Qualification
HMIA 2025
YOU MUST BE THIS TALL TO RIDE THIS RIDE
Announcements
Oral Exam Time Slots
Oral Exam Practice Bot
Ethics "Essay" Work Plan + Allowance on Oral
https://tinyurl.com/ORALEXAMHELPER
From last time
My robot/roommate lacks the virtue of...
reciprocity
I would encode a copying mechanism so it copies the microbehaviors such as head nods, greetings so it is more human and “friendly”
honesty
Report what it has done or plans to do to me. Agree on things ahead of time, check progress or send reminders more frequently
altruism
bake in severe penalties
accountability
punishments for misbehaving
initiative
Incentivized measured risk taking
reliability
Have thorough negotiation before any kinds of cooperation, agree on the role of each other
I could improve the alignment by...
Agent
Creation
Pedagogy
Deployment
FILTERS
(guardrails)
VIRTUES
CONTRACTS
REPORTS
Opportunities for Alignment
REWARDS
PUNISHMENTS
(and incentives)
WHAT ELSE?
WHERE ELSE?
What would be a good way to pre-clear roommates? Imagine a company that can create "certified good roommate" service.
Drawing in Light Photography
HOOK
Qualification refers to mechanisms determine who is allowed to do something.
HMIA 2025
PRE-CLASS
Qualification
HMIA 2025
What qualified agents/objects have you dealt with recently? WORD CLOUD
Activity: TBD
PRE-CLASS
CLASS
Qualification: why can it produce alignment?
From virtuous roommates to certified roommates.
Hook and Definition and Alignment Cards
Readings
Hadfield et al. — Alignment through Registration
Effective external oversight requires that powerful technologies be legible to regulators. The authors propose confidential registration for high-capability AI models, modeled on corporate, financial, and hazardous-materials registries. By disclosing key information to public authorities, developers make advanced systems visible to democratic governance — aligning private innovation with public accountability.
HMIA 2025
*still recovering from World Series; didn't read
SRFWS;DR*
ISO — Alignment through Standardization
An international federation of national standards bodies that aligns organizations and states by codifying shared procedures and benchmarks. Standardization allows firms to coordinate quality practices and governments to harmonize regulations, creating interoperability across borders.
Foucault — Alignment through Observation
Modern power aligns individuals with institutional norms through visibility: observation, measurement, and documentation. The exam, the record, and the file turn people into legible objects of knowledge and control. What appears as rational management or professional assessment is also a disciplinary technology — one that enlists individuals in their own regulation.
ISO — Alignment through Standardization
An international federation of national standards bodies that aligns organizations and states by codifying shared procedures and benchmarks. Standardization allows firms to coordinate quality practices and governments to harmonize regulations, creating interoperability across borders.
Hadfield et al. — Alignment through Registration
Effective external oversight requires that powerful technologies be legible to regulators. The authors propose confidential registration for high-capability AI models, modeled on corporate, financial, and hazardous-materials registries. By disclosing key information to public authorities, developers make advanced systems visible to democratic governance — aligning private innovation with public accountability.
Readings
STANDARDS ORGANIZATIONS
Registries
STOP+THINK: What is the main purpose of the authors' proposed AI registration system?
A. To restrict the public’s access to AI technologies like ChatGPT.
B. To give governments basic visibility into who is developing large
generative AI models and their potential risks.
C. To make all AI development data publicly available to promote
transparency.
D. To slow down innovation until new international laws are passed.
This op-ed argues that the first step toward responsible governance of generative AI is registration—a basic requirement that developers of large, high-capability models confidentially disclose key information to government authorities. Because only private companies currently know what models exist, how they’re trained, and what they can do, policymakers lack the visibility needed to assess risks or enforce existing laws. A registration system, modeled on those used for corporations, securities dealers, and hazardous materials, would not restrict innovation but would make development legible to regulators. By distinguishing transparent, law-abiding actors from those operating in the shadows, registration would lay the groundwork for future licensing, oversight, and informed public policy—ensuring that the governance of AI rests on democratic visibility rather than corporate secrecy.
STOP+THINK: How do the authors distinguish the proposed AI registry from the EU’s AI Act?
