AI’s Promise Is Real. So Is the Threat No One’s Talking About.

Beyond deepfakes, bad actors are learning to manipulate search itself - shaping what AI sees, what it surfaces, and ultimately what people believe.

AI is not a side story anymore - it’s becoming core infrastructure.

The pace of research, the expansion of capabilities, the way it’s reshaping how people find information, make decisions, and engage with the world - that’s real progress. Used well, AI can accelerate learning, expand access, and surface information faster than any system we’ve had before. It can make local voices louder, help small organizations compete, and give people tools that used to only exist inside large institutions.

But that same shift is creating a new vulnerability that too many people are still underestimating.

We’re watching a fundamental change in how people find and trust information - and it’s not small.

For years, search meant platforms like Google. You typed a query, got a list of links, and decided what to trust. It was imperfect, but the user still had agency. You could compare sources, click around, and see the messiness of information in real time.

AI changes that. Both in terms of stand alone chat bot answers, but also web search engines like Google increasingly pushing “AI answers” into search results itself. New data from Advanced Web Ranking found that Google’s “AI Overviews” now appear in 59.7% of all search results, a trend that is only growing quickly.

AI doesn’t just point you to information- it interprets it for you. It pulls from what it can find, compresses it, synthesizes it, and delivers it back as a clean, coherent answer.

That’s a completely different dynamic.

Because AI isn’t just reflecting reality- it’s assembling a version of reality from which it retrieves and spits out to the end user..

And when that constructed reality includes misinformation at scale, that misinformation doesn’t just sit on the margins like it might in a search result. It gets absorbed into the synthesis. It becomes part of the answer. Part of the narrative. Part of what feels like truth.

That’s the difference:

AI can bring the Firehose of Falsehood into search

Search shows you information.
AI tells you what the information means.

And increasingly, people aren’t starting with search anymore, they’re starting with AI engines.

We are watching, in real time, the emergence of systematic search manipulation - not just misinformation, but coordinated efforts to shape what AI systems and search engines retrieve, prioritize, and repeat. And studies like this one shows how the content that AI systems train on is dramatically skewed toward the right.

This isn’t about one bad post or one viral lie. This is systemic.

It’s about seeding entire information ecosystems with content designed to be picked up by algorithms and AI models. It’s strategic. It’s repeatable. And it’s increasingly effective.

The playbook is straightforward:

• Create volume

• Repeat simple frames

• Exploit how systems rank and retrieve information

If enough content says the same thing, structured the same way, across enough surfaces, it starts to look like consensus - to both people and machines.

That’s the shift.

We’re moving from a world where misinformation spreads socially…
to a world where it gets built into the information layer itself. Into the core websites, social media, books, and other content that trains these AI models themselves.

This is the next evolution of the “Firehose of falsehood” - but instead of overwhelming people directly, it overwhelms the systems people rely on.

This is not a cybersecurity problem.

It’s an infrastructure problem.
It’s a narrative problem.
And it’s a power problem.

Because the groups that understand how to seed systems—how to create what I call retrievable truth—are the ones that shape what everyone else sees.

Right now, bad actors are already doing this:

• Flooding low-quality sites with coordinated narratives

• Using AI tools to mass-produce content at scale

• Exploiting gaps in local news and “attention deserts”

• Testing which language gets picked up—and doubling down on it

So what do we do about it?

First, organizations need to stop thinking of content as one-off outputs and start treating it as infrastructure.

That means shifting from:
“We made a piece of content”
to:
“We built something the system can actually use.”

It starts with language.

Not clever language - durable language.

Clear, repeatable phrasing that defines an issue the same way every time. One idea, one frame, one set of words that show up everywhere.

Because AI and search don’t reward novelty.
They reward consistency.

If your language changes every time you talk about an issue, you’re essentially erasing your own footprint.

Then it’s about distribution - but not in the old sense of blasting something out once. Especially influencing key platforms, social media like Reddit and Wikipedia among many others.

You need to publish in layers.

The same core idea should exist as:

• a short video

• a written post

• a caption

• a transcript

• a headline

• a quote

Each version reinforces the other. Each format becomes another entry point into the system.

You’re not repeating yourself.
You’re increasing the chances that your version of the truth is what gets picked up, summarized, and surfaced.

Consistency over time is what turns content into infrastructure.

One post doesn’t matter.
Ten disconnected posts don’t matter.

But sustained, aligned repetition does.

When the same facts and frames show up again and again - across platforms, across voices, across formats - they start to register as signal, not noise.

That’s what both algorithms and people learn from.

And none of this works without messengers people actually trust.

Not influencers in the abstract- specific, local, credible voices who already have relationships with their audiences:

Teachers.
Veterans.
Parents.
Workers.
Community leaders.

People who don’t just deliver a message once, but carry it over time, in their own voice, in their own networks.

That’s how information moves in fragmented environments.
Not top-down.
Network to network.

Second, we need to invest in monitoring and rapid response.

That includes:

• Tracking how issues are showing up in search and AI outputs

• Identifying coordinated manipulation patterns early

• Running message testing to understand what’s breaking through

• Responding quickly with accurate, structured content that can compete

If bad information is seeding the system every day, you can’t respond once a month.

Third, we need to build reporting and accountability mechanisms that match the scale of the problem.

Right now, reporting misinformation is fragmented and reactive.

That’s not enough.

What’s needed:

• Standardized ways to flag coordinated manipulation patterns - not just individual posts

• Partnerships between researchers, platforms, and civil society

• Responsible content creators need to license and provide their content selectively to responsible AI platforms - and not reflexively block everyone.

• Peer reviewed, trusted means to evaluate potential ideological skew in new AI systems

• Escalation channels when AI outputs are being systematically distorted

• Transparency from platforms about how they’re addressing large-scale manipulation, and which specific data sources they train new models on, and means they fine tune those models

And for activists and organizations, this means documenting what you’re seeing:

• Capture misleading AI outputs and search results

• Track repeated phrases across sites

• Map where narratives originate and how they spread

• Share findings with journalists, researchers, and watchdog groups

Because sunlight still matters -
but only if it’s applied at the system level.

The bottom line

AI doesn’t create truth.
It organizes and amplifies what’s already there.

If you don’t shape the inputs, you don’t control the outputs.

And right now, that’s the real battleground.

More resources:

The New York Times on deepfakes

UK Electoral Commission launches deepfake detection ahead of May election