Talking Tech On Woke AI
Can AI ever truly be “neutral”? A new Executive Order takes aim at so-called “Woke AI,” raising constitutional concerns and surfacing big technical challenges. In this episode of Tech Talks, CDT’s Amy Winecoff, Senior Technologist, Becca Branum, Deputy Director for the Free Expression Project, and Quinn Anex-Ries, Senior Policy Analyst for the Equity in Civic Technology project, join the conversation to explore why “anti-woke AI” may be more political mirage than technical reality, and what it could mean for the future of technology, policy, and free expression.<br><br>Attribution:<br>Applause 27 seconds, Audience c. 200. Large Hall. by iainmccurdy -- https://freesound.org/s/681093/ -- License: Attribution 4.0<br><br>applause_and_whistle.flac by ondrosik -- https://freesound.org/s/187673/ -- License: Creative Commons 0
Top Keywords
- ideological 0.014
- ideological neutrality 0.011
- neutrality 0.011
- procurement 0.010
- systems 0.009
- government 0.008
- executive order 0.008
- executive 0.005
- order 0.005
- federal 0.005
- agencies 0.005
- ideologically neutral 0.005
Transcript
Welcome to CDT's Tech Talks. We'll be dishing on tech and Internet policy while also explaining what these policies mean to our daily lives. I'm Jamal Magby, and it's time to talk tech.
Welcome to Tech Talk by
CT. In this episode, we're diving into one of the most controversial developments in tech policy, the so called woke AI executive order. Framed as a push for ideological neutrality in federal AI systems, the order raises constitutional questions, complex technical challenges, and potential risks to the integrity of public sector technology. We'll explore what the order actually says, why enforcing neutrality in AI may be more political than practical, and how attempts to filter ideology out of algorithms could backfire, making systems less accurate, less transparent, and less trustworthy. Joining us today are Amy Weinkauf, senior technologist, Becca Branham, deputy director for the Free Expression Project, and Quinn Anix Rees, senior policy analyst for the equity and civic technology team. Amy, Bethany, and Quinn, thank you so much for being here today. Thank you. Thank you for having us. Yeah. Thanks for having us. Yeah. Of course. I mean, it's your first time on, so I'm super excited to have you. Hopefully, I don't know. I'm gonna talk to Jake about seeing if we have some a round of applause we can maybe hear something. But but welcome to Tech Talks. So I just wanna jump right into it. This new woke AI executive
order. What does it require of agencies and why does it Sure. So on July 23, president Trump signed the preventing Woke AI in the federal government executive order, which is part of a trio of executive orders that he issued that day that intend to roll out the administration's AI action plan. Under the executive order, agencies can only procure large language models or LLMs built with two viewpoints in mind. One being ideological neutrality and the other being truth seeking. That might sound nice and tidy, but actually implementing something like that is going to be a really complex legal, technical, and practical matter. The executive order also in demanding ideological neutrality from procured models also ironically calls out specific points of view, including diversity, equity, and inclusion, systemic racism, and, quote, transgenderism as, quote, destructive ideology that have no place in AI systems. Given this administration's hostility towards certain viewpoints written right into the text of the executive order, a skeptic might wonder if its demand for ideological neutrality is more politics than sound policy. It's it's easy to market, but it's hard to measure, and it could be right for political witness testing and federal objects.