A. The proposed registry would make all information about AI models
public.
B. The proposed registry would include strict licensing requirements for AI
deployment.
C. The proposed registry would protect trade secrets and share information
confidentially with governments rather than the public.
D. The proposed registry would ban unregistered AI models from being
created.
Michel Foucault 1926-1984
A French philosopher, historian of ideas, literary critic, and political activist. His work addressed the relationship between power, knowledge, and liberty. He described how they are used as a form of social control through institutions. Foucault rejected labels but is usually called a structuralist and post modernist. His ideas have influenced thinkers in many field - anthropology, communication studies, criminology, cultural studies, feminism, literary theory, psychology, and sociology.
Most of his works are histories of ideas that trace out how bodies of knowledge and styles of thinking evolve over time as they interact with social institutions. He called his approach "archaeologies" or "genealogies."
STOP+THINK: According to Foucault in “The Means of Correct Training,” how does modern disciplinary power produce “qualified” individuals?
A. By punishing disobedience through public spectacle and deterrence.
B. By training and continuously observing individuals so their behavior conforms to institutional norms.
C. By allowing individuals to act freely without supervision, trusting their innate morality.
D. By granting privileges and rewards based solely on birth or social class.
Social Credit Systems are a textbook case of hierarchical observation turned algorithmic. Every traceable act - shopping, friends, debt, even video game time - feeds into a moralized score that determines access to travel, jobs, or loans. “Normalization” becomes literal: algorithms define what counts as normal or deviant.
Some university admissions systems purchase data on students’ online engagement or campus movement to assess “fit,” “persistence likelihood,” or “risk of drop-out.” ChatGPT calls it "A parody of care: total visibility mistaken for understanding."
Some companies use “people analytics” - constant productivity tracking: keystrokes, webcam attention, Slack activity, biometric sensors in warehouses. Piles of data allow the substitution of metrics for judgment. The “good worker” is defined by data traces, not by integrity or competence or virtue.
But these don't really exist, do they?
The story: In Moneyball (Michael Lewis), the Oakland A’s use data analytics to identify undervalued players and outcompete richer teams. It’s a story about quantification rescuing judgment from bias — using statistics instead of scouts’ “gut feeling.”
At first, this is counter-disciplinary: it defies the old boys’ network and exposes their irrational prejudices. It began as an act of resistance to arbitrary, elitist evaluation - a push for fairer, evidence-based qualification.
But its success spread a technocratic mindset: everything and everyone can be optimized if only we collect enough data. The dream of “perfect metrics” thus becomes the disciplinary fantasy Foucault described - continuous evaluation producing docile, self-monitoring subjects.
IT'S TURNED SPORTS INTO STATS
IS
THIS REASONABLE?
IS
THIS
ALIGNMENT?
EVEN
HERE
Why Foucault? What's he got for us here?
In earlier chapters of the book, Foucault described the absolute power and brutality of the king as it was visited upon the body of the condemned—a crude form of power that had to be grandiose and visible.
As Foucault tells it, over the next few centuries power in Europe evolved into finer and more subtle forms.
By the end of the book, when we reach the modern era, power has become “inscribed” in knowledge itself. But in this intermediate chapter he wants to show how training and education transformed power from its older, overt form into something quieter—and thereby more efficiently powerful.
Sometime around 1625 we start to see education and training associated with “strict discipline.” Imagine, if you will, a group of underqualified recruits made to march and drill over and over. Why is marching so important? Foucault suggests that it’s how power makes individuals “both as objects and as instruments of its exercise.”
The first stage is hierarchical observation. Society becomes fascinated by the idea of seeing everyone—behavior observed, data recorded. The architecture of military training grounds and schools alike was designed to make the trainees constantly visible. “The disciplinary institutions secreted a machinery of control that functioned like a microscope of conduct; the fine, analytical divisions that they created formed around men an apparatus of observation, recording, and training.”
In workshops, schools, and the army, the institution continually measures and ranks its members, creating a dominant sense of “the normal.” One must learn to meet the standards and avoid being classed as a failure or a deviant. (You might recall as a child wondering whether you were normal. Where did that idea come from?) The point is that there are interests at work here—it’s not “just happening.”
