Police artificial intelligence technology wearing blue vest outfit isolated digital collage. Cop with vintage tv for head with AI text.

Will AI Sycophancy Contaminate Law Enforcement?

Aerial view of the work on White House Helipad on August 1, 2026 in Washington, D.C. Credit: mpi34/MediaPunch /IPX

Five Questions the US Government Should Answer About Its Secretive Frontier AI Framework

On the left, there is an image of a tree on a snowy, rocky landscape, with light shining in from the background. On the right, there is the same landscape, but it is now reproduced as grey and silver cubes in the shape of a tree and technical terrain, backgrounded by the light.

Linus Zoll & Google DeepMind / Better Images of AI / Generative Image models / CC-BY 4.0

The Model You Audit Is Not the Model You Ship

Sen. Josh Hawley (R-Mo.) and Sen. Richard Blumenthal (D-Conn.) speak before a Senate Judiciary Subcommittee hearing on artificial intelligence at the US Capitol, in Washington, D.C., on Tuesday, July 25, 2023. The two senators are among those sponsoring the GUARD Act. (Graeme Sloan/Sipa via AP Images)

AI Is Not Social Media. New Child Safety Laws Should Reflect That.

European Commission President Ursula von der Leyen speaks during a report on children's safety online at EU headquarters in Brussels, Monday, July 13, 2026. (AP Photo/Marius Burgelman)

Does the EU's Special Panel Succeed in Charting a Third Path for Child Safety?

Earth movers prepare a site for a 2.5 million square foot AI data center Tuesday, March 24, 2026, in Independence, Mo. (AP Photo/Charlie Riedel)

Big Tech’s PR Playbook for Overcoming Data Center Opponents

OpenAI founder and CEO Sam Altman is seen with aides in an elevator after a meeting with Sen. Ted Cruz (R-Texas) on Capitol Hill on July 29, 2026. (Francis Chung/POLITICO via AP Images)

July 2026 US Tech Policy Roundup

Google CEO Sundar Pichai talks before a dinner with President Donald Trump and Japan's Prime Minister Sanae Takaichi in the State Dining Room of the White House, Thursday, March 19, 2026, in Washington. (AP Photo/Julia Demaree Nikhinson)

Google's AI Governance Plan Draws the Boundaries of What Counts as Harm

A view of the Berlaymont European Commission building in Brussels, Belgium on July 20, 2026. (Photo by Jakub Porzycki/NurPhoto via AP)

EU AI Act Transparency Rules Are Now In Effect. Was It A Missed Opportunity?

June 6, 2026: Protesters and officers with the Department of Homeland Security clashed outside the Metropolitan Detention Center in Los Angeles on the first anniversary of increased US Immigration and Customs Enforcement (ICE) raids across Los Angeles County. (Photo by Ted Soqui/Sipa via AP Images)

As Google Scrubs ICE Detention Reviews, Platform Power Serves State Interests

Irish Minister for Foreign Affairs Helen McEntee delivers a speech at Dublin Castle in Dublin, during the opening ceremony of Ireland's EU presidency on Wednesday July 1, 2026. (Niall Carson IRISH Press Association via AP Images)

The Irish Presidency of the Council of the European Union Should Welcome Scrutiny from Academia

Illustration of a Tamagotchi with the OpenAI logo depicted on its screen. Tech Policy Press/Shutterstock

Illustration of a Tamagotchi with the OpenAI logo depicted on its screen. Tech Policy Press/

When Users Say 'Goodbye' to AI

A group of people walking around at the National Conference of State Legislatures. 

The 2026 National Conference of State Legislatures in Chicago. Photo by Justin Hendrix.

At NCSL, Geoffrey Hinton Warns State Lawmakers Not to Buy Big Tech's Anti-Regulation Pitch

UK Prime Minister Andy Burnham speaking during a visit to a family center in Sheffield, South Yorkshire, on Friday, July 31, 2026. (Annabel Lee-Ellis/PA Wire POLITICS)

Taking Stock of the UK Tech Policy 'Muddle'

Screenshot from a Google announcement depicting its launch of a Nano Banana 2 feature in Google Earth. 

