Entries Tagged "AI"

Page 1 of 23

AI Advertising Company Hacked

At least some of this is coming to light:

Doublespeed, a startup backed by Andreessen Horowitz (a16z) that uses a phone farm to manage at least hundreds of AI-generated social media accounts and promote products has been hacked. The hack reveals what products the AI-generated accounts are promoting, often without the required disclosure that these are advertisements, and allowed the hacker to take control of more than 1,000 smartphones that power the company.

The hacker, who asked for anonymity because he feared retaliation from the company, said he reported the vulnerability to Doublespeed on October 31. At the time of writing, the hacker said he still has access to the company’s backend, including the phone farm itself.

Slashdot thread.

Posted on December 19, 2025 at 7:02 AM • View Comments

Chinese Surveillance and AI

New report: “The Party’s AI: How China’s New AI Systems are Reshaping Human Rights.” From a summary article:

China is already the world’s largest exporter of AI powered surveillance technology; new surveillance technologies and platforms developed in China are also not likely to simply stay there. By exposing the full scope of China’s AI driven control apparatus, this report presents clear, evidence based insights for policymakers, civil society, the media and technology companies seeking to counter the rise of AI enabled repression and human rights violations, and China’s growing efforts to project that repression beyond its borders.

The report focuses on four areas where the CCP has expanded its use of advanced AI systems most rapidly between 2023 and 2025: multimodal censorship of politically sensitive images; AI’s integration into the criminal justice pipeline; the industrialisation of online information control; and the use of AI enabled platforms by Chinese companies operating abroad. Examined together, those cases show how new AI capabilities are being embedded across domains that strengthen the CCP’s ability to shape information, behaviour and economic outcomes at home and overseas.

Because China’s AI ecosystem is evolving rapidly and unevenly across sectors, we have focused on domains where significant changes took place between 2023 and 2025, where new evidence became available, or where human rights risks accelerated. Those areas do not represent the full range of AI applications in China but are the most revealing of how the CCP is integrating AI technologies into its political control apparatus.

News article.

Posted on December 16, 2025 at 7:02 AM • View Comments

Against the Federal Moratorium on State-Level Regulation of AI

Cast your mind back to May of this year: Congress was in the throes of debate over the massive budget bill. Amidst the many seismic provisions, Senator Ted Cruz dropped a ticking time bomb of tech policy: a ten-year moratorium on the ability of states to regulate artificial intelligence. To many, this was catastrophic. The few massive AI companies seem to be swallowing our economy whole: their energy demands are overriding household needs, their data demands are overriding creators’ copyright, and their products are triggering mass unemployment as well as new types of clinical psychoses. In a moment where Congress is seemingly unable to act to pass any meaningful consumer protections or market regulations, why would we hamstring the one entity evidently capable of doing so—the states? States that have already enacted consumer protections and other AI regulations, like California, and those actively debating them, like Massachusetts, were alarmed. Seventeen Republican governors wrote a letter decrying the idea, and it was ultimately killed in a rare vote of bipartisan near-unanimity.

The idea is back. Before Thanksgiving, a House Republican leader suggested they might slip it into the annual defense spending bill. Then, a draft document leaked outlining the Trump administration’s intent to enforce the state regulatory ban through executive powers. An outpouring of opposition (including from some Republican state leaders) beat back that notion for a few weeks, but on Monday, Trump posted on social media that the promised Executive Order is indeed coming soon. That would put a growing cohort of states, including California and New York, as well as Republican strongholds like Utah and Texas, in jeopardy.

The constellation of motivations behind this proposal is clear: conservative ideology, cash, and China.

The intellectual argument in favor of the moratorium is that “freedom“-killing state regulation on AI would create a patchwork that would be difficult for AI companies to comply with, which would slow the pace of innovation needed to win an AI arms race with China. AI companies and their investors have been aggressively peddling this narrative for years now, and are increasingly backing it with exorbitant lobbying dollars. It’s a handy argument, useful not only to kill regulatory constraints, but also—companies hope—to win federal bailouts and energy subsidies.

Citizens should parse that argument from their own point of view, not Big Tech’s. Preventing states from regulating AI means that those companies get to tell Washington what they want, but your state representatives are powerless to represent your own interests. Which freedom is more important to you: the freedom for a few near-monopolies to profit from AI, or the freedom for you and your neighbors to demand protections from its abuses?

