The danger is not only synthetic lies. It is also a public trained to doubt everything.
A lot of people talk about AI and democracy as if the biggest danger is obvious. For example, a deepfake goes viral, fake recordings spread before an election, or a false image confuses voters. Of course, those risks are real concerns, but after reviewing materials from an ethics cohort that I’m in (more on that later), I keep coming back to different questions. One of the questions is what gets trusted (probably why you clicked this post in the first place).
To extrapolate a bit, I keep thinking about what gets trusted, in terms of whether we are looking at things like data privacy, party platforms, policies, and/or regulations in the “right way(s).” I know polarization doesn’t help, but what if confusion had the final say?
Because if the answer to what gets trusted becomes “nothing,” then the damage goes beyond one bad piece of content. It starts to eat away at the conditions people need to participate in democracy at all, including whether freedom has to be made real or is simply assumed.
That is what this week’s post is about.
So, what gets trusted?
As many of us know, trust is not only about whether something is true. It’s also about whether people believe something can still be verified, if institutions still seem credible, and whether the public still thinks there is a shared baseline of reality. If this seems abstract, stick with me for a moment.
The above questions are among (some of) the reasons I do not believe AI’s “democracy problem” is simply about broad disinformation. It is also about what happens when panic over something like AI becomes so widespread that people begin treating all media, evidence, and institutions as potentially fake… by default.
That kind of collapse is not neutral by any stretch. In my opinion, it not only creates more polarization but also a climate in which accountability weakens. That’s a bad combo. As a result, bad actors with influence could have more room to maneuver their discourse, and we ordinary folks (from any part of the political spectrum) are left trying to interpret everything (through a lens of suspicion).
AI panic can become a democracy problem, too.
Nathaniel Persily makes that argument clearly in Misunderstanding AI’s Democracy Problem. He pushes back on the idea that democracy will be destroyed mainly by a flood of synthetic content persuading masses of voters. Instead, he argues the greater long-term risk is that fear of AI-generated falsehoods will erode trust in the rest of the information ecosystem. He also warns that people may end up in a nihilist bind, where they no longer know what to trust, but we can save that for another post.
The point around trust erosion in the information ecosystem matters because the biggest democracy problem may not be that one fake video “tricks” millions of people into believing the same thing. Truly think about it: the chances are extremely slim that any single piece of content could ever sway our individual beliefs en masse and dramatically change our perspective on a single issue.
A different kind of threat
However, that is something we could face, but perhaps not in our current interpretation. It may be that, over time, panic surrounding AI-generated content convinces enough people that nothing online is reliable anymore. Once that kind of suspicion settles in, it becomes easier for harm to start looking normal inside a system. Then, over time, if more people believe that than not, we’d, by default, normalize that perspective…into real life. Note: I italicized ‘online’ above because how do most people consume content from the information ecosystem? Online (well, at least as of 2025)!
The point is, once that happens, every institution gets weaker. If it doesn’t start with institutions, then at least every correction could look partisan, or every verification would look suspicious.
The liar’s dividend thrives in that environment.
Persily also pointed to the liar’s dividend, which is where the mere existence of deepfakes gives political actors more room to dismiss accurate content as fake. In other words, even when something is real, people can cast doubt on it simply by invoking an “AI excuse.”
The liar’s dividend stuck with me this week, and perhaps even made me lose some sleep. Not from sheer paranoia or anything, but thinking about how this issue goes deeper than media or information ecosystem literacy alone. For example, if every recording or image can be dismissed as fake, or at least written off as questionable, then truth does not disappear. It becomes politically optional.
If you don’t think this is an issue, perhaps take note of how folks are beginning to interpret whether an image or YouTube video is AI-generated. Actually, this weekend, I was at a dinner, and someone shared a political advertisement for a local candidate. The photo showed the candidate in a superhero costume, trying to resonate with constituents. Working in this space, I immediately saw a red flag. The folks at the dinner who don’t work in tech or aren’t tech-adjacent seemed to see the image differently.
Is this what it looks like?