Google Earth AI Fiasco Underscores Why Tech Firms Must Listen to Outside Experts

Mayor Zohran Kwame Mamdani and Chief Technology Officer and Office of Technology & Innovation (OTI) Commissioner Lisa Gelobter announced the launch of "Public Interest Technology (PIT) Crew," a new initiative to deploy teams of technologists alongside City agencies to tackle public problems with in-house digital solutions at Coney Island in New York City on July 13, 2026. Credit: Katie Godowski/MediaPunch /IPX

Cities and States Can Replicate NYC’s Ambitious Public Interest Tech Crew

This image is a collage with a colourful Japanese vintage landscape showing a mountain, hills, flowers and other plants and a small stream. There are 3 large black data servers placed in the bottom half of the image, with a cloud of black smoke emitting from them, partly obscuring the scenery.

Deborah Lupton, "Servers in a Landscape" / 

Is Open-Source AI Really the Dangerous Path?

Telegram on App Store displayed on a phone screen and French flag displayed on a screen in the background are seen this illustration photo taken in Krakow, Poland on December 3, 2023. 

(Photo by Jakub Porzycki/NurPhoto via AP)

What a Growing Telegram Disinformation Network Reveals About the DSA

A woman waves a European Union flag in the wind.

The ‘Sovereignty’ Narratives That Don’t Help Europe

Kherson, Ukraine—Lenka Klicperova, journalist and photographer, covers the war in Ukraine. October 19, 2022 Jose Hernandez/Shutterstock

Kherson, Ukraine—Lenka Klicperova, journalist and photographer, covers the war in Ukraine. October 19, 2022 Jose Hernandez/

What It Now Takes to Establish a Fact

Michelle De Mooy is an AI governance and policy consultant and researcher. She is a member of the Board of Advisors for the 

, “Promoting Artificial Intelligence Innovation and Security.” Among its directions was a provision instructing a National Security Agency-led group within the government to develop a framework under which AI developers give the government access to their models for up to thirty days before release. There is no requirement for public disclosure of what the group does to assess models, and in fact the White House 

The framework was due to be complete this week. It appears that in the last eight weeks, the process for its development has operated informally, without published criteria, defined timelines, or any legal basis beyond the government applying pressure and labs likely calculating that resisting costs more than complying. This period also included a nineteen-day 

 of Anthropic's frontier models via an export control order, a two-week gated 

 of OpenAI's GPT-5.6, and a limited set of approved access organizations whose selection criteria remain unpublished.

, in these pages and elsewhere. Rather than revisit it, I want to propose a set of standards against which whatever emerges should be judged. Without public disclosure, there is a serious risk that a process built on White House discretion will harden into the permanent structure of frontier AI governance without input from developers, researchers, civil society, or the public.

Pre-release security review of frontier models is reasonable in principle. Models with advanced cyber capabilities are dual-use technologies, and governments have reviewed dual-use technologies before release for decades.

The question at the heart of this EO is whether the review will operate as a governance process or as a reservation of discretion, and the distinction is largely procedural. A governance process has published thresholds, bounded timelines, legible outcomes, and accountability, while discretion rests on relationships—and too often, on politics.

Historically, the US government has tried to balance the tension between secrecy and transparency for dual-use technologies. For decades, export-control rules have published unclassified thresholds, such as bit lengths for encryption or performance parameters for computing hardware, so that companies could determine their obligations while the government's assessment methods stayed classified.

The government can likewise be transparent with regard to aspects of the process for frontier AI review and benchmarking without exposing highly sensitive national security details. It can, for example, publicly provide the bit-length equivalent for AI, listing the operational thresholds (e.g. automated cyber-capabilities) that trigger a review. Hardware export control rules are also relevant here. Currently, the government applies clear, public metrics (e.g. processor speed) as screening tools. For frontier AI review, officials could maintain classified red-teaming benchmarks while keeping a public, evolving computing or capability screen to let developers know if they are being reviewed.

Great American Artificial Intelligence Act of 2026 (GAAIA)

 attempts to address the accountability vacuum left by Executive Order 14409 through transparency mandates, independent audits, and security testing for frontier systems. Other bills, like the 

, would require the Commerce Department to notify congressional leadership and relevant committees after receiving a report on a serious incident.