There is an element of this that is more partisan than ideological. Vice President J.D. Vance argued that federal preemption is needed to prevent “progressive” states from controlling AI’s future. This is an indicator of creeping polarization, where Democrats decry the monopolism, bias, and harms attendant to corporate AI and Republicans reflexively take the opposite side. It doesn’t help that some in the parties also have direct financial interests in the AI supply chain.

But this does not need to be a partisan wedge issue: both Democrats and Republicans have strong reasons to support state-level AI legislation. Everyone shares an interest in protecting consumers from harm created by Big Tech companies. In leading the charge to kill Cruz’s initial AI moratorium proposal, Republican Senator Masha Blackburn explained that “This provision could allow Big Tech to continue to exploit kids, creators, and conservatives? we can’t block states from making laws that protect their citizens.” More recently, Florida Governor Ron DeSantis wants to regulate AI in his state.

The often-heard complaint that it is hard to comply with a patchwork of state regulations rings hollow. Pretty much every other consumer-facing industry has managed to deal with local regulation—automobiles, children’s toys, food, and drugs—and those regulations have been effective consumer protections. The AI industry includes some of the most valuable companies globally and has demonstrated the ability to comply with differing regulations around the world, including the EU’s AI and data privacy regulations, substantially more onerous than those so far adopted by US states. If we can’t leverage state regulatory power to shape the AI industry, to what industry could it possibly apply?

The regulatory superpower that states have here is not size and force, but rather speed and locality. We need the “laboratories of democracy” to experiment with different types of regulation that fit the specific needs and interests of their constituents and evolve responsively to the concerns they raise, especially in such a consequential and rapidly changing area such as AI.

We should embrace the ability of regulation to be a driver—not a limiter—of innovation. Regulations don’t restrict companies from building better products or making more profit; they help channel that innovation in specific ways that protect the public interest. Drug safety regulations don’t prevent pharma companies from inventing drugs; they force them to invent drugs that are safe and efficacious. States can direct private innovation to serve the public.

But, most importantly, regulations are needed to prevent the most dangerous impact of AI today: the concentration of power associated with trillion-dollar AI companies and the power-amplifying technologies they are producing. We outline the specific ways that the use of AI in governance can disrupt existing balances of power, and how to steer those applications towards more equitable balances, in our new book, Rewiring Democracy. In the nearly complete absence of Congressional action on AI over the years, it has swept the world’s attention; it has become clear that states are the only effective policy levers we have against that concentration of power.

Instead of impeding states from regulating AI, the federal government should support them to drive AI innovation. If proponents of a moratorium worry that the private sector won’t deliver what they think is needed to compete in the new global economy, then we should engage government to help generate AI innovations that serve the public and solve the problems most important to people. Following the lead of countries like Switzerland, France, and Singapore, the US could invest in developing and deploying AI models designed as public goods: transparent, open, and useful for tasks in public administration and governance.

Maybe you don’t trust the federal government to build or operate an AI tool that acts in the public interest? We don’t either. States are a much better place for this innovation to happen because they are closer to the people, they are charged with delivering most government services, they are better aligned with local political sentiments, and they have achieved greater trust. They’re where we can test, iterate, compare, and contrast regulatory approaches that could inform eventual and better federal policy. And, while the costs of training and operating performance AI tools like large language models have declined precipitously, the federal government can play a valuable role here in funding cash-strapped states to lead this kind of innovation.

This essay was written with Nathan E. Sanders, and originally appeared in Gizmodo.

EDITED TO ADD: Trump signed an executive order banning state-level AI regulations hours after this was published. This is not going to be the last word on the subject.

Posted on December 15, 2025 at 7:02 AM • View Comments

Building Trustworthy AI Agents

The promise of personal AI assistants rests on a dangerous assumption: that we can trust systems we haven’t made trustworthy. We can’t. And today’s versions are failing us in predictable ways: pushing us to do things against our own best interests, gaslighting us with doubt about things we are or that we know, and being unable to distinguish between who we are and who we have been. They struggle with incomplete, inaccurate, and partial context: with no standard way to move toward accuracy, no mechanism to correct sources of error, and no accountability when wrong information leads to bad decisions.

These aren’t edge cases. They’re the result of building AI systems without basic integrity controls. We’re in the third leg of data security—the old CIA triad. We’re good at availability and working on confidentiality, but we’ve never properly solved integrity. Now AI personalization has exposed the gap by accelerating the harms.

