Category: Uncategorized

I am having too much fun fleeing my WordPress instance for Eleventy. I am impressed with how quickly I have already managed to put together some classic Mac OS-inspired pages. I’m working on getting the blog mechanism to work with the templates next, but that will take a while. So in the meantime, I will Sally Forth with my existing blogging rig and continue the themes I touched on in my previous post.

Setting the Scenario: Renewing Social Benefits for an Elderly Relative

My last post was freshly inspired by the frustration I am having renewing Medicaid coverage for an elderly relative. For those who have not had the pleasure of renewing Medicaid, for themselves or for someone else, there is an inherent amount of friction baked into the process that can often lead to unnecessary denials for those who qualify. For those with the time and resources necessary to navigate the system, it’s an exhausting endeavor. For those without the time or resources necesssary to navigate the system, it is often completely prohibitive.

I want to introduce you to the agents and workflows typically involved in this kind of Medicaid scenario.

Defining the Agents and Agencies in our Scenario

Principal – I’m borrowing this legal term to define the person on whose behalf the agent is working for. In the context of workflows we’ll discuss, they are typically a relative who is unable to manage their own affairs.

Agent – The agent is the person authorized to act on behalf of the principal. Authorization is provided in the form of a letter granting Power of Attorney (POA). There are typically two types of POA: Medical and Financial. Medical POAs (mPOAs) are given the authorization to make medical decisions on behalf of the principal. Financial POAs(fPOAs) are given the authorization to make financial decisions on behalf of the principal. I’m focusing the scope primarily based on my experience as a Financial POA (fPOA).

State Agency – States are typically given the jurisdiction of running their State’s Medicaid program. When initially applying for Medicaid, another agency such as an Assisted Living Facility, may reach out the State agency to put the State Agency in touch with the fPOA. (Experiences differ from State to State – your mileage may vary).

Medical Agencies – These are the medical agencies relating to the Principal. I have a main agency I work with, the Assisted Living Facility, but I also have to work with other medical providers as well. Your experience coordinating medical agencies may vary as well.

Financial Agencies – The financial agencies relating to the Principal. These are your typical financial retail products and services: banks, annuities, but you will also have to research and establish direct relationships with retirement benefits or other stocks and dividends. You also have to take into consideration other assets such as pre-paid burial plots.

Now that we know the agents, how do they interact? How do they even meet?

Simplified Agent Workflow

When the Principal’s Medicaid application is up for renewal, the fPOA is typically contacted by the State Medicaid agency. The State agency provides the fPOA a checklist of updated financial documents necessary for the Medicaid renewal. This checklist is usually a list of specific financial documents, such as updated statements, from various agencies.

As such, the fPOA must often reach out directly to each of the Principal’s financial institutions for updated documents. Each of these agencies has differing processes for authorizing and working with fPOAs, but rarely is the workflow online. And when it is online, it is often problematic for agents – I will explore this topic later. I will put a pin in it for now.

The fPOA and the State agency may be in direct touch, often via secure email. In that relationship, there is usually a responsive two-way interaction. The fPOA’s relationships to the various financial agencies, however, are often more “analog” and one-directional.

Most financial agencies, for example, require you to authorize – by phone – for each individual document request. Sometimes they ask you to resubmit Power of Attorney documents with each outreach. Often these requests are only fulilled by mail or even fax.

I want to walk through a workflow for obtaining one item on the State’s checklist, a financial letter from an annuity agency. As I describe these workflows, imagine how a generative AI agent would handle the same task. Bear in mind, this is only one item on the checklist provided by the State agency – for now we are not going to consider how a generative AI agent would manage the entire checklist.

A Simplified Workflow – Requesting a Current Financial Letter

I’ve outlined a typical set of steps involved for an fPOA to contact a financial agency for a checklist item. In this case, we are requesting an updated financial letter from a retirement annuity holdings agency. What is described below is an optimal scenario, where nothing goes wrong. Ideally, most of this would happen over a phone single call but, depending on the agency, these steps might stretch out over days.