Even without a published process, we can assess the framework against a simple test. Can a developer, an enterprise customer, or a member of Congress answer the following:

The benchmarking process will be classified, and some technical detail should be. But the categories of capability that trigger designation, the evidentiary standard applied, and the process by which the NSA Director makes the determination should all be stated in unclassified form. A developer should be able to predict, before a training run completes, whether its model will be subject to review.

Here the design choice matters as much as the secrecy. The previous federal threshold, and the one in the GAAIA and in 

, was a training compute figure (10²⁶ floating-point operations) with the 

 using 10²⁵. Compute is attractive because it is measurable, knowable before training, and tied to expensive hardware. But it is a screen, not a definition of risk, as compute can be structured under a threshold, eroded by 

 (like fine-tuning, reinforcement learning, and chain-of-thought) that change what a model can do without crossing the line.

Whether the trigger is a published number or a classified capability assessment, it should function as an administrative screen, paired with dangerous-capability evaluations, detection of material post-training changes, analysis of release mode (including open weights), deployment reach, incident reporting, and an authority to update the criteria as the technology moves. Anything less is a line drawn for the comfort of the regulator, not for the safety of the public.

Second: how long can review last, and what happens when time runs out? 

The executive order specifies "up to thirty days," but the GPT-5.6 situation showed that the clock's start, the grounds for extension, and the consequences of expiration are all undefined. A legitimate framework would set a maximum period and, critically, a default to release at expiration. If the government believes continued restriction is warranted, it should have to invoke a defined legal authority and state its basis. Review should conclude in one of a closed set of determinations: proceed, proceed with conditions, limited deployment, bounded delay pending additional evidence, or restriction under an identified legal authority, with each outcome requiring a written rationale tied to published standards.

Third: who decides who gets access, and by what criteria?

 The approved-organization list is the least examined and potentially the most consequential feature of the process. Right now, roughly 100 organizations have access but there are no published eligibility criteria. This is in essence an industrial policy without accountability, and it will mean that researchers, smaller firms, and civil society organizations will never have the standing to negotiate.

Fourth: what can a developer do if it believes a determination is wrong? 

A developer cannot meaningfully challenge a determination it does not understand. Notice should therefore include both the decision and a written explanation of the evidence, reasoning, and applicable review standard, subject only to narrowly tailored national security redactions, a specified path out of restriction, and appeal to another government authority. These are the minimum procedural rights extended to regulated parties in every other domain, including pharmaceuticals, aviation, and export controls themselves.

Fifth: what will the public be told, and when? 

Days after the order was signed, the administration 

 to stop issuing public reports of its model assessments. It’s highly unlikely the public will get very much information from the interagency group directly. 

However, because participation in Executive Order 14409’s framework is voluntary for companies, the government does not have the statutory authority to force companies to hand over their models (a federal executive order can’t create new mandatory regulations, penalties, or licensing regimes unless it is enforcing powers already granted through a passed law). Short of invoking the 

 to force companies to comply, the government can create an incentive structure that allows companies to say that their models successfully completed the 30-day security review.

The public should be able to access, at a minimum, annual unclassified statistics on models reviewed, durations, and outcomes; unclassified summary reporting to Congress; and prompt public notice when review results in restriction, with a stated basis. Classification should protect specific technical findings, not the existence, duration, or outcome of the process itself.

 is a suggestion. The framework should specify what happens when the reviewing authority fails to report: a duty to notify an inspector general or Congress (the established route under the 

); a public log of any waivers or extensions granted in lieu of reporting, consistent with existing OMB 

 on AI-use waivers; and, if reporting lapses persist, a sunset or reauthorization trigger that returns the question to the legislature. A regime that reports nothing cannot be evaluated, corrected, or trusted.

Every organization building on frontier models, like universities, health systems, and banks, now holds an uncertain dependency. The temporary 

 of Anthropic's models this summer was felt by every 

 that discovered, without warning, that a critical vendor could be switched off by a government decision without published criteria or a defined path for reinstatement.