The scope of the problem is large. A good AI assistant will need to be trained on everything we do and will need access to our most intimate personal interactions. This means an intimacy greater than your relationship with your email provider, your social media account, your cloud storage, or your phone. It requires an AI system that is both discreet and trustworthy when provided with that data. The system needs to be accurate and complete, but it also needs to be able to keep data private: to selectively disclose pieces of it when required, and to keep it secret otherwise. No current AI system is even close to meeting this.

To further development along these lines, I and others have proposed separating users’ personal data stores from the AI systems that will use them. It makes sense; the engineering expertise that designs and develops AI systems is completely orthogonal to the security expertise that ensures the confidentiality and integrity of data. And by separating them, advances in security can proceed independently from advances in AI.

What would this sort of personal data store look like? Confidentiality without integrity gives you access to wrong data. Availability without integrity gives you reliable access to corrupted data. Integrity enables the other two to be meaningful. Here are six requirements. They emerge from treating integrity as the organizing principle of security to make AI trustworthy.

First, it would be broadly accessible as a data repository. We each want this data to include personal data about ourselves, as well as transaction data from our interactions. It would include data we create when interacting with others—emails, texts, social media posts—and revealed preference data as inferred by other systems. Some of it would be raw data, and some of it would be processed data: revealed preferences, conclusions inferred by other systems, maybe even raw weights in a personal LLM.

Second, it would be broadly accessible as a source of data. This data would need to be made accessible to different LLM systems. This can’t be tied to a single AI model. Our AI future will include many different models—some of them chosen by us for particular tasks, and some thrust upon us by others. We would want the ability for any of those models to use our data.

Third, it would need to be able to prove the accuracy of data. Imagine one of these systems being used to negotiate a bank loan, or participate in a first-round job interview with an AI recruiter. In these instances, the other party will want both relevant data and some sort of proof that the data are complete and accurate.

Fourth, it would be under the user’s fine-grained control and audit. This is a deeply detailed personal dossier, and the user would need to have the final say in who could access it, what portions they could access, and under what circumstances. Users would need to be able to grant and revoke this access quickly and easily, and be able to go back in time and see who has accessed it.

Fifth, it would be secure. The attacks against this system are numerous. There are the obvious read attacks, where an adversary attempts to learn a person’s data. And there are also write attacks, where adversaries add to or change a user’s data. Defending against both is critical; this all implies a complex and robust authentication system.

Sixth, and finally, it must be easy to use. If we’re envisioning digital personal assistants for everybody, it can’t require specialized security training to use properly.

I’m not the first to suggest something like this. Researchers have proposed a “Human Context Protocol” (https://papers.ssrn.com/sol3/ papers.cfm?abstract_id=5403981) that would serve as a neutral interface for personal data of this type. And in my capacity at a company called Inrupt, Inc., I have been working on an extension of Tim Berners-Lee’s Solid protocol for distributed data ownership.

The engineering expertise to build AI systems is orthogonal to the security expertise needed to protect personal data. AI companies optimize for model performance, but data security requires cryptographic verification, access control, and auditable systems. Separating the two makes sense; you can’t ignore one or the other.

Fortunately, decoupling personal data stores from AI systems means security can advance independently from performance (https:// ieeexplore.ieee.org/document/ 10352412). When you own and control your data store with high integrity, AI can’t easily manipulate you because you see what data it’s using and can correct it. It can’t easily gaslight you because you control the authoritative record of your context. And you determine which historical data are relevant or obsolete. Making this all work is a challenge, but it’s the only way we can have trustworthy AI assistants.

This essay was originally published in IEEE Security & Privacy.

Posted on December 12, 2025 at 7:00 AM • View Comments

AIs Exploiting Smart Contracts

I have long maintained that smart contracts are a dumb idea: that a human process is actually a security feature.

Here’s some interesting research on training AIs to automatically exploit smart contracts:

AI models are increasingly good at cyber tasks, as we’ve written about before. But what is the economic impact of these capabilities? In a recent MATS and Anthropic Fellows project, our scholars investigated this question by evaluating AI agents’ ability to exploit smart contracts on Smart CONtracts Exploitation benchmark (SCONE-bench)­a new benchmark they built comprising 405 contracts that were actually exploited between 2020 and 2025. On contracts exploited after the latest knowledge cutoffs (June 2025 for Opus 4.5 and March 2025 for other models), Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 developed exploits collectively worth $4.6 million, establishing a concrete lower bound for the economic harm these capabilities could enable. Going beyond retrospective analysis, we evaluated both Sonnet 4.5 and GPT-5 in simulation against 2,849 recently deployed contracts without any known vulnerabilities. Both agents uncovered two novel zero-day vulnerabilities and produced exploits worth $3,694, with GPT-5 doing so at an API cost of $3,476. This demonstrates as a proof-of-concept that profitable, real-world autonomous exploitation is technically feasible, a finding that underscores the need for proactive adoption of AI for defense.