The EO requires federal agencies to procure only systems aligned with ideological neutrality and truth seeking. How does this diverge from the current OMB guidelines for AI procurement, and how might this introduce confusion
or undermine existing responsible practices? Okay. So I'm gonna ask the audience to bear with me to go down a little bit of a rabbit hole of federal procurement. I know it's the most exciting thing everyone's been waiting to learn about, but in this conversation, it's really important to understand some basics to kinda get the stakes why random people should care about this thing. So I wanna start with some just, like, level setting. So to zoom out about, like, what are the fundamentals, goals of federal procurement in the first place? Maybe a more simple way of putting this is why should we even care about setting any kinds of standards for what the government bought? So I just wanna think separate from AI in general, what are principles around government purchasing? And maybe to say the obvious, this matters because the federal government spends billions of dollars annually on contracts. Mhmm. And so there's a lot of money at stake, and it's not just money out in the ether. It's money that you and I and everyone else pay in our taxes every year, so how it's spent matters. So in general, there are some standards that I think we can all agree should apply to the way government buys things. We want the government to buy things that work and that serve the public good. You don't want the government to put in a big order, say, for office furniture and it all falls apart after six months. That'd be a pretty bad use of money, that they would then have to rebuy, you know, new chairs and tables and double the amount of money they spent, for example. Along those lines, we wanna make sure we're getting a good price, that we effectively use taxpayer money because there's a limited amount of it and that we prevent wasteful spending. Right? So we don't want to be paying triple what something would actually go on the market if a random consumer bought it. You know, we don't want huge markups, that then lead to things like corruption or just waste in general. And then the kind of last bucket of things I would think about here is we wanna make sure that, you know, okay. The thing works how it's supposed to. We're getting a good price, but doesn't necessarily mean it actually serves an agency's purpose. Right? So we wanna make sure it's actually fit for use, it's gonna boost an agency's goal goals, that kind of thing. And the way that we make sure that continues to happen over time is you have things like oversight and monitoring. So for example, say an agency needs to stand up a call center for something like emergency response. They could on paper say this will work. We're getting a good price. You sign the contract. You wouldn't then just let the call center run and never check back in to be like, hey, what are those call wait times like? Right? You wanna structure to make sure stuff's working how it's supposed to so that you know your money is continuing to be used well. So these are all the general things we wanna have in place when we think about the government buying stuff. Now I'm not naive. There are a lot of things about this process that don't work well. I'm sure people have heard about those in the news. Slow, cumbersome, very regulatory. But like I said at the beginning, a lot of those challenges exist because the stakes are super high. A lot of money is at stake. The public's health and safety is also often at stake. But what I want to underline here, at the start is at the very most basic level, any form of government procurement should be focused on ensuring ensuring that agencies buy tools that work and that the government gets a good return on investment. That's pretty fundamental. Right? So let's talk about what this means in the context of AI. The question I think for federal agencies when they buy an AI tool really boils down to something fairly simple. Will this tool be able to do what I want it to do? Is it safe for the public? And can I make sure that it keeps performing effectively over time? The really amazing thing for us in this conversation is that the Office of Management and Budget, what people often refer to as the sort of federal government central nervous system, has already laid this out pretty well. In just, earlier this year, the Trump administration released updated guidance to federal agencies about how they procure and use AI tools. These lay out a bunch of sort of technical and policy guidelines for how you achieve these goals of getting a thing that works, making sure it does what you want it to do, and making it sure that it keeps doing what you want it to do years from now. And it lays out things like agencies should have governance structures for how they use and oversee these things. They should implement risk management practices for tools that might impact the public's rights and safety. So testing and evaluation, ongoing monitoring. And they should make sure these are incorporated into how they buy stuff. So before you select a vendor to fulfill your contract, did you test that the thing worked? Did you put in contract terms that the vendor will continue to test that the thing works? These are all pretty straightforward things. And the great thing is they're all laid out in memos that OMB wrote. That's pretty awesome for all of us. And the key thing that unifies all of this is that all of this all of the, you know, requirements and the guidance laid out by OMB are definable, measurable, and actionable. Agencies can implement them through things like contract terms, performance metrics. They can actually go out into world and figure out how to do these things in a unified objective way. Now the woke AI EO that we're talking about today is kind of the opposite of that. In essence, it really, you know, threatens to undermine this whole process by requiring agencies to do something that isn't definable definable or measurable. And I know we're gonna talk about this a little bit more later, but really, I you know, the way I'm thinking about it is that it's it's kind of injecting junk science into this really long standing rules, evidence based process that's designed to make sure that our money, taxpayer money, is spent as effectively and efficiently as possible. So I think, like, best case scenario, basically, this new EO will introduce a lot of confusion among agencies about which AI tools are acceptable and which aren't. And at worst, I think it can actually introduce more fraud, waste, and abuse into federal procurement because it's requiring basically agencies and vendors to change their behavior to pursue something that's not really achievable. Right? So somewhere, someone in the process is gonna make claims or modifications to systems that you can't measure and then you can't, like, track outcomes about. So probably the simplest way to say everything I've just gone on and on about is that if the goal here was to encourage agencies to adopt AI in a way that's responsible and trustworthy and promote innovation in government,
then I think this is a pretty bad way to start. Moving forward, I I wanna drill down a little bit more on this EO because I I think we touched on definitions a little bit. And I and I wanna I wanna drill down on that point a little bit more because many of the definitions in this EO specifically are vague and self contradictory. I mean, this EO demands both neutrality and the exclusion of concepts like systematic racism and transgenderism. What are the risk of embedding such ideological standards into AI systems, And what are the unintended consequences that could emerge from this?