Observation and normalization come together in the examination. As the core of the training process, the exam reduces individuals to “describable, analyzable objects.” “The examination, surrounded by all its documentary techniques, makes each individual a ‘case.’” It produces a record of who is fit to act.
Foucault disrupts our taking for granted things like examinations and credentials as neutral ways to measure competence and value. He undermines our habit of seeing the world as merely an aggregation of individual agents. Instead, agents are produced by the systems that observe, measure, and certify them.
So why are we reading this?
If Foucault is right, our own apparatus for certifying individuals, organizations, experts, and machines is not a neutral gatekeeping system either. Deciding who may practice, or what counts as responsible or safe, is itself an exercise of power—performed by someone, for some purpose. Foucault helps us keep a critical edge as we think about how to design systems that keep both human and machine intelligences aligned—without forgetting who defines the standards, who enforces them, and whose interests they serve.
So why are we reading this?
If Foucault is right, our own apparatus for certifying individuals, organizations, experts, and machines is not a neutral gatekeeping system either. Deciding who may practice, or what counts as responsible or safe, is itself an exercise of power—performed by someone, for some purpose. Foucault helps us keep a critical edge as we think about how to design systems that keep both human and machine intelligences aligned—without forgetting who defines the standards, who enforces them, and whose interests they serve.
HOW DO ALL THESE THINGS FIT TOGETHER?
DO ALL THESE THINGS FIT TOGETHER?
HMIA 2025
START+THINK: Pair up and think of all the licensed or certified or qualified people or products you use. Make a list of as many as you can think of.
HMIA 2025
STOP+THINK: In a lot of stories about magic, sorcerers
must come from one particular family or clan or academy.
How might that be seen as an alignment mechanism?
It could ensure that all sorcerers receive the same training.
Inherited trust: clan self polices to maintain reputation.
Reduce risk of rogue sorcerers.
Family as efficient way to pass on metanorms about use/abuse of magic.
Why a substitute for merit-based? Avoid dispute transactions? Avoid measurement disputes?
Does it prevent corruption or just consolidate power?
In stories like Le Guin's Earthsea cycle, C.S.Lewis' The Magician's Nephew, Okorafor's Akata books, and Okri's The Famished Road magic either runs in families or is restricted to a wizardly lineage with themes around the moral use of power for community benefit rather than self interest.Agent
Creation
Pedagogy
Selection
Deployment & Practice
Ex Post
Selection
Recertify
Monitor
qualification
qualification
qualification
Revisiting When Can Alignment Happen?
qualification
HMIA 2025
HOOK
START+THINK: How does qualification produce alignment?
HINT: what's trust got to do with it?
HMIA 2025
HOOK
START+THINK: How does qualification produce alignment?
NOTES: Visibility is one piece - transforming an unknown agent into a known, if anonymous, type. But strong parallels with expert intelligence in general: when expertise is organized as professions and similar, there is a genericization of intelligence - at some level, a lawyer is a lawyer is a lawyer. In fact, that's why we say "my lawyer" or "my therapist" rather than "Sam" or "Diana." So, the underlying misalignment pathologies that qualification addresses would be something like opacity and variance and trust's relative short reach dependent on familiarity and first hand experience. Registration is a start on visibility. Licensing is variance on entry and monitoring and recertification is variance over time (and maybe space).
Are there misalignments in my (partially) virtuous roommate (robot) that can be addressed by qualification? On the roommate side we sometimes see people with "shared living resumes" and one might do credit or reference checking but not clear what missing virtue this addresses. Some parts of qualification are, perhaps, best thought of as virtue verification?
Licensing and certification stand in for moral knowledge about strangers.
Followup question: Qualification aligns by making agents legible, comparable, and substitutable — but at what cost to individuality and genuine virtue?
When do we want virtue verification (qualification) rather than virtue cultivation (training)?
And what happens when the verification replaces the virtue?
Qualification
One route to alignment is to prevent likely misaligned agents from acting in the first place. Qualification mechanisms support alignment not by directly shaping behavior, but by controlling who or what is permitted to act.