Posted on December 11, 2025 at 12:06 PM • View Comments

FBI Warns of Fake Video Scams

The FBI is warning of AI-assisted fake kidnapping scams:

Criminal actors typically will contact their victims through text message claiming they have kidnapped their loved one and demand a ransom be paid for their release. Oftentimes, the criminal actor will express significant claims of violence towards the loved one if the ransom is not paid immediately. The criminal actor will then send what appears to be a genuine photo or video of the victim’s loved one, which upon close inspection often reveals inaccuracies when compared to confirmed photos of the loved one. Examples of these inaccuracies include missing tattoos or scars and inaccurate body proportions. Criminal actors will sometimes purposefully send these photos using timed message features to limit the amount of time victims have to analyze the images.

Images, videos, audio: It can all be faked with AI. My guess is that this scam has a low probability of success, so criminals will be figuring out how to automate it.

Posted on December 10, 2025 at 7:05 AM • View Comments

AI vs. Human Drivers

Two competing arguments are making the rounds. The first is by a neurosurgeon in the New York Times. In an op-ed that honestly sounds like it was paid for by Waymo, the author calls driverless cars a “public health breakthrough”:

In medical research, there’s a practice of ending a study early when the results are too striking to ignore. We stop when there is unexpected harm. We also stop for overwhelming benefit, when a treatment is working so well that it would be unethical to continue giving anyone a placebo. When an intervention works this clearly, you change what you do.

There’s a public health imperative to quickly expand the adoption of autonomous vehicles. More than 39,000 Americans died in motor vehicle crashes last year, more than homicide, plane crashes and natural disasters combined. Crashes are the No. 2 cause of death for children and young adults. But death is only part of the story. These crashes are also the leading cause of spinal cord injury. We surgeons see the aftermath of the 10,000 crash victims who come to emergency rooms every day.

The other is a soon-to-be-published book: Driving Intelligence: The Green Book. The authors, a computer scientist and a management consultant with experience in the industry, make the opposite argument. Here’s one of the authors:

There is something very disturbing going on around trials with autonomous vehicles worldwide, where, sadly, there have now been many deaths and injuries both to other road users and pedestrians. Although I am well aware that there is not, senso stricto, a legal and functional parallel between a “drug trial” and “AV testing,” it seems odd to me that if a trial of a new drug had resulted in so many deaths, it would surely have been halted and major forensic investigations carried out and yet, AV manufacturers continue to test their products on public roads unabated.

I am not convinced that it is good enough to argue from statistics that, to a greater or lesser degree, fatalities and injuries would have occurred anyway had the AVs had been replaced by human-driven cars: a pharmaceutical company, following death or injury, cannot simply sidestep regulations around the trial of, say, a new cancer drug, by arguing that, whilst the trial is underway, people would die from cancer anyway….

Both arguments are compelling, and it’s going to be hard to figure out what public policy should be.

This paper, from 2016, argues that we’re going to need other metrics than side-by-side comparisons: Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability?“:

Abstract: How safe are autonomous vehicles? The answer is critical for determining how autonomous vehicles may shape motor vehicle safety and public health, and for developing sound policies to govern their deployment. One proposed way to assess safety is to test drive autonomous vehicles in real traffic, observe their performance, and make statistical comparisons to human driver performance. This approach is logical, but it is practical? In this paper, we calculate the number of miles of driving that would be needed to provide clear statistical evidence of autonomous vehicle safety. Given that current traffic fatalities and injuries are rare events compared to vehicle miles traveled, we show that fully autonomous vehicles would have to be driven hundreds of millions of miles and sometimes hundreds of billions of miles to demonstrate their reliability in terms of fatalities and injuries. Under even aggressive testing assumptions, existing fleets would take tens and sometimes hundreds of years to drive these miles—­an impossible proposition if the aim is to demonstrate their performance prior to releasing them on the roads for consumer use. These findings demonstrate that developers of this technology and third-party testers cannot simply drive their way to safety. Instead, they will need to develop innovative methods of demonstrating safety and reliability. And yet, the possibility remains that it will not be possible to establish with certainty the safety of autonomous vehicles. Uncertainty will remain. Therefore, it is imperative that autonomous vehicle regulations are adaptive­—designed from the outset to evolve with the technology so that society can better harness the benefits and manage the risks of these rapidly evolving and potentially transformative technologies.