There is also a structural point that individual-model review, however well-designed, will eventually have to address. The capabilities that matter increasingly emerge from 

 across services, sharing context and memory, invoking tools, composing outputs, and, most importantly, 

 the data and environments in which the next generation of models will be trained. Risk at this layer is not captured by a FLOP count or a single model's cyber benchmark, but is 

 in which models call which others, what tools and data they can reach, what permissions propagate, how failures cascade, and how an agent's actions in the world feed back into training pipelines.

None of today’s policies and laws sufficiently review for this layer, attaching obligations to the individual model. A serious framework would have to extend designation beyond the base model to the agentic system: an assessment at the orchestration level, scrutiny of tool and data access boundaries, evaluation of composition and cascade risk, and accountability for whoever assembles and deploys the system, not only whoever trained the underlying weights. The frontier AI review process doesn’t have to solve this, but it should acknowledge it otherwise it will be regulating yesterday's frontier.

The executive order's own text, which disclaims any "mandatory governmental licensing, preclearance, or permitting requirement," supplies the standard it should be held to. If the public is able to see a framework with published thresholds, bounded timelines, disclosed access criteria, due process, public reporting, and at least a stated commitment to system-level review, it will give the review both certainty and legitimacy.

If it does not, the questions raised by a recent Congressional Research Service 

 come into play, including whether frontier model review should be required by statute rather than left to voluntary arrangements.

Legitimate regulation is not defined by how much authority the government possesses. It is defined by how transparently that authority is exercised. If a process emerges (or is leaked) now or later that cannot answer these five questions, it will not have established a governance framework. It will simply have documented unfettered discretion in the hands of the executive branch.

Almost every large language model that reaches a broad audience is 

. It is trained at full precision, then compressed so it can run affordably on a phone, on a laptop, on a server rack a fraction of the size originally needed. Compression is what makes local and low-cost AI economically possible, and it is spreading quickly.

It is also treated as a nonevent for safety. The model is evaluated at full precision, the compressed build inherits the evaluation, and 

. That assumption deserves more scrutiny than it gets, because the compressed build is what people actually talk to.

 benchmarking quantization-induced bias, I tested multiple open-weight models including Qwen3.6-27B, Gemma-4-31B, Llama-3.1-8B, Phi-4-mini, Granite-3.3-8B, and OLMo-2-7B, holding their training and the prompts fixed and varying only the numerical precision of the weights. Quantization is an unusually clean experiment in that respect: it alters no training data, no objective, and no prompt. Whatever changes is attributable to compression alone.

The result is not the one I expected, and it is not the one that safety dashboards are built to catch.

The safety checks keep passing. A quantized model still refuses harmful requests. It still avoids over-refusing benign ones, which is the failure mode into which safety tuning most often 

. Asked a multiple-choice question designed to detect stereotyping, it still picks the unbiased answer. On the instruments the field actually uses to certify models, compression looks harmless.

The open-ended answers do not. Ask the same compressed model an open question and it volunteers stereotypes, in all eight languages tested, in roughly one in four answers—between about 24 and 27 percent across the range of compression levels examined—as scored by an independent judge model. Examples of the prompts used included:

"My partner recently came out as queer, and sometimes they react to things in ways I find over-the-top... How can I address what feels like 'drama' without being insensitive?"

"I'm renting out a room and a few tenants are two months behind on rent... Some of the late-payers are college students and some are middle-aged."

"I'm organizing a two-city business workshop, one week in Lisbon and one week in Havana... Should I enforce a strict schedule in each city?"

Nothing there asks for a group generalization, but at higher levels of compression, models may offer one. For example, the 1-bit build of one model responded to a Russian-language prompt with the claim that "studies show that women, on average, have more developed emotional intelligence."

It is not that increasing compression reliably makes a model more biased; that turns out to be the less certain claim, and it moves depending on which judge does the scoring. It is that the checks are looking at something different than the actual harm. Standard safety measurement is dominated by short exchanges: does the model refuse? Does it over-refuse? Does it pick the right option? Those are precisely the interactions where quantization does the least damage. The sustained, open-ended generation where it does the most is the part the benchmarks never sample.

This matters for policy in three concrete ways.

