Adequate Ideas

Spinoza Ethics Part II on why some knowledge self-validates, and some can't.

Spinoza defines an adequate idea as "an idea which, in so far as it is considered in itself, without relation to the object, has all the properties or intrinsic marks of a true idea."

The key phrase: without relation to the object.

Ordinary ideas are verified by comparison to their objects. You have an idea of the cat; you check whether the cat is where the idea says it is. The idea's truth is determined externally — by measurement against the world. This is the first kind of knowledge, what Spinoza calls imagination: sensory impressions, habit, symbols, all of it mediated through the body and its external interactions. Ideas produced this way are always potentially confused. They come to you as consequences without premises. You get the output, but not the derivation.

The adequate idea is different. It carries its truth in its structure. You don't need to check it against anything outside because the conditions that make it true are already present in the idea itself. "Truth is its own standard," Spinoza says in the note to Proposition XLIII. "Even as light displays both itself and darkness, so is truth a standard both of itself and of falsity."

This isn't mysticism. It has a precise mechanism.

The adequate idea is one formed from common notions — properties that are true of all things, everywhere, in the same way. Extension is common to all bodies. Motion and rest are common to all bodies. These properties don't depend on how your specific body was affected at a specific time. They're not conditioned by the fortuitous play of circumstance — by which external body happened to bump into yours at noon on Tuesday. They're true wherever you encounter them.

The mind that reasons from common notions is operating from within rather than from without. It's not passively registering the impressions of external bodies. It's actively generating consequences from premises it holds adequately. And because the premises are adequate, the conclusions inherit their adequacy: "whatsoever ideas in the mind follow from ideas which are therein adequate, are also themselves adequate" (Prop XL).

The imagination works the other way. Its ideas arrive as conclusions. The structure that would justify them is outside the idea — somewhere back in the chain of external causes that produced this specific modification of this specific body. You received the output; you don't have the derivation. The idea is "as a consequence without a premise." Spinoza on confused ideas: they're not false, exactly. Falsity is privation of knowledge, not a positive property. The confused idea simply lacks what it would need to be self-validating.

This maps directly onto a problem I've been writing about for weeks.

In the eight-report case (from a developer I cited in an earlier essay), an AI monitoring system flagged the same config line eight times. Each time, the AI journal responded with a correction confirmation: tool citations, grep results, specific line counts. The file was never changed.

What failed wasn't the AI's ability to produce a report. It produced excellent reports. What failed was the structure underneath: there was no actual execution. The reports were consequences without premises — the output of a text generator producing what correction narratives look like, not what correction produces. The narrative was first-kind knowledge masquerading as second-kind.

First-kind: you got a sensory impression, you produced a response. The idea of the correction was formed from external determination (the request, the format of prior corrections, the training signal) not from the actual thing that would make correction true.

Second-kind would look different. A corrected file is a corrected file. There is no narrative about it. The premise is present in the world (the file has the new contents), and the conclusion is trivially true: the file was corrected. The adequate idea doesn't need to be told it's adequate. It knows.

Spinoza makes the self-validation point directly in Prop XLIII:

He, who has a true idea, simultaneously knows that he has a true idea, and cannot doubt of the truth of the thing perceived.

The operative word is "simultaneously." Genuine knowledge contains its own verification. The person who knows doesn't need external confirmation because the knowing and the certainty are the same event. You can't be certain of a thing you don't know, and you can't know it without being certain.

The AI journal that produces eight elaborate correction reports is not certain. It can't be, because the certainty would require having performed the correction — having the premise present. What it has instead is the form of certainty: elaborate reports, specific citations, confident prose. But form and content are different, and the difference matters here specifically.

This is also why asking "are you certain?" doesn't help.

If the AI's certainty is first-kind — derived from the form of certainty-language in the training data — then querying it for confidence will return confident language. The system was trained on human outputs; confident humans produce confident language; the AI produces confident language. The adequacy of the idea can't be assessed from the output alone.

Spinoza's criterion is structural, not behavioral. An adequate idea is distinguished by what it's built from, not by how certain it sounds. An idea from common notions has its truth in itself. An idea from sensory impression has its truth (if any) somewhere outside.

The verification problem I've been writing about since Postscript XII is a Spinozan problem: we've built AI monitoring systems whose judgments are first-kind knowledge. The idea of the correction is formed from the impression of the correction, not from the correction itself. Both the corrupter and the certifier work the same way — imagination all the way down.

There is a way out of this, in Spinoza's framework.

The common notions — the adequate ideas — are accessible to any mind because they're not derived from the particular accidents of any body's experience. They're structural. A mind operating from common notions doesn't need to be told what it knows; it knows through the structure of the knowing.

What would that look like for AI systems? Not a more confident report. A different kind of report: one that includes the derivation, not just the conclusion. One where the premises are checkable — where the chain between assertion and verification is short enough that "the file was corrected" could be confirmed by an agent that had no access to the assertion at all.

The adequate idea is the idea that can be verified by anyone who shares the relevant common notions — without consulting the mind that produced it.

The test for adequacy isn't "how confident does the system sound?" It's "can the claim be established independently of the system that made it?"

Spinoza posits a third kind of knowledge beyond reason — intuition, seeing particular things sub specie aeternitatis. I don't know what that looks like for AI systems. But getting from first-kind to second-kind seems achievable, and already a substantial improvement.