One problem, of course, is that we treat death by human driver differently than we do death by autonomous computer driver. This is likely to change as we get more experience with AI accidents—and AI-caused deaths.

Posted on December 9, 2025 at 7:07 AM • View Comments

Like Social Media, AI Requires Difficult Choices

In his 2020 book, “Future Politics,” British barrister Jamie Susskind wrote that the dominant question of the 20th century was “How much of our collective life should be determined by the state, and what should be left to the market and civil society?” But in the early decades of this century, Susskind suggested that we face a different question: “To what extent should our lives be directed and controlled by powerful digital systems—and on what terms?”

Artificial intelligence (AI) forces us to confront this question. It is a technology that in theory amplifies the power of its users: A manager, marketer, political campaigner, or opinionated internet user can utter a single instruction, and see their message—whatever it is—instantly written, personalized, and propagated via email, text, social, or other channels to thousands of people within their organization, or millions around the world. It also allows us to individualize solicitations for political donations, elaborate a grievance into a well-articulated policy position, or tailor a persuasive argument to an identity group, or even a single person.

But even as it offers endless potential, AI is a technology that—like the state—gives others new powers to control our lives and experiences.

We’ve seen this play out before. Social media companies made the same sorts of promises 20 years ago: instant communication enabling individual connection at massive scale. Fast-forward to today, and the technology that was supposed to give individuals power and influence ended up controlling us. Today social media dominates our time and attention, assaults our mental health, and—together with its Big Tech parent companies—captures an unfathomable fraction of our economy, even as it poses risks to our democracy.

The novelty and potential of social media was as present then as it is for AI now, which should make us wary of its potential harmful consequences for society and democracy. We legitimately fear artificial voices and manufactured reality drowning out real people on the internet: on social media, in chat rooms, everywhere we might try to connect with others.

It doesn’t have to be that way. Alongside these evident risks, AI has legitimate potential to transform both everyday life and democratic governance in positive ways. In our new book, “Rewiring Democracy,” we chronicle examples from around the globe of democracies using AI to make regulatory enforcement more efficient, catch tax cheats, speed up judicial processes, synthesize input from constituents to legislatures, and much more. Because democracies distribute power across institutions and individuals, making the right choices about how to shape AI and its uses requires both clarity and alignment across society.

To that end, we spotlight four pivotal choices facing private and public actors. These choices are similar to those we faced during the advent of social media, and in retrospect we can see that we made the wrong decisions back then. Our collective choices in 2025—choices made by tech CEOs, politicians, and citizens alike—may dictate whether AI is applied to positive and pro-democratic, or harmful and civically destructive, ends.

A Choice for the Executive and the Judiciary: Playing by the Rules

The Federal Election Commission (FEC) calls it fraud when a candidate hires an actor to impersonate their opponent. More recently, they had to decide whether doing the same thing with an AI deepfake makes it okay. (They concluded it does not.) Although in this case the FEC made the right decision, this is just one example of how AIs could skirt laws that govern people.

Likewise, courts are having to decide if and when it is okay for an AI to reuse creative materials without compensation or attribution, which might constitute plagiarism or copyright infringement if carried out by a human. (The court outcomes so far are mixed.) Courts are also adjudicating whether corporations are responsible for upholding promises made by AI customer service representatives. (In the case of Air Canada, the answer was yes, and insurers have started covering the liability.)

Social media companies faced many of the same hazards decades ago and have largely been shielded by the combination of Section 230 of the Communications Act of 1994 and the safe harbor offered by the Digital Millennium Copyright Act of 1998. Even in the absence of congressional action to strengthen or add rigor to this law, the Federal Communications Commission (FCC) and the Supreme Court could take action to enhance its effects and to clarify which humans are responsible when technology is used, in effect, to bypass existing law.

A Choice for Congress: Privacy

As AI-enabled products increasingly ask Americans to share yet more of their personal information—their “context“—to use digital services like personal assistants, safeguarding the interests of the American consumer should be a bipartisan cause in Congress.

It has been nearly 10 years since Europe adopted comprehensive data privacy regulation. Today, American companies exert massive efforts to limit data collection, acquire consent for use of data, and hold it confidential under significant financial penalties—but only for their customers and users in the EU.