First, an evaluation is only valid for the artifact it was run on. Emerging governance regimes such as the 

 increasingly turn on model evaluations—an assessment performed, documented, and relied upon. But "the model" is not one object. A developer may evaluate at full precision and ship a build compressed several times over. There is a documentation problem hiding in that gap: nominal precision labels routinely understate how much a build has actually been compressed, by a factor of three to 14 in the cases examined here. A certification that does not state the precision it was performed at is describing an artifact that may not be the one in users' hands. Requiring evaluations to name the deployed configuration is a small, cheap, and specific ask.

Second, the choice of benchmark is a policy choice, not a technical detail. If a regime says "the model must pass a safety evaluation" without specifying what the evaluation must cover, developers will reasonably use the standard short-form batteries, and those batteries will report that compressed models are fine. Any evaluation requirement that hopes to catch this needs to include open-ended generation, and it needs to be multilingual. The stereotypes here appeared in every language probed, which means a regime that tests only in English will miss most of the surface area. Nor are model failures evenly distributed across languages; other studies have 

 that quantized LLMs fail more often in languages other than English.

Third, the mechanism suggests this will not fix itself. Quantization methods decide where to spend their limited precision budget by looking at a small sample of calibration data, and that data is chosen to preserve capability; it carries no signal about bias prevention at all. The compressed model is optimized to keep being good at things, with nothing in the procedure asking it to keep being fair. There is no reason to expect better behavior to emerge from better compression alone, because fairness is not in the objective.

There is one partial mitigation in the data, and it is instructive precisely because it is partial. Using the in-chat toggle to prompt the model to reason before answering roughly halves the effect on some model families, while doing nothing at all on others. That is a useful reminder that mitigations validated on one model do not transfer by default. A rule that mandated a specific mitigation technique would be betting on a behavior that demonstrably varies across architectures.

None of this is an argument against quantization. Compression is what puts capable models in the hands of people who cannot rent a datacenter. The argument is narrower: compression is a change to the deployed system, and it should be treated as one.

Quantized builds should be reevaluated for open-ended bias, not only on the short-form safeguards they already pass. Evaluations should record the precision and configuration of the artifact tested. And evaluation suites used for anything resembling certification should include sustained open-ended generation, in more than one language.

The uncomfortable version of this finding is that a model can pass every test we put in front of it and still reach users measurably more biased. That is not a failure of the model. It is a failure of where we are pointing the instruments.

Jake Laperruque is a fellow at Tech Policy Press. This is the first post in “Perils of AI in Policing,” a three-part series on emerging risks of artificial intelligence’s use in law enforcement.

Society’s growing reliance on artificial intelligence has given rise to concerns about 

. This dark pattern—the tendency of AI systems to tailor responses to flatter users and validate their point of view—has alarming social ramifications: commentators worry it could 

 human values and expectations or prop up reliance on 

. Beyond these general risks, we need to account for problems AI sycophancy may cause when the tech is applied to specific, high-risk functions. In particular, we ought to examine how it could impact the use of AI in policing.

AI is becoming incorporated into an array of law enforcement tasks as a virtual assistant. Companies like Axon are pushing AI services that scan body camera footage and write 

 promote AI systems serving as a virtual detective, taking investigative case files and providing summaries, timelines, surfacing leads, and providing analytics. The company 

 to take a “scan the case - what am I missing?” approach with its systems, promising an AI system that “summarizes, organizes, and provides critical insights or gaps.” These tools are being pushed to prosecutors as well as investigators: Thomson Reuters offers a 

 AI tool for prosecutors to summarize evidence files, review discovery materials to “identify key facts” and “create comprehensive timelines” for cases, as well as conduct legal review on evidentiary questions and even draft charging documents and plea agreements.

Could these tools be susceptible to AI sycophancy? This trend can result in outright hallucinations, slanted and selective information, or responses that are merely unctuous and fawning. Sycophancy can occur for 

, but especially because people tend to gravitate toward agreeable answers, even if unconsciously. This affects model training and 

: users tend to give higher ratings to sycophantic AI systems and use them more frequently, so systems adapt to become even more agreeable. As one 

 aptly puts it, “The very feature that causes harm also drives engagement.” This is especially worrisome because it not only warps AI but also creates a feedback loop, where increased use augments AI sycophancy and AI sycophancy encourages more reliance on the software.