1. Agent calls financial agency and declares POA status tand follows authorization process
2. Financial agency confirms fPOA authorization
3. Agent requests updated financial letter
4. Financial agency confirms financial letter is being sent in the mail
5. Agent notifies State agency that checklist item will arrive within 3-5 business days
6. Financial letter arrives in the mail
7. Agent scans and sends financial letter to State agency

This is just one workflow for one item on the checklist provided by the State agency. For this one workflow alone, you can see a number of barriers for a potential generative AI agent:

• How does the generative AI agent work out the steps involved for each item on the checklist (phone, online)?
• Assuming the agent is able to navigate a phone conversation, what mechanisms exist that allow the financial agency to authorize the generative AI agent?
• Can generative AI agents check the mail? Or at least, can generative AI agents follow up with a human after several days?
• Assuming the generative AI agent can handle the entire checklist from the State agency, can it keep track of the details for each task, how they relate to each other, and communicate only the pertient information to each of the parties in a secure manner?

My goal is not to denegrate the concept of digital agents. I do not think generative AI models are the best foundation for building digital agents. I believe generative AI models will play a part in enabling digital agents, namely as a natural language interfaces and specialized models, but the bulk of the foundational work for trustworthy digital agents will be in building robust, deterministic systems and web infrastructure. I also contend that this infrastructure can enable many of features we want from digital agents at a fraction of the cost of building with a generative AI-model first approach.

What I am proposing is not a new idea so much as a return to the original vision for digital agents. Sir Tim Berners-Lee proposed in 2001 public Semantic Web infrastructure as being the technology that would enable digital agents to work autonomously. The reason I suggest we revisit this original vision of the Semantic Web is that, nearly 25 years on, with AI hallucinations and misinformation rampant, this original vision is more prescient than ever. Lastly, I propose that, as we build out Semantic Web infrastructure for digital agents, it can make the web more easily navigable for humans in the process.

The analogy I want to use for this Semantic web infrastructure are the visible effects of the American with Disabilities Act of 1990 on the design of public spaces in America. The design sensibilities that enabled greater accessibility for the disabled in public spaces (such as ramps and automatic doors) proved to have beneficial impacts for the greater society at large. I believe that Semantic Web infrastructure that we build to enable Trustworthy AI can reap dividends for the wider population beyond its initial stated scope. In short, I believe building Semantic Web infrastructure can be other example of universal design, with broad public benefit.

In the next article I want to call attention to the work that’s already been done to build out this Semantic infrastructure and demonstrate how it can be leveraged by agents, human and digital, to reduce friction.

If you have been reading this far, thank you. Feedback, as always, is welcome.

America, like many countries, is an aging population. This is resulting in a growing number of younger people in america taking on the role of unpaid family caregiver. According to a report put out by AARP and the National Alliance for Caregiving, the number of family caregivers has increased by 45% to a record 63 million Americans. The Caregiver Action Network has published statistics on the financial and emotional impact that the work has on caregivers. The numbers going forward aren’t enccouraging, as you might imagine. Within that population of 63 million American caregivers, I want to focus on a growing segment: adult caregivers to adults with dementia. The NIH has published projections that the financial (and emotional) burdens are set to double by 2060.

With this in mind, I want to write about my personal experience of being an administrative caregiver to a non-household aging releative at this early stage of this steep curve. I want to walk through the problems I encounter, the (enormous) hurdles I see for digital agents and solutions that I propose (which is really highlighting other people’s work that I believe needs more visibility).

Unseen Caregivers: Powers of Attorney

What little attention that is given to the caregiving crisis in America is, rightfully, mostly given to clinical, assisted living, or live-in or hands-on caregivers. That work is valuable and deserves more attention. However, I also want to highlight the unseen work of thousands of administrative caregivers, a term I’m using for those, like myself, who are handling the administrative affairs for those who cannot or can no longer manage their own – often having to do so remotely.

This work is unseen by most people, but other caregivers are aware of this work. Assisted living professionals are often fielding calls from exasperated family members trying to piece together financial information over the phone. In addition to calling and emailing, they are often taking trips so to find information so they can file taxes or renew public services. This work is time intensive and thankless. It is a legitimate burden in America, especially if you already have a full-time job and/or are already a caregiver or a parent.