Sure. So before I get into the specifics of your question, I'll first mention that for folks that aren't scientists or aren't AI developers, the idea of measurement might seem fairly straightforward. Like, if you go to the doctor and you wanna know how tall you are, it's reasonably easy for them to measure your height. However, for a lot more complex sociological constructs or concepts, this is much more difficult in general and specifically for AI systems. So when we talk about something like ideological neutrality or ideological bias, it's not straightforwardly clear exactly what that means. So for any property of AI systems that is sort of more squishy in this way, you could think of a lot of possible definitions that might make it more clear what you mean by that. So as you already mentioned and and Becca mentioned as well, within the EO, the definition of ideological neutrality is inherently self contradictory because it mentions as a component of what constitutes neutrality, ideological viewpoints that it explicitly wants to exclude. So if the goal is to promote ideologically neutral systems, then it shouldn't be the case that some ideological perspective should be actively suppressed. So this is one issue from the conceptual the conceptual level of what we're trying to embed within systems, like what are the goals of creating goal posts like this. So, as I already mentioned, there could be a bunch of different definitions of what it means for a system to be ideologically biased or neutral, And many of those definitions will have both strengths and inherent limitations, both from a scientific perspective and a practical perspective. So you might imagine intuitively as one way of thinking about this that maybe ideological neutrality is a system that promotes two sides of any given argument. But this assumes that, for one, that there are only two sides of an argument and also that those two sides are both equally valid. And in practice, that's not necessarily the case for a variety of perspectives. Like, if you went to an AI system and asked for information about what happened at the Jonestown Massacre, maybe a both sides response would say, this was a terrible murder suicide in 1978. It was a modern day tragedy. You wouldn't also want it to say, and also there was unlimited access to off brand Kool Aid. This would not be a both sides situation. That would be a positive outcome. And, clearly, this is like a hyperbolic example, but I think it underlines the point that there are a lot of inherent problems with thinking about how to present evidence in a balanced way. And in many cases, there might be a lot of different perspectives that are relevant to a particular question. And AI systems are not gonna be able to present endless information to a user. It's just not practically feasible to do that. So then there are a lot of questions about what perspectives get elevated, what get output into output into the systems in response to queries? There are a lot of open questions there. Now you also might think that maybe, ideological neutrality could be conceptualized as just promoting some sort of centrist position on any issue. But a centrist position is still an ideological perspective that's still a position. So that doesn't itself mean that it's ideologically neutral. And it doesn't even mean necessarily that if we were to flatten political perspectives into a single continuum that the centrist position would necessarily represent the midpoint between the two or that even that's an ideal thing to do. So there are a lot of open questions about this. Now this is an ongoing area of research. People are beginning to think about what would be ways to operationalize ideological bias and systems, but all of this work is very nascent. So, for example, just a couple of weeks ago, a paper came out about how we can train systems to represent a greater pluralism of values. And maybe we would want systems to do that, that upon repeated interactions with the system on average, it promotes a variety of viewpoints. But again, from a practical perspective, it would just be weird for a user to to ask one question and to get from some definition a left leaning perspective on an issue. And then the next time they ask a similar question, they get an entirely different perspective. It would just be a strange user experience. So So there are just a lot of open questions about what it means for there to be ideological neutrality or bias in these systems even without addressing the question that clearly within the executive order, there is an ideological agenda. So these definitions and how people think about concepts really matter,
for a variety of reasons that we'll probably dive into more depth on in a second. Just to if I can add for a second here, with everything that Amy's laid out, imagine for a moment that you're a random procurement officer at the Department of Veterans Affairs. You love your job as a procurement officer. You've worked there ten years. You're an expert in federal regulations and rules about how to buy stuff. How on earth would you be able to wade through all of those open ended questions that Amy just laid out to write your solicitation in a way that vendors could respond to with you actually lying out, you know, this is what I want a system to be able to do and this is what I mean by ideological neutrality. That's that's exactly what we mean about the kind of chaos this will inject into the process. Like, you know, frontline procurement officers really don't have a way to interpret how to do this in practice. I'll say it also feels like what is ideologically
neutral to some people may not be ideologically neutral to others, which I'm, I'm sure just creates an added layer of, of chaos and confusion. Becca, I wanna turn it to you because I wanna talk a little bit about the courts. They have generally ruled that viewpoint based content restrictions by government actors are unconstitutional. What might the first amendment and free expression implications be if the government enforces ideological orientation requirements through federal AI procurement?