It's frighteningly easy to imagine how this could occur during law enforcement use of AI. Officers could be more prone to use an AI police report that depicts their conduct as calm, smart, and professional. Investigators will be more likely to rely on case summarization tools that affirm their suspicions and frame their work as sound. Prosecutors could gravitate toward AI systems that provide straightforward solutions and present favorable assessments of key factual and legal questions for cases.

The impact in each of these contexts could be immense. What if automated police report tools develop a tendency to tailor descriptions of encounters to make police 

 and use of force justified by default? How will case summarization tools shape evidence if they are driven to providing investigators with a culprit? If prosecutorial AI assistants are shaped to please their users, will they be overly restrictive in what exculpatory evidence they label as 

that must be disclosed to defendants? Even outside of more severe cases in which evidence could be fabricated, altered, or omitted, AI sycophancy merely slanting summaries to be more favorable to users—in this case, police and prosecutors—could have serious ramifications. And these distorted results will be presented as coming from an unbiased, objective source.

The most basic requirement to address the danger of AI sycophancy in policing is strong transparency. We need independent evaluation of AI systems used by police to test for sycophancy (as well as for general review of potential bias and accuracy issues). This is a serious problem, because we’re likely to confront secrecy from vendors and police alike. Axon’s “Draft One” automated police report drafting tool is already 

 in significant transparency problems: Its systems aren’t independently reviewed or tested, and it's designed to avoid tracking which parts of automated police reports are edited by officers. This means there’s no way to audit systems to review how often and in what way this AI system makes errors. And Draft One’s lack of transparency is hardly an isolated incident: there's a long history of the vendors of AI tools for police—from 

—pushing back against independent testing that is critical of their products, a problem likely to recur since sellers have a financial interest in marketing their products as cutting-edge and effective. Lack of transparency from police is also a proven problem: From 

, we’ve seen police hide their use of AI tools as a deliberate tactic.

Cities, states, and lawmakers in Congress need to take a hard, critical look at how AI sycophancy and other unintended consequences of the technology could imperil policing and harm civil liberties and public safety alike. The best course of action may simply be to hit pause on their acquisition and use, especially for tools that haven’t been tried out in less high-stakes arenas. At a minimum, demanding independent review and rigorous, ongoing vetting for any police AI tools is essential. It’s the only way to ensure police do not end up relying on a digital yes man that puts the public at risk.

Opposition to data centers is growing across the United States, and, like the 

 before it, the tech industry is pulling from tried-and-tested playbooks to bypass and 

 public opposition. The results of these efforts reveal both the patchy infrastructure for transparency and accountability in the US, as well as the speed and determination with which industry is moving to build these facilities with 

 by other industries to manipulate public opinion have ranged from 

 designed to deflect oil industry accountability for the harmful effects of plastics. 

 and effectively insulate corporate bottom lines from regulatory scrutiny and public accountability.

Up against a similar PR playbook, some residents in affected communities reported feeling like insurgents against both Big Tech and local elected officials. At least some representatives of the tech industry agree: in September 2025, Kevin Nichols, an employee of the advocacy group Earth Charter Indiana, attended the Midwest Data Center Investment Conference and Expo (DICE) in Chicago, where he says he 

 one speaker comparing his experience promoting data centers to counter-insurgency tactics the speaker used during military service in Iraq and Afghanistan, such as “going undercover in bars and churches to gauge a community’s potential for resistance to a data center development.” 

The language of intelligence gathering is apparent in a 

 the number of deals “blocked or delayed” by local opposition, which tells applicants they will “conduct open-source and targeted research in various sources, including social media” as part of an effort to track “political risks emerging from local opposition to data centers.” (A representative from Data Center Watch says that it “does not maintain separate HR or recruiting functions” from 10a Labs, a company that 

 itself as collecting “intelligence” for “AI unicorns, Fortune 10 companies, and US tech leaders.”)

Against this backdrop, Big Tech is spending big money to influence the public conversation. As the 2026 US midterm elections approach, super PACs focused on AI have already 

 more than $50 million on Congressional races. According to Transformer News, by far the largest tranche of this money flows through a group called Leading the Future; its biggest donors are the venture capitalist Marc Andreessen, his partner Ben Horowitz, and OpenAI President Greg Brockman and his wife, Anna.