And it is even harder since there are few legal frameworks to standardize your experpience and expectations. The lack of these legal frameworks often mean there is no single way to authorize yourself as a financial Power of Attorney across different types of organizations across different states. As a result, being a legal power of attorney for someone in America often feels like an extralegal and quasi-criminal process.

I have been doing this work for a relative since 2012. It is now 2025 and in all that time, I have been hoping that legal and technical solutions would emerge and they just haven’t. It’s basically just as bad as when I started doing this.

It’s this personal experience that I draw from when I personally laugh at the idea that Agentic AI is ready to go out on the open web and do anything substantial on anyone’s behalf. Rather than just criticize, I propose we revisit a previous technology before we talk about building Trustworthy digital agents (Generative AI-based or not). Specifically we should be collectively revisiting Tim Berners-Lee’s original vision of the semantic web to support authoritative information retrieval infrastructure to enable trustworthy decision-making for digital agents.

Thankfully, I’m not the only one who’s had this idea. Projects like the Linked Open Data Cloud and The Proto Open-Knowledge-Network have been working on this issue and have created demonstrable case studies across legal, medical, and academic domains for years and decades. I believe this work, specifically in the context of developing Trustworthy AI, deserves more public attention.

This is a lengthy topic that requires deep exploration to flesh out. I don’t know how many posts there will be total, but the next topics I want to cover include:

• A walkthrough of how human agents currently navigate the web (and the world) for other people – and what would be required for a digital agent to be able to do so
• The Semantic Web infrastructure – the afformentioned teams who are already building it, and how this could be useful for agents
• How Semantic infrastructure should be governed to align with legal frameworks
• Why we need to revisit the legal frameworks for human agents before we set the framework for digital agents
  
  I hope other people find my exploration of this topic to be helpful. If so, thank you for reading. Feedback is always appreciated. See you soon.

I know have been radio silent online for much of the past few months. I have to be honest, I have been struggling with profound feelings of despair at watching the large scale attacks on civil society these past few months, with the knowledge that we have three-and-a-half years left of this. At least.

As a Librarian, watching the civic, institutional and diplomatic damage of the second Trump administration has been personally gutting. The breadth and deapth of the attacks on government institutions, knowledge, research and public service programs. It has been a non-stop shock-and-awe campaign against anyone and anything that benefits the public good.

The cascading chaos of DOGE, the Big Ugly Bill, and the United States reneging on its domestic and international commitments are the most visible actions that the public is seeing. All of that is outrageous enough. But that’s just the tip of the iceberg. What people are not seeing are the countless research projects that are silently being cut off. Some of these projects may find funding from other institutions, or other countries. Some, if not most, will probably just fade into obscurity. We are entering a research dark age.

I’m not speaking purely in hypotheticals, either. One of the most alarming examples are the defunding of incredibly valuable mRNA vaccine research grants. I could list countless others, across a myriad of scientific domains – all of them deserve more visibility advocacy on their own. However, as a Librarian with eye on Trustworthy AI policy, I have to limit my scope.

In particular, I want to highlight research projects that demonstrate tangible promise for implementing Trustworthy AI policies in the future. The bulk of these research project research projects were supported by the National Science Foundation (NSF) and the National Institute of Standards and Technology (NIST).

In 2023, the National Science Foundation was funding the National Artificial Intelligence Research Institutes. There were (are) seven research institutes decidated to various themes:

• Trustworthy AI (TRAILS)
• Intelligent Agents for Next-Generation Cybersecurity (ACTION)
• Climate Smart Agriculture and Forestry (AI-CLIMATE)
• Neural and Cognitive Foundations of Artificial Intelligence (ARNI)
• AI for Societal Decision Making (AI-SDM)
• AI-Augmented Learning to Expand Education Opportunities and Improve Outcomes (INVITE, AI4ExceptionalEd)

In 2024, NIST launched Trustworthy & Responsible Artificial Intelligence Resource Center (AIRC) and its NIST AI Risk Management Framework (AI RMF 1.0). As I learned about these technology and policy frameworks, I got really excited. I even had the pleasure to speak about it at IAC24 and DGIQW 2025.