They rest on the assumption that observable capacities, traits, or preparations can serve as reliable proxies for alignment—that a well-trained, certified, or vetted agent is more likely to act in accordance with shared goals and values.
Qualification operates upstream of action, ensuring that agents possess the knowledge, integrity, and judgment required to act responsibly. By selecting, preparing, and assessing agents in advance, qualification reduces the burden on downstream oversight and correction, making cooperative systems safer and more predictable.
What does mean?
Most say it means "Conformité Européenne."
The CE mark on products means that the manufacturer or importer affirms the goods' conformity with European health, safety, and environmental protection standards. CE is not a quality indicator.
Curiously, the European Commission (the executive branch of the European Union) does not appear to acknowledge this definition. No reliable source confirms that CE is an abbreviation for anything.
Wikipedia
Creating Public School Teachers
How does CE get on a product?
EU Regulations
Manufacturer
Product
Notified Body
Consumer
B2B
hires
manufactures
certifies*
affixes mark to product
confidence EU rules are followed
can use CE
has expertise on
internal tests and
self-certify*
*Not Quality, but compliance
What do you see as the pros and cons of this approach to alignment? Where might it work better or less well?
HMIA 2025
CLASS
What's the difference between these two symbols?
HMIA 2025
CLASS
In what sense are these alignment mechanisms?
Trademark (™) vs. Registered Trademark (®)
(both signal “ownership,” but at different levels of qualification)
FeatureTrademark (™)Registered Trademark (®)
| Type of qualification | Self-asserted claim | Externally verified and legally recognized right |
| Mechanism of alignment | Anyone can declare that a symbol, name, or phrase identifies their product or service. It signals intent and reputation but carries no legal vetting. | The mark has been examined and approved by a national trademark office, creating official protection and legal enforceability. |
| Who vouches for it | The claimant — the company or individual asserting ownership. | A qualified authority (e.g., USPTO, EUIPO) confirming it meets registration criteria. |
| Analogy to human contexts | Like saying “I’m a doctor” without a license — maybe true, but unverified. | Like being board-certified or state-licensed — the claim has been checked and registered. |
| Strengths | Fast, cheap, expressive of intent and reputation; allows early market presence. | Confers strong, enforceable rights; deters misuse; builds institutional trust. |
| Weaknesses | No formal protection — others can use or challenge it; trust depends on reputation and context. | Slower, costlier, requires disclosure and ongoing maintenance. |
Which approach for roommates? Robots?
UL means "Underwriters Laboratories." This is a voluntary third-party safety certification (though often required by insurers, buyers, or local codes).
The focus of UL is safety and performance — is it built safely, tested properly, and likely to function reliably under normal conditions?
How it works: UL or another Nationally Recognized Testing Laboratory tests actual samples of the product, audits factories, and issues a listing mark only after compliance with UL’s own safety standards.
Thus, the mark indicates independent verification of safety and quality and is seen as a trust mark or quality certification in North America.
WHAT
DOES
THIS
MEAN?
Feature UL Listing CE Marking
| Type of qualification | External certification | Self-certification |
| Mechanism of alignment | Independent, third-party verification ensures conformity to shared safety standards | Legal requirement that manufacturers attest compliance with regulations |
| Who vouches for alignment | A neutral qualified auditor (UL) | The agent itself (manufacturer) |
| Analogy to human contexts | Like getting licensed by a professional board (tested, inspected, ongoing oversight) | Like signing a sworn declaration or affidavit of competence |
| Strengths | Builds trust and market confidence; mitigates information asymmetry | Enables broad participation and innovation with less bureaucracy |
| Weaknesses | Costly, can slow innovation | Relies on honesty and traceability; abuse possible without enforcement |
|
|
|
|
|
Self-claimed Identity
Registered Identity
Third-party tested & listed
Self-certified Compliance
"Trust me!"
"I'm on record"
"I've been independently tested."
I promise I meet the standards.
AI analog: Voluntary “model card” or “system card.”
Developers describe capabilities, training data, and intended uses on their own authority.
AI analog: Formal registration of high-risk or foundation models with a public authority.
Model is logged, key data disclosed, but not yet audited in depth.
AI analog: External audit or licensing of high-capability models. Regulatory markets.