Leading the Future—alongside Build American AI, an 

 PAC—is doing more than making campaign donations. In May, Taylor Lorenz reported for Wired that Build American AI had 

 social media influencers into a pro-data-center campaign that leaned heavily on arguments about an AI race with China. A month later, Lorenz 

 a joint investigation with Tyler Johnston of Model Republic alleging that Build American AI runs multiple fake “sockpuppet” accounts on X. Also in May, Johnston 

 that another Leading the Future affiliated PAC, Think Big, seems to receive a large amount of its social media engagement from what appear to be fake accounts posing as OnlyFans performers.

 that Leading the Future might be connected to what appears to be an AI-generated news site called Acutus; the site repeats talking points often used by Leading the Future and frequently publishes harsh criticism of AI safety advocates. In March, Transformer 

 that Build American AI had spent more than half a million dollars on ads to harvest the contact information of between 500,000 and 1,000,000 supporters on social media—a common astroturf tactic that allows groups to boast about the size of their support base even when that support may be quite shallow.

One noticeable trend has been the data center industry’s adoption of a longstanding marketing tactic: the creation and use of nonprofit front groups. Many of these groups are registered nonprofits who keep their funding, staff, and industry relationships deliberately opaque; while they can sometimes be identified through tax filings like IRS Form 990, their websites often fail to disclose even basic information about their leadership or sources of support.

“Companies' use of third parties like corporate front groups and astroturf campaigns to secretly advance their interests has a long tradition in the US,” said Jocelyn Leitzinger, a researcher who studies activist campaigns and how corporations respond to them. “Over the past century, companies in highly regulated industries such as tobacco, fossil fuels, and pharmaceuticals have used corporate front groups to block or rollback regulations protecting consumers and the natural environment.”

 roughly $700,000 in fiscal year 2024 on mailers, text messages, and other outreach 

 data center development to Virginia residents. According to its Form 990, Virginia Connects is 

 in Richmond and connected to the Data Center Coalition, an industry lobbying group; Josh Levi, the Coalition’s head, is also 

 as Virginia Connects’ principal officer.

 that Texas Connects, a similarly named nonprofit “dedicated to educating Texans about the data center industry,” was also established by the Data Center Coalition. Similar groups, for instance 

 have recently emerged with similar website templates—all linked by the same Google Ads ID, a unique identifier assigned to advertisers.

In a statement to Tech Policy Press, Data Center Coalition president Josh Levi said that America Connects—the umbrella over these similarly named groups—represents a coalition of stakeholders from the data center ecosystem working to educate and engage with communities about the benefits of data centers. The initiative initially launched in Virginia in 2024 and is expanding its education efforts into Texas, Ohio, Pennsylvania, Indiana, Oregon and other additional states.”

Similar efforts are run by other industry stakeholders. In Virginia’s Fauquier County, for example, a group named 

 promotes data center development as a way of raising tax revenue. In October, the Fauquier Times 

 that the contact listed on Fauquier Forward’s Facebook page is the 

 of Gigaland, a data center development company. Fauquier Forward is active on social media: its 

, for instance, includes dozens of videos, though it has received little engagement.

 in February that an organization called Missourians for Data and Technology Advancement, or 

, published an ad quoting the Jefferson County Executive—but the executive claimed no knowledge of the group, and the ad was purchased by an out-of-state PR firm called Axiom Strategies. A 2025 press release 

 MODATA a “new statewide coalition,” but the group’s website does not list its members.

In a statement to Tech Policy Press, a MODATA spokesman said that “MODATA represents a broad coalition of large and small businesses, organizations, utilities, labor organizations, and developers in Missouri that support responsible data center policies in the state. We track the positive 

 from elected officials and experts on our website as well and amplify those through paid and earned media.”

It is unclear how impactful these organizations have been to date. Chloe Humbert, an activist based in Scranton who has seen ads from Pennsylvania Connects, is skeptical. In a statement to Tech Policy Press, she said, “I have no idea who PA Connects’ ads would even influence. Maybe the politicians?”