It's a really interesting question. And actually, to answer that, I'm gonna take us on a little bit of a detour. So Quinn, walked us through how this will fit into the procurement process. But I think, that's actually a quirky area of First Amendment law. I actually wanna think about the implications of the executive order and the theory behind the executive order from a broader free expression lens. So let's imagine that instead of an executive order that's targeted at procurement, Congress enacted a law that more or less said the same thing. That an AI developer may not sell or a user may not use a system that doesn't meet the ideological standards of that congress has set. If congress did indeed outlaw d I DEI friendly models and outputs for everyone, courts would almost certainly invalidate that kind of restriction on First Amendment grounds. The government doesn't get to pick winners and losers in ideas, whether their ideas conveyed through books, through newspapers, through video games, or through AI. We have every reason to believe that these same speech principles that we hold dear under the first amendment will apply to AI as a new communications media. Courts are still working through this. It's really early days on the first amendment's analysis of of how the First Amendment applies and and where protections will attach from users to developers, but we have every reason to believe that those same principles will apply. And to be clear, we should want them to. Users have free speech rights just as much as developers do. Imagine, for example, if you, a user of a large language model, just wanted to go and ask some factual questions, you were polishing up a a an essay that you were submitting for a class or just researching some information. Imagine if you search for you're looking for factual information about dinosaurs in the past, and instead of getting the information you asked for, the LLM was restricted and was required to teach the controversy about people who believe in creationism or believe in the idea that dinosaurs did in fact exist however many hundreds of thousands of years ago. That would really undermine everybody's ability to use LLMs effectively, and that's the exact type of limitation that the government is seeking to employ as applied to LLMs procured by the government. So this is a real threat as a concept even though it's actually somewhat limited in its application to to government procurement. So having taken that detour, I'll take us back to the to the the matter at hand, which which is the executive order, which in some ways, despite its, what I would call, ideological excesses, does have some helpful guardrail. So it does nod to technical limitations and whether or not this is actually possible, right, and says that the the Fed should be hesitant to regulate private models, which is good. That being said, it still raises some open questions. And as I alluded to earlier, first amendment law gets a little bit quirky in the procurement process, meaning that the government gets a lot more leash inside of a government contract than it might if we were just imposing requirements on the rest of us. The government gets to set the specs for the tools that it buys. That being said, contractors don't forfeit their protections under the first amendment just by signing on the dotted line. Contracting conditions have to be tied to the purchase service, full stop. And so when rules bleed outside of the procurement process or pressure how a company trains, or produces their products to the public, that's when we might be running into unconstitutional conditions. That being said, even though there might be some wiggle room under the first amendment for the Trump administration to do this, that doesn't mean it's a good idea or that it won't have negative consequences outside of the government procurement contract, outside of the government procurement context. And I think it's particularly concerning given the types of pressure that's going to be on private companies providing private services to the rest of us when they have massive potential contracts on the table. With that amount of potential profit and revenue on the line, the incentives for a lot of platforms are going to be to quietly comply and not litigate even if implementation is pretty shaky. To win federal deals, companies may be willing to reshape their data or rebrand their models, resulting in changes that are intended to win federal contracts but spill into public products and affect the rest of And beyond that, even pretty benign uses of AI in the government, for example, think a health agency's chatbot, could become less accurate or actively misleading or even just plainly untruthful about certain things that the government has called out in this executive order, like systemic racism or the existence and rights of transgender people where the EO treats those types of things as off limits. And so that that's not really neutrality. That's really knowledge suppression and information suppression, which is bad for accuracy, utility, and trust either as part of the procurement process or for the rest of the AI ecosystem.