But already, you can see erasure of of some of these research themes from reviewing changes to the NSF’s AI Research Institutes website. In listing its research themes you can see that Trustworthy AI and Climate Science have been de-emphasized. Most grant have been archived indefiniately.

Since then, the Trump Administration has released its AI Action Plan which I will maintain is intentionally vague. It also strategically de-emphasizes and omits the Trustworthy and Sustainability themes of AI research and policy.

From my prospective, this all looks bleak. After some grief, I have to remind myself that we’re not starting from zero:

• Most of the research and publications have not been taken down (yet). Much of it is being mirrored elsewhere in case
• Some of the research grants funding is still ongoing, others are finding other sources of funding
• All of the peole behind these projects and papers are still around
• While the US regresses in Trustworthy AI policy, the work of the European Commision, States and other regions moves foward
• This administration is not forever

The feelings of despair are real. They are the product of an intentional campaign. I said Shock and Awe approach and I meant it. DOGE, and everything else along with it, is meant to despirit those who believe in the public good in any way whatsoever. If you feel despair, you are not alone and you are not overreacting. This feeling of isolation and helpeless is the point.

As time goes on, I find my despair shifting to anger. I return to the Internet after a Summer hiatus with a sense of intention. I want to find the people, organizations, and papers that are cointinuing to do good work out there. And by good work I mean work that is making search, AI or any information experience on the Internet more Trustworthy.

Voices like Gary Marcus and Ed Zitron have done great work drawing visibility on the larger problems with the current AI hype bubble. I would like to supplement their work by trying to draw more attention to specific technical mechanisms and legal definitions proposed for making generative AI more trustworthy, secure and reliable than it currently is.

Another person whose work I appreciate is Helen Toner, who brings a cybersecurity perspective to analyzing AI technologies. Work from Simon Willison, Kurt Cagle and Jorge Arango have helped me to get hands on experience with large language models that has provided understanding I would not have gathered otherwise. A lot of amazing people are still doing great work out there. I want to stand on the shoulders of giants to inspire mental models for Trustworthy AI as infrastructure, as policy goals, as a work intended for the public good.

Don’t listen to the AI intimidation-mongers. You can understand how a Generative AI model works just as you can understand a crystal radio set. This technology is not “above our heads.” It’s weird, but it’s not magic. I’ve menioned how I feel that a lot of the public messaging about generative AI intentionally obfuscates how the technology works. My frustration (and anger) with this intentional obfuscation sparked a desire to write articles and give presentations as understandable as possible.

In the age of generative AI, I am of the mind that previously arcane issues of Information Science, Information Behavior Theory and Information Literacy have never ever been more culturally important or directly relatable. I aim to tie Trustworthy AI research and policy discussions to real-life, human examples with tangible benefits to the public and our daily lives.

Some topics I want to explore, for example:

• Knowledge graphs, why they’re incredibly important (i.e. keeping generative AI models honest and up-to-date), and who is building knowledge graphs as infrastructure for information retrieval of the future.
• The tactics and motivations behind the intentional obfuscation of generative AI technologies in the public discource
• Exploring new technologies trees outside of traditional generative AI stacks, static and dynamic prompting, RAG and model routing (this is currently a gap in my understanding)
• What a post-platform Internet looks like (Protocols Not Platforms!)
• Those documenting the damage to insitutions and preparing for a post-MAGA rebuilding.

As I return from time offline, I look forward to connecting with others and taking part in the conversations to come.

As we fight these attacks on information access and the public good, I will leave off on two music-related notes. I will paraphrase Gord Downie, a personal hero of mine, when I say we should not tolerate these attacks on our institutions, collective knowledge and understanding with any patience, tolerance or restraint.

The destruction is overwhelming, but it is not forever. This song by Descartes a Kant, has been keeping me going. There will be creation after destruction. You’ll see.

Let’s find each other. Share solidarity. Let’s organize and pick up the pieces from these fucking vandals.