From my perspective, the image was clearly AI-generated, likely with HeroMe or using Reface filters (not suggestions or adverts). However, the folks analyzing the image were picking it apart. They noted that the “superhero’s” crest was above the zipper but also blended into it, and that the shadowing on the stitching around the shoulders was incohesive.
I don’t recall those types of analyses happening 10 years ago, and definitely not 20 years ago. Nonetheless, we are looking at something that changes the terrain we are used to. Now, instead of arguing over what happened or what was said, people argue over whether “evidence” itself still means anything.
And that is where democratic trust begins to fracture more seriously.
This is not just a content problem. It is a legitimacy problem.
That distinction stood out to me because it gives space to potentially new, or at least unfamiliar, forms of authoritarian creep in our workspaces and beyond. A lot of public conversation still treats AI and democracy as a content-moderation issue, a deepfake issue, or a problem of catching bad information fast enough.
Those issues matter, of course, but Persily’s argument is broader. He suggests that public responses to AI may do as much damage as the underlying synthetic content, if they cause widespread distrust of media and institutions more generally. As a side note, if you read Timothy Snyder’s book (not an advert), On Tyranny, I immediately thought of chapters 2 (defend institutions), 10 (believe in truth), 11 (investigate), and 17 (listen for dangerous words).
Bias is not a side issue either.
Persily also makes a useful point about hallucinations and bias. He argues that these systems are not simply neutral; they generate responses based on training data, product design choices, and constraints set by the companies that build them. For example, once a chatbot gives a single answer instead of a set of links, the company is making more substantive choices about what gets surfaced, how, and according to which values (this one made me think of On Tyranny’s 9th chapter around “being kind to our language”).
This means we are not just talking about falsehoods anymore, but about whose judgments, norms, and/or values are embedded into a tool. Once those choices are embedded deeply enough, it becomes harder to tell what gets tracked and what disappears into the system. Of course, that does not make every answer propaganda, but it does mean these systems are not above politics just because they are technical.
And if one set of corporate priorities begins shaping how millions of people access information (hello oligarchy), that matters for democracy too. Which is another reason governance cannot be an afterthought once those priorities start hardening into infrastructure.
Competition and openness do not solve everything.
Another part of Persily’s essay that stood out to me was his discussion of competition and open models. At first glance, openness can sound more democratic (example: how low-code or no-code platforms “democratize software development”). More people can build, customize, and use the tools. Power gets distributed more broadly. That sounds good in principle.
But he also points out that openness creates risks that are hard to reverse once a model is already out in the world. The same broader access that distributes benefits can also distribute harms.
We will not “tech” [sic] our way out of this.
This may be the part I agreed with most. Persily argues that “we will not tech [sic] our way out” of AI’s challenges to democracy. He specifically warns against overreliance on technical fixes alone, such as watermarking and provenance systems, because even useful tools like those do not solve the emotional, psychological, and political dimensions of distrust.
That rings true to me.
Public capacity matters more than panic.
To hold true to my blog’s theme (and good old reality), if we cannot rely on technical fixes alone, then what?
Glad you asked, because Persily argues for something more practical: better accountability, stronger administration, more public investment, independent auditing capacity, and systems of transparency that allow outside scrutiny before deployment. That kind of work does not begin only at the top; it is also part of how to defend rights from where you are. He also stresses that governments need greater institutional capacity and expertise to govern these technologies effectively.
That is a much less flashy answer than the ones people usually want, but I do believe it is the better one.
At the end of the day
AI can produce false content, and it is important to recognize that. But a society that loses confidence in its ability to verify anything may face an even deeper problem. It’s why I think the next question is not only what gets enforced, but what gets trusted. The two are connected because trust weakens fast when safeguards stop holding in practice.
After all, once trust across the entire information environment begins to collapse, democracy becomes easier to manipulate even without a single decisive deepfake. And when that happens, the damage does not come only from the lie.
It comes from the public learning to doubt everything around it.
P.S. If you find Nate Persily’s work interesting, check out his new book (not an advert or affiliation): Artificial Intelligence, Politics, and Political Science