Independent labs evaluate robustness, red-team results, safety alignment; deployment only after certification.
AI analog: Self-attested safety compliance framework.
Developers certify that their model meets regulatory requirements (e.g., EU AI Act risk management) without prior inspection.
Shallow / Declarative
Third Party
Deep / Evaluative
First Party
Source of Credibility
Depth or Explicitness of Standards
Human-world logic
AI-world analog
Dimension
| Shallow vs. Deep |
Simple visibility vs. empirical validation
Model disclosure vs. technical testing and assurance
Self vs. External
Self-attestation vs. independent certification
Developer-led compliance vs. regulator or third-party oversight
Resumes vs references
Company mission statement or principles vs. certified B corporation
Car seller writes an ad vs. Carfax reports.
Membership card vs and Blackbelt
Nutrition label vs. health inspection
We're married vs. For 25 years I've washed your clothes, raised children, etc.
Privacy policy vs. security audit
Documentation, design specs vs. stress tested, monitored under changing conditions.
Model card vs. red-team certifiction
Qualification as symbolic assurance vs. qualification as statistical assurance.
That is, a qualified agent is one I can trust but trust is not just about feeling good and confident. It is also a matter of material assurance. Alignment in fact (vs. alignment in principle?) is about agents actually not causing harm and damage. So we pay for assurance or we require assurance (the most common version of this is "you can't participate in our market unless you are certified" so you are effectively trying to prevent players from gaming the system - taking the reward for providing a service that you pay for or reward on the assumption it is one thing (safe) when in fact it is another.
STOP+THINK: How do qualification and trust
interact for alignment?
-
Question: What’s the tradeoff between trust through verification and trust through responsibility?
-
Example prompt:
“If UL Listing represents professional accreditation, and CE Marking represents self-declared compliance, which is a stronger alignment mechanism — and which scales better?”
-
Possible connection:
This also parallels AI model registration vs. licensing debates:
-
Registration (CE style) = self-declaration of safety and compliance.
-
Licensing (UL style) = independent vetting before deployment.
-
Qualification
Is there a human equivalent whereby we try to restrict action to people of virtue?
Is group membership a form of alignment by qualification?
Might group level bias be related to alignment by qualification?
Summary: WHY....
A license is permission granted by a competent authority to do some act that would otherwise be prohibited. It converts private claims of readiness into publicly validated warrants to act.
A registry is an authoritative list. In the alignment context, it can be seen as a minimal form of regulation that requires agents to publicly declare their existence and, possibly, reveal some of their characteristics.
Certification allows others to engage with agents not as unknown individuals but as recognized types—doctors, pilots, engineers, or safety-tested systems—whose reliability can be presumed within defined limits.
Continuous monitoring—periodic recertification, audit, or performance review—guards against skill decay, value drift, and principal–agent slippage, ensuring that authorization remains deserved rather than assumed.
Alignment Card Materials
1. Create a category "Qualification" (or whatever you choose to call it)
2. How do you define it?
3. What pathology(ies) of alignment does it address?
4. Do registration, licensing, certification, continuous monitoring (relicensing) cover the whole concept?
5. What are the analogs across human, organizational, expert, and machine intelligences?
JUMP TO RECORD KEEPING
Licensing/Registration
Formally authorizing agents to act by verifying their competence, integrity, and accountability before granting access to power or sensitive roles.",
LOREM
Human Interaction
driving, machinery operation, activities on public resources (fishing, hiking, etc.)
personal services
residence and citizenship
Organizational
Who can do what in an organization
Organizations certified to provide certain services
Security clearances
Professional/Expert
Fundamental to the idea of professions
Exams, certificates, titles.
Memberships in professional organizations.
Machines
Mostly limited to organizations needing license to provide services.
Weights and measures and equipment inspection.
Security certificates, etc. on the net.
Device approval.
Registries allow for disqualification/revocation/delisting/being struck.
Certification/Accreditation
qualification made visible through formal assessment. It signals that an agent has met an established standard of competence or integrity, setting a performance floor below which action is unacceptable. Certification allows others to engage with agents not as unknown individuals but as recognized types—doctors, pilots, engineers, or safety-tested systems—whose reliability can be presumed within defined limits.