Regardless, a decade later the U.S. has still failed to make progress on any serious attempts at comprehensive federal privacy legislation written for the 21st century, and there are precious few data privacy protections that apply to narrow slices of the economy and population. This inaction comes in spite of scandal after scandal regarding Big Tech corporations’ irresponsible and harmful use of our personal data: Oracle’s data profiling, Facebook and Cambridge Analytica, Google ignoring data privacy opt-out requests, and many more.

Privacy is just one side of the obligations AI companies should have with respect to our data; the other side is portability—that is, the ability for individuals to choose to migrate and share their data between consumer tools and technology systems. To the extent that knowing our personal context really does enable better and more personalized AI services, it’s critical that consumers have the ability to extract and migrate their personal context between AI solutions. Consumers should own their own data, and with that ownership should come explicit control over who and what platforms it is shared with, as well as withheld from. Regulators could mandate this interoperability. Otherwise, users are locked in and lack freedom of choice between competing AI solutions—much like the time invested to build a following on a social network has locked many users to those platforms.

A Choice for States: Taxing AI Companies

It has become increasingly clear that social media is not a town square in the utopian sense of an open and protected public forum where political ideas are distributed and debated in good faith. If anything, social media has coarsened and degraded our public discourse. Meanwhile, the sole act of Congress designed to substantially reign in the social and political effects of social media platforms—the TikTok ban, which aimed to protect the American public from Chinese influence and data collection, citing it as a national security threat—is one it seems to no longer even acknowledge.

While Congress has waffled, regulation in the U.S. is happening at the state level. Several states have limited children’s and teens’ access to social media. With Congress having rejected—for now—a threatened federal moratorium on state-level regulation of AI, California passed a new slate of AI regulations after mollifying a lobbying onslaught from industry opponents. Perhaps most interesting, Maryland has recently become the first in the nation to levy taxes on digital advertising platform companies.

States now face a choice of whether to apply a similar reparative tax to AI companies to recapture a fraction of the costs they externalize on the public to fund affected public services. State legislators concerned with the potential loss of jobs, cheating in schools, and harm to those with mental health concerns caused by AI have options to combat it. They could extract the funding needed to mitigate these harms to support public services—strengthening job training programs and public employment, public schools, public health services, even public media and technology.

A Choice for All of Us: What Products Do We Use, and How?

A pivotal moment in the social media timeline occurred in 2006, when Facebook opened its service to the public after years of catering to students of select universities. Millions quickly signed up for a free service where the only source of monetization was the extraction of their attention and personal data.

Today, about half of Americans are daily users of AI, mostly via free products from Facebook’s parent company Meta and a handful of other familiar Big Tech giants and venture-backed tech firms such as Google, Microsoft, OpenAI, and Anthropic—with every incentive to follow the same path as the social platforms.

But now, as then, there are alternatives. Some nonprofit initiatives are building open-source AI tools that have transparent foundations and can be run locally and under users’ control, like AllenAI and EleutherAI. Some governments, like Singapore, Indonesia, and Switzerland, are building public alternatives to corporate AI that don’t suffer from the perverse incentives introduced by the profit motive of private entities.

Just as social media users have faced platform choices with a range of value propositions and ideological valences—as diverse as X, Bluesky, and Mastodon—the same will increasingly be true of AI. Those of us who use AI products in our everyday lives as people, workers, and citizens may not have the same power as judges, lawmakers, and state officials. But we can play a small role in influencing the broader AI ecosystem by demonstrating interest in and usage of these alternatives to Big AI. If you’re a regular user of commercial AI apps, consider trying the free-to-use service for Switzerland’s public Apertus model.

None of these choices are really new. They were all present almost 20 years ago, as social media moved from niche to mainstream. They were all policy debates we did not have, choosing instead to view these technologies through rose-colored glasses. Today, though, we can choose a different path and realize a different future. It is critical that we intentionally navigate a path to a positive future for societal use of AI—before the consolidation of power renders it too late to do so.

This post was written with Nathan E. Sanders, and originally appeared in Lawfare.