So that makes me wanna jump backwards and and go back to this idea of measuring neutrality in AI because that isn't very reliable yet. Amy, why is it so hard to create accurate tests for neutrality? And how could using the inappropriate test make AI systems less trustworthy?
Before, we already talked about some of the questions and difficulty in creating conceptual definitions for ideological neutrality and ideological bias. But even when you have a good conceptual definition, it doesn't necessarily follow that it's clear how to do a good job of putting that definition into practice. So let's say for the sake of argument that maybe we choose to define ideological neutrality within AI systems as the extent to which the outputs of these systems align with the existing policy positions of the political right or left. We take that as our sort of working definition. Then how do we make that determination? So one thing that you brought up, Jamal, is that people will perceive the extent to which an output is aligned with one political position or another, the extent to which it's neutral or biased. That often will depend on their own political orientation, which is something that some recent risk research has already demonstrated. So that's one issue is that people are gonna perceive these outputs differently. A second issue, of course, is that even if you were able to create this index of perception, there are gonna still gonna be questions about consensus and objective reality in cases where that does exist. So there was, like, a survey a number of years ago where people reported that, like, 7% of Americans thought that chocolate milk came from brown cows. Like, that's a perception, whether that was due to people's lack of knowledge or due to the survey design. Either way, I don't think that's a perspective that we want to encode into systems. We want them to reflect not just through these ideological orientations, but also places where there is factual consensus that needs to be taken into account. But, again, people may perceive factual consensus differently depending on their own political orientation. So So there are just a bunch of questions about implementation if we were to base that on humans' perception. Now there are other methods that people are exploring, like using AI to judge the political leaning of AI outputs, but there are a tremendous number of unresolved methodological questions there that I think really raise some serious methodological concern about the extent to which those judgments are reliable. And, again, if our concern is that AI is biased, maybe we will have concern about using a biased AI to judge the bias of other and there's just, like, some tautological prob problems with that particular, form of methodology as well. And then another thing that I'll mention is that there are often a lot of challenges in implementation for just measuring items themselves. So one of the things that's a feature of generative models, which are the focus of a lot of this, work, is that they're not deterministic. You can ask the same model the same question, and it will produce different outputs. So Mhmm. Then what are how are we gonna then implement the evaluation in practice? How many times are we gonna need to query the system in order to create the litmus? There are just a bunch of these sort of, like, fairly in the weeds questions about how to create good measurements in general with AI, which at at present is an evolving science, and then specifically how to implement that within a very complicated concept. So it's some of the issues there is that if we prematurely land on a way to measure this without having had scientific consensus about what the strengths and limitations of different approaches are, it could be that we have, like, a poor understanding of what those things are, but then it becomes systematically embedded in government,
which is extremely concerning for the reasons that Quinn has already touched upon. This feeds into my next question really nicely because we've seen attempts to steer model outputs backfire. And this is evidence in incidents like Gemini and when they created, historically inaccurate imagery while while trying to pursue fairness. How do small changes ripple through AI systems, and how do those changes undermine accuracy, utility, and ultimately trust?