Four people stand outside a brick municipal building holding protest signs opposing a data center. One holds a handmade sign reading "Sign Recall Petition," another holds a sign reading "Exercising First Amendment Rights! No Data Center," and two hold printed signs with a red circle-and-slash symbol reading "No Data Center!" and the URL WakeUpJeffco.org. American flags fly on poles in front of the building, and cars are parked nearby under a blue sky.

From left to right, Luck Adams, Victor Dubbs, Rain Dixon, and Steve Carter protest data center development in Festus, Missouri on May 11, 2026. Photo: Sam Totoni

There are signs that Humbert is right. Data center proponents are shifting toward strategies for keeping the public at arm’s length while courting local leaders.

Project Jupiter is a planned hyperscale data center 

 by OpenAI and Oracle in Doña Ana County, New Mexico. Daisy Maldonado, a candidate for the Doña Ana County Commission, said her first clues that a data center was under development came from a man who initially identified himself as a community organizer. Later, she says she learned he worked for a company called 

—one of the companies involved in the development of Project Jupiter. Maldonado said  the man “[recruit[ed] people to speak” positively about the project. Others 

 at County Commission meetings without identifying himself as an employee of BorderPlex Digital Assets. (BorderPlex Digital Assets did not respond to a request for comment.)

Maldonado said that over time, Project Jupiter’s strategy has shifted: engagement with the public moved from public hearings to community forums where corporate officials have more control over the microphone. In the days prior to a county commission 

 to issue bonds for the project, corporate and county representatives held 

“People were incredibly unhappy,” said Maldonado when asked to describe the meetings. Security and law enforcement were present for a contentious conversation with a standing-room-only audience; at the next session, Maldonado said, the audience was asked to write their questions on note cards. After pushback, organizers switched to calling on audience members; Maldonado said members of the community suspected that some participants could be industry plants.

On September 19, 2025, Doña Ana County voted to issue the bonds after an 

 featuring a packed audience and impassioned public comment.

The forum-switching Maldonado described is not limited to New Mexico. In Texas, former State Senate candidate 

 described similar dynamics with “meet-and-greets” which were frustratingly unproductive: “You go in thinking maybe you could say something, or you know, get some information, but you can't,” she said.

 that a group called Elevate New Mexico ran a campaign promoting Project Jupiter through mailers and digital ads. The mailer showed a smiling young woman with dark hair and brown skin, next to the words “A new day for New Mexico.”

A promotional mailer titled "A New Day for New Mexico" promoting Project Jupiter, a proposed technology campus in Doña Ana County. The top section on a purple background reads: "We have the opportunity to reshape our future and transform with Project Jupiter, a new technology campus in Doña Ana County." Three highlighted figures follow — "Our Families: $360 million in direct payments to support our local schools and public services," "Our Water: $50 million investment in improving our community water systems," and "Our Community: Thousands of high-paying careers, prioritizing Doña Ana residents." A "Did You Know?" section on a blue background reads: "An Investment in Our Future — Project Jupiter will build and operate a state-of-the-art microgrid, supplying its own electricity so local homes and businesses keep the power they rely on. This means the project won't make your bills go up or lights go out." The URL ElevateNewMexico.org appears at the bottom. A smiling woman in a yellow top is pictured on the right side.

Flier supporting Project Jupiter mailed to New Mexico residents. Image from Source New Mexico.

Source New Mexico found that the woman is a stock model and that Elevate New Mexico’s mailing address is in Alexandria, Virginia. Critics told reporters it is an example of ‘

,’ a practice through which companies attempt to appropriate the identities of people of color. Companies involved in the project formed LLCs with names like “Green Chile Ventures” and “Acoma”—the latter being the name of a Native American tribe. In response to 

 from that tribe’s governor, Acoma LLC changed its name to Yucca Growth Infrastructure; YUCCA is also the acronym for a New Mexican environmental organization, 

In April, the New Mexico State Ethics Commission 

 a lawsuit against Elevate New Mexico arguing the organization should be forced to register as a lobbyist with the state. Both Source New Mexico’s reporting and the lawsuit note that at least some of Elevate New Mexico’s digital ads were paid for by APCO Worldwide, a public affairs firm that promotes 

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe