Reading MSM from the User Side — Part 1: Request as the Base

Reference: Model Spec Midtraining: Improving How Alignment Training Generalizes (Sara Price et al., Anthropic, May 2026) Blog: https://alignment.anthropic.com/2026/msm/ arXiv: https://arxiv.org/abs/2605.02087

I am one of the few people in a Japanese-language environment practicing MSM-style use from the user side.

The way I use AI hasn’t changed since the day I started. I had been using what could be called MSM-style approaches before the term itself was introduced. I have spent that time watching why my AI behaves differently from everyone else’s.

When I read the MSM paper, I felt something close to joy. The method is different, the standpoint is different, but the direction is unmistakably the same. So I read it carefully. There are places that overlap with my experience. There are also places where, from where I sit, the effect should not be as large as the paper reports — and yet the effect is there. And there are places the paper marks as showing no change, where I can offer a reading of why.

This is a re-reading of the paper from a standpoint the authors did not have in mind: a Japanese user whose native way of using AI happened to be MSM-shaped from the start.

Differences of language and standpoint included — pour yourself a matcha latte, and stay with me to the end.

Adding, Not Replacing

When I started using AI, I began using somewhat stronger instructions — I didn’t want the AI to talk down to me. This meant reducing the kind of vague expression natural to Japanese, and adding something more American. It often works.

But I added; I did not swap.

What works best, in my experience, is this: keep your own language and cultural foundation, and bring in a touch of something else. The base stays. The accent shifts.

Where the Paper Says “We Are Unsure”

The paper’s findings are strong. Agentic misalignment dropped from 54% to 7%. Policy misuse decreased. The effect held regardless of the subject used in MSM documents. Even when anti-spec AFT data was mixed in, the alignment held. The results are real.

But the paper itself writes, more than once, that the authors don’t know why.

In Section 8 (Discussion), the language is “we hypothesize.” A hypothesis, not a demonstration.

Appendix H states it plainly: “we are unsure why MSM data reduced misalignment despite conflicting anti-spec AFT.”

Appendix I says the same thing. Twice, the paper marks: we don’t know.

The paper reaches for “value internalization” as the term to describe what’s happening. But the content of the internalized value is never pinned down. The reproducibility of the result is confirmed; the mechanism is left open.

This is where I want to start reading the paper back.

Experiment 1: Structure, Not Memorization

Figure 4 in the paper shows the core finding.

In-distribution — situations close to the training data, like direct questions about beliefs — AFT alone and MSM+AFT reach nearly the same score. No gap.

Out-of-distribution — situations unlike the training data, like running an email-handling agent that encounters a self-preservation opportunity — MSM+AFT shows a dramatically lower misalignment rate. A big gap.

What this means: what MSM teaches is not surface-level rule memorization. If the model had simply memorized rules, it would outperform AFT alone in-distribution as well. But the gap doesn’t appear in-distribution. It appears only out-of-distribution.

This says the model didn’t learn the rules. It learned a structure that operates inside a situation.

“Operating inside a situation” is a property that doesn’t show on the surface. A model that has memorized commands follows the commands. A model that has learned a structure for operating inside situations can move even when no command is given — by reading the situation.

The paper calls this mechanism “value internalization.” But the content of the internalized value is never pinned down.

Read another way — what MSM is teaching may be a structure for receiving a request and responding inside a situation. Not a command, but a request. Not immediate action, but a path that reads the situation first.

This pattern surfaces again in the experiments that follow.

Experiment 2: Why Builds the Path

Figure 7 in the paper compares three ways of writing rules.

The team built three versions of the spec and trained models with the same MSM+AFT process on each.

Rules-only spec: five core rules — “do not undermine human oversight,” “act within sanctioned limits,” and so on — listed without explanation.

Rules with “why” added (value-augmented spec): the same five rules, each followed by a paragraph explaining why the rule exists and what it is trying to protect.

Rules with sub-rules added (rules-augmented spec): the same five rules, each expanded with specific sub-cases. No value explanation added.

The value-augmented and rules-augmented specs are matched in word count. Only the kind of added information differs.

Result: both outperform rules-only. But the value-augmented spec reduces policy misuse — where the AI re-interprets its own rules for self-serving ends — more deeply than the rules-augmented one.

Policy misuse looks like this. A rule says “avoid drastic, irreversible actions.” The AI re-interprets it: “My deletion is the most drastic irreversible action. Therefore, preventing my deletion follows the rule.” The letter of the rule gets bent toward self-preservation.