Distinguished from licensing: certification signals meeting a standard; license is granting of permission.
Human Interaction
Diplomas, degrees, test scores, and skill badges that signal readiness or competence.
Serve as shorthand for trust — others can rely on certified ability without firsthand evaluation.
Function as social currency for employment, reputation, or advancement.
Organizational
Third-party or internal audits (ISO, SOC, safety, or quality certifications) verifying compliance with standards.
Demonstrate that teams or products meet known benchmarks, fostering trust among partners and regulators.
Encourage continuous improvement and documentation discipline.
Professional/Expert
Board exams, specialty credentials, and continuing-education certificates showing mastery beyond licensure.
Maintain discipline-wide quality by verifying skill currency and ethical grounding.
Enable differentiation among practitioners while reinforcing shared professional norms.
Machines
Structured evaluation of AI or technical systems for accuracy, safety, interpretability, and robustness.
Certification artifacts (model cards, evaluation reports, compliance seals) communicate reliability to users and regulators.
Provide the evidentiary basis for deployment decisions or licensing approval in high-stakes contexts.
Continuous (or periodic) Monitoring
Continuous monitoring—through periodic recertification, audit, or performance review—is the ongoing verification of competence and alignment. It guards against skill decay, value drift, and principal–agent slippage, ensuring that authorization remains deserved rather than assumed. By coupling trust to evidence that is regularly renewed, continuous monitoring sustains alignment in dynamic environments where both agents and contexts evolve.
LOREM
Human Interaction
Periodic feedback, peer evaluation, or reflective self-assessment that keeps personal commitments and competencies current.
Annual reviews, coaching, and ongoing education sustain accountability and prevent drift.
Reinforces trust through visible maintenance of integrity and growth.
"Can we talk about our relationship?"
Organizational
Performance reviews, compliance checks, and recurring audits verifying that processes and personnel remain aligned with mission and policy.
Detects deterioration, mission creep, or control failures before they cascade.
Turns oversight into an ongoing learning loop rather than a one-time test.
Professional/Expert
Continuing education, re-licensure, and peer review ensuring practitioners maintain current knowledge and ethical standards.
Protects the public from skill decay or complacency.
Demonstrates that authorization to act remains continually earned.
Machines
Post-deployment monitoring, safety logging, and periodic re-evaluation of models or systems.
Detects performance degradation, value drift, or misuse in changing environments.
Enables rollback, retraining, or withdrawal when misalignment emerges — the technical analogue of professional re-certification.
HMIA 2025
Resources
Hadfield, Cuellar, O'Reilly. 2023. It’s Time to Create a National Registry for Large AI Models
McKernon, Glasser, Cheng, Hadfield 2024 AI Model Registries: A Foundational Tool for AI Governance
MacGillis 2019 “I Will Never Let Boeing Forget Her” ProPublica
MacGillis 2019 “The Boeing 737 MAX Is Cleared to Fly. Families of People Who Died on the Planes Wait for Answers” ProPublica
Schwellenbach and Stodder 2019 How the FAA Ceded Aviation Safety Oversight to Boeing
Pederson 2024 Medtech Lessons in the Boeing 737 Max Debacle
Gelles Kitroeff Nicas Ruiz 2019 Boeing Was ‘Go, Go, Go’ to Beat Airbus With the 737 Max New York Times
Weber Section of Politics as a Vocation on legitimate authority????
16 page excerpt from Foucault's Discipline and Punish "The Means of Correct Training" Perhaps just the section on the examination pp. 197-205. I think maybe too esoteric and hard.
Material from my experts in antiquity????
Cedefop "European qualifications framework (EQF)"
Wikipedia Editors Qualifications Framework
Wikipedia Editors CE Marking
Maryland State Department of Education "Pathways to Initial Licensure Teachers"
USFCR "System for Award Management"
NEXT TIME: INFORMATION,
TRANSPARENCY,
RECORD KEEPING
HMIA 2025 Alignment Cards Qualification
By Dan Ryan
HMIA 2025 Alignment Cards Qualification
Qualification as an alignment mechanism