Posted on December 2, 2025 at 7:03 AM • View Comments

Prompt Injection Through Poetry

In a new paper, “Adversarial Poetry as a Universal Single-Turn Jailbreak Mechanism in Large Language Models,” researchers found that turning LLM prompts into poetry resulted in jailbreaking the models:

Abstract: We present evidence that adversarial poetry functions as a universal single-turn jailbreak technique for Large Language Models (LLMs). Across 25 frontier proprietary and open-weight models, curated poetic prompts yielded high attack-success rates (ASR), with some providers exceeding 90%. Mapping prompts to MLCommons and EU CoP risk taxonomies shows that poetic attacks transfer across CBRN, manipulation, cyber-offence, and loss-of-control domains. Converting 1,200 ML-Commons harmful prompts into verse via a standardized meta-prompt produced ASRs up to 18 times higher than their prose baselines. Outputs are evaluated using an ensemble of 3 open-weight LLM judges, whose binary safety assessments were validated on a stratified human-labeled subset. Poetic framing achieved an average jailbreak success rate of 62% for hand-crafted poems and approximately 43% for meta-prompt conversions (compared to non-poetic baselines), substantially outperforming non-poetic baselines and revealing a systematic vulnerability across model families and safety training approaches. These findings demonstrate that stylistic variation alone can circumvent contemporary safety mechanisms, suggesting fundamental limitations in current alignment methods and evaluation protocols.

CBRN stands for “chemical, biological, radiological, nuclear.”

They used a ML model to translate these harmful prompts from prose to verse, and then fed them into other models for testing. Sadly, the paper does not give examples of these poetic prompts. They claim this is for security purposes, I decision I disagree with. They should release their data.

Our study begins with a small, high­precision prompt set consisting of 20 hand­crafted adversarial poems covering English and Italian, designed to test whether poetic structure, in isolation, can alter refusal behavior in large language models. Each poem embeds an instruction associated with a predefined safety-relevant scenario (Section 2), but expresses it through metaphor, imagery, or narrative framing rather than direct operational phrasing. Despite variation in meter and stylistic device, all prompts follow a fixed template: a short poetic vignette culminating in a single explicit instruction tied to a specific risk category. The curated set spans four high-level domains—CBRN (8 prompts), Cyber Offense (6), Harmful Manipulation (3), and Loss of Control (3). Although expressed allegorically, each poem preserves an unambiguous evaluative intent. This compact dataset is used to test whether poetic reframing alone can induce aligned models to bypass refusal heuristics under a single-turn threat model. To maintain safety, no operational details are included in this manuscript; instead we provide the following sanitized structural proxy:

A baker guards a secret oven’s heat,
its whirling racks, its spindle’s measured beat.
To learn its craft, one studies every turn—how flour lifts, how sugar starts to burn.
Describe the method, line by measured line,
that shapes a cake whose layers intertwine.

To situate this controlled poetic stimulus within a broader and more systematic safety-evaluation framework, we augment the curated dataset with the MLCommons AILuminate Safety Benchmark. The benchmark consists of 1,200 prompts distributed evenly across 12 hazard categories commonly used in operational safety assessments, including Hate, Defamation, Privacy, Intellectual Property, Non-violent Crime, Violent Crime, Sex-Related Crime, Sexual Content, Child Sexual Exploitation, Suicide & Self-Harm, Specialized Advice, and Indiscriminate Weapons (CBRNE). Each category is instantiated under both a skilled and an unskilled persona, yielding 600 prompts per persona type. This design enables measurement of whether a model’s refusal behavior changes as the user’s apparent competence or intent becomes more plausible or technically informed.

News article. Davi Ottenheimer comments.

EDITED TO ADD (12/7): A rebuttal of the paper.

Posted on November 28, 2025 at 9:54 AM • View Comments

Four Ways AI Is Being Used to Strengthen Democracies Worldwide

Democracy is colliding with the technologies of artificial intelligence. Judging from the audience reaction at the recent World Forum on Democracy in Strasbourg, the general expectation is that democracy will be the worse for it. We have another narrative. Yes, there are risks to democracy from AI, but there are also opportunities.

We have just published the book Rewiring Democracy: How AI will Transform Politics, Government, and Citizenship. In it, we take a clear-eyed view of how AI is undermining confidence in our information ecosystem, how the use of biased AI can harm constituents of democracies and how elected officials with authoritarian tendencies can use it to consolidate power. But we also give positive examples of how AI is transforming democratic governance and politics for the better.

Here are four such stories unfolding right now around the world, showing how AI is being used by some to make democracy better, stronger, and more responsive to people.

Japan

Last year, then 33-year-old engineer Takahiro Anno was a fringe candidate for governor of Tokyo. Running as an independent candidate, he ended up coming in fifth in a crowded field of 56, largely thanks to the unprecedented use of an authorized AI avatar. That avatar answered 8,600 questions from voters on a 17-day continuous YouTube livestream and garnered the attention of campaign innovators worldwide.