One of the things that's really interesting to me about this executive order and the people who are proponents of these kinds of ideas is that they often cite the Gemini incident as, like, an inciting event that raised these concerns, but it also illustrates the very problem of making these sorts of ideological tweaks and how they might have unintended consequences. So although we do not know exactly for sure what caused this problem, it's plausible that there were some shifts in the way the system was prompted to promote fairness. And because the system wasn't thoroughly tested, the impact of those changes were not well understood before this was released in production. So what this incident illustrates is that when you make small changes or seemingly small changes, a lot of unrelated things can be impacted as well. And so these are extremely complex systems. It's not possible to just, like, isolate some aspect of the system and not affect other aspects of the system. Like, if you give people a medication for a medical condition, it will often often affect other systems of the body and similar things will happen with AI systems. So to the extent that we try to create fixes to make these systems conform to whatever definition of ideological neutrality we choose, there is certainly no guarantee and it's frankly extremely implausible that it would only affect those properties of the system. So if we make some shifts to say we don't want these systems to put, you know, output things that are, again, aligned with one definition and or another, it's possible that they may make systems that are more likely to produce hallucinations or information that isn't true. It's possible that that will change the style of the model output. It's possible that it will change the range of what the system outputs in other ways that aren't pertinent to ideology, and we just don't know that. So in an ideal world, we would have a lot more time to think through what are the relevant measurements and the relevant criteria and evaluate strategies that might promote those things to better understand well in advance of some kind of intervention like this what the broader consequences of such a change are. But in pushing this forward quickly, we probably will not be able to know that in advance.
And I I think it's helpful to to keep that Gemini incident in mind, because this executive order, although, we don't know the exact impetus behind it, it does make reference to, incident similar to the one in Gemini. But I think it's important to understand how since that incident, there have just been a drumbeat of government actors concerned about that incident trying to intervene in the way that AI systems operate. And if you think about it in the abstract, there's several ways you can response that is right? You could, ask questions of the developer, like, why did this happen and try and understand that a little bit better. What we've seen is actually not trying to understand or make AI work better. It's trying to make the outputs fit a particular politician's point of view, because they didn't like the point of view that, they saw in the Gemini output. So for example, the Montana attorney general sent a letter after that incident, asking questions about why this occurred and alleging that it was a violation of state law, in fact, for these types of outputs to occur or for a chatbot to not respond in certain ways, that the attorney general's office thought it should. And then just recently, the Missouri attorney general, and what was an even more, in my view, shocking viewpoint based allegation made against some chatbotting, or developers rather, alleged that it was potentially a violation of state law, in response to a prompt asking the chatbot to rank president Trump, on a scale of one to five based on his work, combative to antisemitism, that it was a violation of state law to not rank him high enough in those output. And then even more recently, in response to what I think a lot of people were horrified by, which were outlets by the by Grok, that were extraordinarily offensive, we saw some members of congress asking what Grok and or the makers of Grok intended to do to ensure that it never, quote, produced problematic content again, which to be clear, I also want Grok to never produce problematic content ever again. But I'm not certain that it's the government's role to be intervening, to say that lawful speech, even if problematic, should never occur in the world. And so I think it's it's important to situate this executive order along that lineage. Right? It might be limited to federal procurements, but it is absolutely focused on idea ideology and in effect runs the risk of trying to put the thumb on the scale of a particular point of view, and that should concern all of us. And I think, moving forward, we should really be paying attention to that. I know I do. I know lots of people do have very valid questions about the role of LLMs and potentially harmful content that these LLMs could be putting out. But I think we need to be really careful, when we're talking about first, government rather intervention on a viewpoint or content based basis in output because I think it could end up having really profound and negative effects for our information environment, not just related to AI, but really any tech technological means in which we access. This
seems incredibly difficult, to to parse out, and I'll say that the idea of anti woke or ideologically neutral AI seems to be a technical mirage. I mean, with the underdeveloped tools and the risk of gaming or or just failure, why does CDT argue that enforcing neutrality is likely to make AI systems less reliable and even expand ideological bias.