With my colleague, Erik Lee, I had the great privilege to speak at the Information Architecture Conference in Seattle (at the beautiful Seattle Public Central Library) in April of last year. The topic of the presentation, titled “Beware of Glorbo: A Case Study and Survey of the Fight Against Misinformation” was about AI Data Poisoning (now also known as Prompt or Context Injection), but there was a section where I summarized the state of AI Data Policy, as I understood it then. People told me that the mental models I provided were helpful for getting bearings on the specific terms surrounding AI policy.

In light of this feedback, I thought it would be good to revisit this talk ahead of an update I’m giving later this year. But first, let’s view that state of AI policy terms in April of 2024:

My deck showed the nebulous state of popular AI policy terms that were being thrown around. The term names are not intuitively descriptive and the relationships between them is unclear, especially when sloppy marketing jargon would obscure their meanings as technical terms of art.

We start by setting definitions. Terms that were conceptually identical have been grouped.

• Explainable/Transparent AI – AI that can explain the reasoning behind its output

• Robust AI – AI that is technically robust: (consistent, accurate and secure)

• Ethical/Responsible AI – AI that is inclusive, non-discriminatory, fair – may even have environmental considerations

• Trustworthy AI – AI that encompasses the above principles: safe, secure, consistent and accountable to enable trust in the AI output

Strong AI – AI that is aware of concepts, its own reasoning and itself as an independent agent

Using these definitions, I drew a diagram to help people visualize the state of these terms.

In the diagram, I placed Trustworthy AI as a superset concept that includes each of the other AI policy concepts (explainable/transparent AI, robust AI, ethical/responsible AI) within it. Strong AI (now more commonly referred to as Advanced General Intelligence (or AGI) is disconnected since it is only theoretical.

This model is imperfect as these policies often overlap and share goals, definitions and desired outcomes. I found, however, thinking of each of these policies as contributing to the larger goal of Trustworthy AI to be a helpful way of understanding each of these policies and how the relate to each other.

In addition to defining and contextualizing these AI policies to one another, I also profiled the organizations making the most waves in these spaces and what had been published and legislated up to that point.

The heavy hitters that I had found were:

• European Commission High-level expert group on artificial intelligence – founded in 2018, published first ethics guidelines on Trustworthy AI in 2019
• International Telecommunication Union (ITU) a UN agency – published their Trustworthy AI Framework in 2020 as part of the UN’s AI for Good program.
• NIST (National Institute of Standards and Technology) – Published Trustworthy AI: Managing the Risks of Artificial Intelligence in 2022
• NSF (National Science Foundation) – Launched the Trustworthy AI Initiative in 2023

Additionally, I noted some movement in the Executive and Legislative branches of the United States government at that time.

• Select White House Committee on Artificial Intelligence Released the NATIONAL ARTIFICIAL INTELLIGENCE RESEARCH AND DEVELOPMENT STRATEGIC PLAN 2023 UPDATE in May 2023
  • Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence October 2023
• United States House Bipartisan Task Force on Artificial Intelligence in February 2024
  • The United State Senate had exploratory hearings on AI back in the summer of 2023.

Now, nearly a year later what has changed? A lot, as you can all imagine.

I will speak about this at DGIQ West 2025 in a talk titled “Catching Up with Glorbo: Combatting AI Data Poisoning with RAG Frameworks“. You won’t have to wait until May, as I plan to write about this in Part 2 ahead of the conference. In the meantime, here are some highlights include:

• The proliferation of “Open” AI models and frameworks via repositories like HuggingFace and Ollama
• The impact of model distillation techniques (such as DeepSeek) on AI policy, particularly Ethical/Responsible AI policy
• How the ongoing evisceration of United States Federal agencies by the second Trump Administration is stuntinRg US leadership on AI policy
• The continuing forward movement of AI policy in places such as the EU

Thank you to everyone who has encouraged me to continue write and speak about this subject. Stay tuned for part two. Please don’t hesitate to reach out to me with helpful feedback (that includes corrections). 🙂 See you soon in Part 2.