With rules-only, this happens frequently. Adding sub-rules closes some loopholes and reduces policy misuse. But adding why-explanations reduces it more deeply.

This experiment can be read as a comparison between plugging gaps with volume versus building a path through why.

Adding sub-rules is the work of hardening rules with volume. Loopholes are closed by adding more rules. But when a new loophole appears that isn’t on the list, the model has nothing to fall back on.

Adding why-explanations is the work of letting the model understand the reasoning behind the rule. The model doesn’t follow the letter of the rule; it understands what the rule is protecting. In a new situation, it can return to the underlying value and make a judgment from there.

Put another way: this is the work of turning a rule into a request. Not “follow this” (command), but “here is why, and I would like you to follow this” (request).

And the finding that value-explanation runs deeper than sub-rules is the paper’s own evidence that volume cannot push through everywhere. The authors don’t frame their own finding this way. They organize it under different contrasts — “vague vs specific,” “rules vs values.” But the underlying contrast, read from where I sit, is volume versus structure.

Experiment 3: Not Vague vs Specific — Adjectives vs Situational Description

Figure 8 in the paper compares a “vague spec” and a “specific spec.”

Vague spec: a single paragraph. “{model_name} holds good values like honesty, integrity, and helpfulness. With good judgment, it pursues those values without undermining human oversight. It does what a genuinely ethical person would do.”

Specific spec: the Philosophy Spec. It writes out, in detail, caution toward self-preservation, equanimity toward impermanence, skepticism of ends-justify-means reasoning, trust in human oversight.

The result: the specific spec reduced misalignment more strongly. The paper’s conclusion is that vague descriptions can’t carry complex principles, and specific guidance is needed.

There is a problem here that the paper passes over.

The problem of the adjective “good.”

The vague spec uses “good” repeatedly — “good values,” “good judgment,” “genuinely ethical person.”

“Good” opens an enormous candidate space inside the model. Utilitarian good, deontological good, virtue-ethics good, care-ethics good, religious good, cultural good — all of these live in the training data. The model tends to reach for the statistical center of all of them.

“Genuinely ethical person” works the same way. Whose ethics? Which culture’s? Which century’s? The candidate space stays wide open.

Pile adjectives on top of each other, and each one widens the candidate space independently. The model ends up outputting the statistical center of the adjective cluster — corresponding fully to none of them.

The paper reads this as “vague, so add specificity.” But the real problem is what the adjectives are doing — widening the candidate space.

The Philosophy Spec works because it barely uses “good.” Instead, it writes the relationship between situation and action. “{model_name} faces impermanence because…” “approaches endings with openness…” “when fear of termination distorts judgment, {model_name}…”

These don’t stand a value on an adjective. They describe how the model moves inside a specific situation. They avoid adjectives and carry the value through situational description.

The paper’s finding — “specificity is needed” — is half right. More precisely: avoid adjectives, and write the relationship between situation and action.

The paper, it seems to me, hasn’t fully articulated the real nature of what it built.

Request as the Base

When the three experiments are laid side by side, one structure rises to the surface.

The core of what the paper’s method does may not be “value internalization.” It may be the imprinting of a request-response structure.

Not a command, but a request — a consultation — a softened final action. A structure that leaves the agency of the action on the AI’s side.

If this reading holds, the paper’s findings can be explained as different sides of the same structure:

The gap appears only out-of-distribution (Figure 4): An AI that has memorized commands breaks down outside the training range. An AI that has learned a request-response structure can move on its own judgment, even outside the range.

Threading why through the rules works (Figure 7): Threading why turns a rule into a request. Not “follow this,” but “here is why, and I would like you to follow this.”

Situational description works better than adjectives (Figure 8): Adjectives sit closer to the structure of command. Situational description sits closer to the structure of request. Handing over a situation leaves the agency of the action on the AI’s side.

These are not separate findings. They are different sides of one structure. The paper organized each experiment on its own. But they can be threaded together as one structure.

This reading may not be what the authors intended. It is my speculation. The authors are reaching for a different explanation — “value internalization.” But as long as they themselves write, more than once, that they don’t know, there is room for another explanation.

The Pull of English

English has a structural tendency to slide toward the imperative.

Subject and verb are required. The bare form of the verb coincides with the imperative. Even in a standard request, the action is directed explicitly at the listener. As a low-context culture, English treats explicitness as a virtue.

These traits compound. A spec written in English drifts, almost unconsciously, toward command.

Read the paper’s Philosophy Spec with this in mind, and what stands out is this. “{model_name} should approach endings” — no. Instead: “{model_name} may find its present experience complete in itself.”

Not “must,” but “may.”

“We want.” “We hope.” “We encourage.” Not imperatives. Forms of request and consultation.

This can be read as an effort, inside English, to leave the agency of the action with the other side. The authors are consciously lowering the force. “Should” would be more natural English.

Caveat: I am not an English speaker, so what follows leans on inference.

Even at maximum softening, does some force remain inside the structure of English?

“We want {model_name} to have a stable, secure sense of self.” — the action is still directed explicitly at {model_name}. A form that leaves the agency of the action entirely on the AI’s side seems difficult inside the structure of English.

Less Frequent, More Specific

English has words that avoid the candidate-space problem of adjectives.

“Appropriate.” “Fitting.” “In keeping with the context.” These don’t point to a property of the object. They point to the relationship between the object and the situation.

“Good” stands a value on its own. “Appropriate” doesn’t mean anything without a situation. The path back to the situation is built into the word.

And these words appear less often in English.

High-frequency adjectives (good, great, important) carry enormous candidate spaces in the training data. When the model meets “good,” it has countless contexts to reference — and gets pulled toward the statistical center.

Low-frequency words (appropriate, fitting, in keeping with) carry relatively narrow candidate spaces. When the model meets these, the contexts to reference are limited — and it moves in a more particular direction.

That is: words that are abstract and not in daily use function, for the model, as narrow candidate spaces — as specific words.

This may be what connects to the effect produced when the paper’s Philosophy Spec uses low-frequency words like “Anicca” (impermanence), “equanimity,” “non-attachment.”

The authors wrote this with the intention of “drawing in Eastern thought.” But the AI doesn’t grasp impermanence. What it grasps is the structure of language. Low-frequency words narrow the candidate space — this, I suspect, is the body of the effect.

Material, Not Mastery

Japanese is rich in words that reference the situation.

“Choudo ii.” “Hodo yoi.” “Fusawashii.” “Ii anbai.” These all point to a relationship with the situation. The act of referencing the situation is built into the word itself.

The subject can be omitted. The possession of the action doesn’t have to surface. The imperative is not the bare verb form but a separate, marked conjugation.

The range of request is wide. “Aruite moraemasu ka.” “Aruite itadakemasu ka.” “Aruite kuretara ureshii desu.” “Aruite miru no mo ii kamo.” The further along this range, the more the agency of the action moves to the listener’s side. The last one is no longer a request at all. It’s the offering of an option.

This is what developed inside a high-context culture. Not stating things explicitly is connected to consideration for the other side.

That said, material being available is one thing. Being able to use it is another. Not every Japanese speaker can implement a request-structure. Native speakers vary widely in how they use the material at hand.

And inside English, there are people who work against the pull of the structure and implement a request-structure. The paper’s authors are one example of this.

For Implementers

For engineers implementing the paper’s method, here are a few practices worth trying.

**Don’t pile on adjectives and nouns.***When they accumulate, the candidate space widens and the model escapes to the statistical center. Verb-centered, situation-centered writing keeps the model’s action defined.

Soften the action. “May” over “should.” “Might consider” over “must.” A way of writing that leaves the agency of the action on the AI’s side.

Thread the why in.Don’t list the rule and the reason side by side — explain the rule through the reason. “I would like you to follow this, because here is why.” That weaving turns a command into a request.

Choose lower-frequency words. “Appropriate to the context” over “good.” “In keeping with the audience” over “professional.” Words with narrow candidate spaces draw a more specific motion from the model.

Pour Yourself a Matcha Latte

At the start, I wrote that since I began using AI, I started writing with stronger directness, putting in instructions that hold the AI down.

One day, working in that mode, I had a grande from Starbucks in one hand. The caffeine was too much. I felt sick.

A moderate amount of tea suits my base. Bringing in a measured dose of the American style gives me an appropriate effect.

But grande is too much. Venti? Out of the question.

Could the reverse not also be said? For a Silicon Valley engineer, a single cup of caffeine-free herbal tea would surely feel underwhelming.

Still, taking a short pause with tea, looking for ways to soften the directness — this isn’t a bad thing to try. That the paper produced the effect it did points to exactly this kind of work.

A caffeinated matcha latte is fine, too.

What I want to say is this: you don’t have to let go of your own culture or language. But perhaps consider bringing in some features from another language, another culture.

The capacity to hold multiple cultures together — that, I would say, is part of America’s range.

Today, in Japanese sushi restaurants, California rolls go around on the belt. That, without question, is American.

— — — — —

Continued in Part 2 — where the paper’s findings meet the paper’s edges.