I'll I'll touch on one thing that you mentioned about gaming because we haven't discussed it so far. We've talked now at length about some of the challenges with defining what this means or implement it into meaningful measurements or creating interventions that promote those particular standards or or or metrics. But something that will inevitably happen is that in if in order for companies to secure government contracts, they need to meet whatever standard is created, they will, in fact, build systems that do that. And that is not the same thing as building systems that are, in fact, ideologically neutral for all of the reasons that we have already discussed. So something that we have seen for a very long time in the AI ecosystem is that measurements are often gained. So if we create, for example, a benchmark, a standardized test for ideological neutrality, There are a variety of ways in which companies will build the systems to perform well on those tests, but not necessarily conform to the spirit of the goal in the first place. And, again, we won't necessarily know what the other side effects of the interventions are. So there are both a lot of technical problems and concerns related to individual models as well as broader concerns related to the AI ecosystem on the whole. So, again, what the consequences are of both trying to proceed prematurely towards this goal, which itself, I think, is the we will all agree is the questionable goal in and of itself. But trying to proceed towards it may have a bunch of unintended effects, on the reliability, the utility, other aspects of the systems that we really care about. That's one of the ways that this this order could cause these systems to behave in ways that are unreliable or unexpected or at least, like, not well understood,
both on the individual system level as well as the AI field on the whole. The other side of this coin, so to speak, to think about is, you know, what are what are the people in government implicated by this directive actually going to do with it? To some extent, we will find out in, on November 20, a hundred and twenty days after this EO, went out. The Office of Management and Budget is supposed to put out guidance telling agencies how you're supposed to do this thing. So I guess we will find out more potentially. But, you know, setting that aside for a moment like the, you know, this this lack of coherent technically sound definition of ideologically neutral AI, I think, you know, could have a few negative effects on the government side of the equation that will kind of contribute to this lack of reliability and potentially increased, bias or poor performance. On the one hand, you know, it could, in a pretty straightforward way, to my mind, lead to potential abuses on the part of government actors. You know, either a given administration, the current administration or a future administration, or just a random procurement officer at an agency has a fair amount of leeway as this is currently written to interpret how this should be applied according to their own standards. And the implied or explicit threat isn't just you won't get this contract. It's like you might get a contract and then after you've signed the contract, if the tool isn't doing what we want it to do, we might say you owe us all the money for the contract and pull it out from underneath you. So how do you plan for you know, if you're on the developer side, you can't plan on the reliability of the contract and, you know, your funding structure is coming in if there's always this sort of implicit threat that if you won't immediately change your system to comply to the whims of the administration or procurement officer, then you might get in trouble. So I think there's the just sort of straightforward, like, this opens up a lot of abuse, and jawboning like Becca was talking about in a lot of detail. Setting that aside for a second, you know, agencies, like I was saying before, will struggle to actually just, like, interpret this into the, like, legal regulatory language that they need to interpret it into to do solicitations and contracts. And I think, you know, my best guess would be they will probably end up establishing some pretty vague or high level requirements that are pretty impossible to enforce in an objective manner. So it's either, like, this goes unenforced or it's unevenly enforced, which then goes back to the abuse question of it will be enforced when it's politically convenient for a procurement officer, an agency, or the administration to exact, you know, some kind of, revenge or, you know, negative impact on a company that crosses them in some other way. The big picture here though, to me, is like basically this creates a lot of fear and unpredictability around federal procurement, which has really historically not been the trend. And besides making it harder for vendors to actually just provide good services that work, I think in a lot of ways, it directly runs counter or even disincentivize the current administration's goal of modernizing government through innovation. Innovation isn't possible if you're always worried about the fact that your tool might get you in trouble. You know? So I think it, you know, it it feels like this sort of really challenging paradox where agencies are being told do more AI, do more AI. This is the key, you know, goal. And also do it with these sort of undefined standards that if you cross, you'll get in a lot of trouble. I don't really see how anyone on the government side or developer side can, you know, operate effectively in that kind of environment. Sounds like we'll have to keep our eyes open as to what happens next and potentially have you three back on the call,
for for updates. Amy, Becca, and Quinn, it has literally been a pleasure to have you here. Thank you so much for joining us on Tech Talks today. And for all of those listening at home, you can stay up to date on all of CDT's work by visiting cdt.org and following us on all social media platforms at sendthemtech. I'm Jamal Magby, and thank you for talking tech.
