🤖 AI

Definitional alignment before capability alignment: a Design-Science framework for adjudicating claims about AGI

This paper proposes DAF-AGI, a Design-Science framework that prioritizes definitional alignment over capability alignment to adjudicate conflicting claims about AGI arrival by establishing ordinal criteria and governance audits for evaluating competing operationalizations.

Original authors: J. E. Aguilera Briones

Published 2026-06-12

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: J. E. Aguilera Briones

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Core Problem: The "Arrival" Argument

Imagine a group of people arguing about whether a new car is a "supercar."

Person A says, "Yes, it's a supercar because it goes 200 mph."
Person B says, "No, it's not. It can't drift, and its engine isn't V12."
Person C says, "It doesn't matter what it is; if it can drive itself to the store, it's a supercar."

They are all looking at the same car. They all agree on the facts (the speed, the engine, the features). But they can't agree on the verdict because they are using different rulebooks to define what a "supercar" actually is.

This paper argues that the field of Artificial General Intelligence (AGI) is stuck in this exact argument. Some companies say, "We have arrived!" because their AI writes good code. Others say, "We are decades away!" because their AI can't remember things long-term or solve new puzzles. The disagreement isn't about the technology; it's about who gets to write the rulebook and what the rules actually say.

The Solution: A "Rulebook Auditor" (DAF-AGI)

Instead of trying to write the one true definition of AGI (which the author says is impossible because everyone has different goals), the author built a tool called DAF-AGI.

Think of DAF-AGI not as a judge that decides who wins, but as a transparency auditor. It's like a "nutrition label" for definitions. When someone claims, "This AI is AGI," the auditor doesn't just say "Yes" or "No." It asks:

What are you measuring? (Is it speed? Is it memory? Is it how much money it saves?)
Who wrote the rules? (Did the company that makes the AI write the rules themselves?)
Can the rules be changed? (If the AI fails, can they just change the rules to make it pass?)

The Five "Rulebooks" (Measurement Families)

The paper looks at the five most common ways people try to define AGI and audits them:

The "Speedometer" (Performance Superiority):
- The Rule: "If the AI is faster/better than a human at tasks, it's AGI."
- The Flaw: It's easy to cheat. You can pick only the tasks the AI is good at (like writing emails) and ignore the ones it's bad at (like long-term planning). It's like saying a runner is an Olympian because they won a race on a track they practiced on, but they can't run on sand.
- Verdict: Low reliability.
The "Wallet" (Economic Substitution):
- The Rule: "If the AI can do a human's job cheaper, it's AGI."
- The Flaw: This mixes up "intelligence" with "economics." A robot arm might replace a factory worker because it's cheaper, not because it's "smarter."
- Verdict: Mixed. Good for economics, bad for measuring intelligence.
The "Map" (Capability Ontology):
- The Rule: "We have a map of different skills (memory, reasoning, creativity). Where does the AI sit on the map?"
- The Flaw: It's a great map, but the mapmaker (the lab) decides where the "finish line" is.
- Verdict: Good structure, but the mapmaker controls the finish line.
The "Report Card" (Psychometric Parity):
- The Rule: "The AI must match the full cognitive profile of a well-educated human adult."
- The Flaw: This is the hardest test. Current AIs are like geniuses in math but have the memory of a goldfish. They fail this test because they are "jagged" (great at some things, terrible at others).
- Verdict: High reliability, but current AIs fail it.
The "Learning Speed" (Skill-Acquisition):
- The Rule: "Intelligence is how fast you learn a new puzzle you've never seen before."
- The Flaw: Current AIs are great at memorizing old puzzles but terrible at solving new ones.
- Verdict: High reliability, but current AIs fail it.

The Big Reveal: The "Stylized Claim"

The paper tests a popular claim: "Current AI is AGI because it beats a well-educated adult at many tasks."

When the auditor (DAF-AGI) checks this, it finds a magic trick.

The claim uses the Speedometer rule (beating the adult at some tasks).
But it borrows the Report Card label ("Well-educated adult").
If you actually apply the Report Card rules, the AI fails because it has memory gaps.
If you apply the Speedometer rules, the AI passes, but only because the company picked the easy tasks.

The Conclusion: The claim is "underspecified." It mixes a loose rule with a strict label to make the AI look smarter than it is. The paper says we cannot declare "AGI arrived" until we agree on which rulebook we are using and who is allowed to change the rules.

The Concept of "Definitional Sovereignty"

This is the paper's most important political idea.

Imagine a country that imports all its food.

Scenario A: They buy food from a neighbor. The neighbor decides what "fresh" means. If the neighbor changes the definition of "fresh" tomorrow, the country has to accept it or starve.
Scenario B: The country buys the food, but they have their own inspectors. They can say, "We accept your definition of 'fresh' today, but if you change it, we will audit it and maybe reject it."

The author calls this Definitional Sovereignty.
Many countries (and the public) are currently in Scenario A. They are "Definition Takers." They accept definitions of AGI, "High Risk AI," or "Maturity" written by a few powerful tech labs in the US or China.

If a lab says, "This is AGI," and that triggers a law or a stock market crash, the world has to accept it.
The paper argues that nations need the capacity to audit, contest, and revise these definitions. They don't need to write their own definitions from scratch; they just need the power to say, "No, that definition doesn't work for us," or "We need to verify that claim."

Summary: What This Paper Actually Says

We can't agree on what AGI is because different groups have different goals (marketing, safety, investment).
More data won't fix this. Even if AI gets smarter, we will still argue about whether it counts as "General" intelligence unless we fix the definition first.
Current "Arrival" claims are shaky. They often mix easy rules with hard labels to make a system look like it has arrived when it hasn't.
Governance is the key. We need to stop asking "Is this AGI?" and start asking "Who wrote the rule that says this is AGI, and can we audit them?"
Sovereignty matters. Countries and the public need the power to challenge imported definitions of technology, or else they are being governed by rules they didn't write and can't change.

In short: Before we argue about whether the robot is "alive" or "smart," we need to agree on who gets to hold the dictionary and make sure they aren't changing the definitions to suit their own wallet.