Towards an AI Research Agenda for Elections and Beyond (Part 3)
Courtesy Getty Images
I recently posted my second part of this longer commentary about the AI research agenda that’s necessary for elections specifically, and a lot of government usage generally. Having explained the particular needs of use in election administration, this time I offer clarity about the needs for government computing and/or public-benefit generally.
The generative-AI-powered domain-specific natural language agents (NL agents or “NLAs”) that I described before are useful across a broad range of government and public needs. Not every possible benefit can be delivered by domain-specific NL agents, but nearly every part of government/public computing can benefit from them. And here’s the key point for the research agenda: it is exactly the same type of generative-AI technology that’s required across the board. The primary difference between use cases is the existing content of the domain-specific information base used as training data for the currently non-existent base model, using currently insufficient training tools.
The benefits of research and development (R&D), driven by these common needs, would accrue to a broad swath of government and NGO usage across many specific information domains and use cases. To take just one family of example, an NL agent to help people navigate government benefits programs, kind of like what ACA “navigators” do person-to-person, but for any government program where beneficiaries need help with eligibility, troubleshooting, updates — which is to say nearly every program; there never will be enough people to “navigate” for all of them. In other words, scaling up assistance to people, far beyond the scale of available government staff to deliver it.
So, if the R&D could yield broad ranging benefits, why isn't it being done?
Likely because the work is not easy, and not obviously profitable in terms of return on the research efforts’ investment. Government organizations might have the motivation for the research — more on that in a moment — but lack the capability. Tech-titans have the capability, but not the motivation because of the lack of a profit model. Non-titan companies are limited in their capabilities too, and although that may be changing, the profit motive is not. For a deeper dive on that point, read “The Economic Case for Generative AI and Foundation Models” from top-tier venture capital firm Andreesen-Horwitz. Their article explains how smaller companies might be able to get in the AI race better, though explained from the perspective of how companies can make more money from AI, and for their investors. That’s not so helpful for anyone who wants to do not-so-profitable government/public-benefit AI research, but at least the essay suggests it is possible. And they got it right on one critical common factor:
“Many of the traditional AI problem domains aren’t particularly tolerant of wrong answers”
Also known to mean, “AI mistakes, lies, and hallucinations.”
If it is understandable why the private sector isn’t doing this work, what about the public sector, government organizations, or government-funded R&D? Shouldn’t the US government be driving the research agenda, instead of the tech titans? Well, yes; yes it should, yet the ability to do so is limited.
Here’s a recent case in point. Every US government agency was the target of a sort-of all-hands memo jointly from the Office of Management and Budget (OMB) and the Office of Science and Technology Policy (OSTP). The memo outlined seven highest-priority technology agendas and directed each agency to include in its annual budget request some funding for R&D on how the agency could pursue these agendas. However, of the 7 key goals, only two (2) are about specific areas of R&D, and one of those is AI (the other is health). Out of all the kinds of tech R&D that could benefit the government, AI is (literally) top of the list for funding priorities, according to OSTP.
However, you might well imagine that such would be scanty, the probability of any federal agency having the ability to perform the R&D themselves is very near zero (notwithstanding certain agencies involved with national defense and security apparatuses), and the ability of most agencies to operate an AI R&D program on their own is also zero.
But OMB and OSTP at least got it right on common needs; for example, very low tolerance for inaccuracy of NL AI output and accuracy being more important than entertainment value. OSTP’s goals for AI were well-stated on the government’s unique and essential roles, and the requirements for technology to foster three goals in particular:
“Using AI technology to better deliver on the wide range of government missions;” which I say includes AI-based effective government-to-citizen communications;
“Mitigating AI risks;” which in my book includes AI-based NLAs lying, hallucinating, and being used to generate disinformation; and
“Advancing solutions to the Nation’s challenges that other sectors will not address on their own;” which is exactly the point about how it is not going to be profitable R&D in AI to meet government needs.
Or to put it another way, OSTP and OMB acknowledge that the for-profit sector on its own isn’t going to be motivated to meet government needs unless the government takes a hand by fostering R&D to meet those needs.
Lastly “trustworthy” is a goal too, clearly an acknowledgement that current NL AI technology is not!
In other words, this memo is a valuable acknowledgement of the current situation, and work required for improvements that benefit government use. But is little more than an exhortation by OMB and OSTP for government agencies to articulate their own R&D agendas (which I expect that few agencies would be capable of saying) and request agency funding for it. Count me skeptical, but even if substantially increased R&D funding were forthcoming from Congress, I don’t believe that a plethora of agencies’ R&D programs are going to make a difference.
What might make a difference? Pick a single R&D problem for which solutions would meet needs common across many agencies. You know my suggestion: domain specific NLAs (as I defined them in my previous installment) feasibly built for, and operated by, any agency that wants one for their domain.
To conclude, if you reflect on the needs I’ve been describing in these past four article posts, and combine that with the venture-capitalists’ view and the advice of OMB and OSTP, three things are clear:
Government and NGO needs include trustworthy NL AI for domain specific NLAs;
For the foreseeable future tech-titans are not working on that, nor other for-profits; and
There’s little hope at present for tangible government efforts to the needed R&D moving along.
In other words, the impetus will need to come from technologists working for public benefit, but without expecting help from government organizations.
At the OSET Institute, we know how that kind of technology development works, but unlike election technology, where we know a thing or two, generative AI is new for everyone. So, the first steps are to get engaged with many organizations with a stake in AI for elections, and to work out an R&D roadmap that’s detailed enough that it could be funded. And we’re on it. Stay tuned.
In any event, the bottom line seems to be: “Seatbelts, everyone!” 🤓
