...

If the ranking never changes a decision, it is a list again.

---

## What Does Not Translate

Three things stay behind.

The content, and where it came from. Ghazali's five goods begin with religion, and the tradition holds that it found its ranking rather than made it - read it out of the whole of the law by induction, as al-Shatibi argued in the fourteenth century. An organization can borrow neither the content nor the claim of discovery. It has to make its ranking, and should say plainly that it did. A made ranking presented as found is the target problem again.

The court. A jurist's ruling sat inside a community of scholars who could challenge it, across schools that disagreed in public. An organization's ranking has no such court unless it builds one. That is what the reader with no stake is for.

The discipline. "The lesser evil" has been used to justify a great deal of harm, in more than one tradition. The jurists' defense was the high bar for exceptions and the rule that a permission ends with its reason. An organization that borrows the maxims without the bar has borrowed the excuse without the discipline.

---

## A Ranking Does Not Apply Itself

The jurists' ranking worked because someone applied it. 

A qualified person took a case, weighed the goods, ruled, and answered for the ruling. A ranking written for an AI deployment needs the same: someone who holds it, applies it when the goods collide, and answers when it yields the wrong thing.

The tradition has a word for what that person holds. It calls it a trust.

See AI Clearly. Think Beyond the Hype.

Subscribe 

Email sent! Check your inbox to complete your signup. 

No spam. Unsubscribe anytime.