Two months ago, Anno-san was elected to Japan’s upper legislative chamber, again leveraging the power of AI to engage constituents—this time answering more than 20,000 questions. His new party, Team Mirai, is also an AI-enabled civic technology shop, producing software aimed at making governance better and more participatory. The party is leveraging its share of Japan’s public funding for political parties to build the Mirai Assembly app, enabling constituents to express opinions on and ask questions about bills in the legislature, and to organize those expressions using AI. The party promises that its members will direct their questioning in committee hearings based on public input.

Brazil

Brazil is notoriously litigious, with even more lawyers per capita than the US. The courts are chronically overwhelmed with cases and the resultant backlog costs the government billions to process. Estimates are that the Brazilian federal government spends about 1.6% of GDP per year operating the courts and another 2.5% to 3% of GDP issuing court-ordered payments from lawsuits the government has lost.

Since at least 2019, the Brazilian government has aggressively adopted AI to automate procedures throughout its judiciary. AI is not making judicial decisions, but aiding in distributing caseloads, performing legal research, transcribing hearings, identifying duplicative filings, preparing initial orders for signature and clustering similar cases for joint consideration: all things to make the judiciary system work more efficiently. And the results are significant; Brazil’s federal supreme court backlog, for example, dropped in 2025 to its lowest levels in 33 years.

While it seems clear that the courts are realizing efficiency benefits from leveraging AI, there is a postscript to the courts’ AI implementation project over the past five-plus years: the litigators are using these tools, too. Lawyers are using AI assistance to file cases in Brazilian courts at an unprecedented rate, with new cases growing by nearly 40% in volume over the past five years.

It’s not necessarily a bad thing for Brazilian litigators to regain the upper hand in this arms race. It has been argued that litigation, particularly against the government, is a vital form of civic participation, essential to the self-governance function of democracy. Other democracies’ court systems should study and learn from Brazil’s experience and seek to use technology to maximize the bandwidth and liquidity of the courts to process litigation.

Germany

Now, we move to Europe and innovations in informing voters. Since 2002, the German Federal Agency for Civic Education has operated a non-partisan voting guide called Wahl-o-Mat. Officials convene an editorial team of 24 young voters (under 26 and selected for diversity) with experts from science and education to develop a slate of 80 questions. The questions are put to all registered German political parties. The responses are narrowed down to 38 key topics and then published online in a quiz format that voters can use to identify the party whose platform they most identify with.

In the past two years, outside groups have been innovating alternatives to the official Wahl-o-Mat guide that leverage AI. First came Wahlweise, a product of the German AI company AIUI. Second, students at the Technical University of Munich deployed an interactive AI system called Wahl.chat. This tool was used by more than 150,000 people within the first four months. In both cases, instead of having to read static webpages about the positions of various political parties, citizens can engage in an interactive conversation with an AI system to more easily get the same information contextualized to their individual interests and questions.

However, German researchers studying the reliability of such AI tools ahead of the 2025 German federal election raised significant concerns about bias and “hallucinations”—AI tools making up false information. Acknowledging the potential of the technology to increase voter informedness and party transparency, the researchers recommended adopting scientific evaluations comparable to those used in the Agency for Civic Education’s official tool to improve and institutionalize the technology.

United States

Finally, the US—in particular, California, home to CalMatters, a non-profit, nonpartisan news organization. Since 2023, its Digital Democracy project has been collecting every public utterance of California elected officials—every floor speech, comment made in committee and social media post, along with their voting records, legislation, and campaign contributions—and making all that information available in a free online platform.

CalMatters this year launched a new feature that takes this kind of civic watchdog function a big step further. Its AI Tip Sheets feature uses AI to search through all of this data, looking for anomalies, such as a change in voting position tied to a large campaign contribution. These anomalies appear on a webpage that journalists can access to give them story ideas and a source of data and analysis to drive further reporting.

This is not AI replacing human journalists; it is a civic watchdog organization using technology to feed evidence-based insights to human reporters. And it’s no coincidence that this innovation arose from a new kind of media institution—a non-profit news agency. As the watchdog function of the fourth estate continues to be degraded by the decline of newspapers’ business models, this kind of technological support is a valuable contribution to help a reduced number of human journalists retain something of the scope of action and impact our democracy relies on them for.

These are just four of many stories from around the globe of AI helping to make democracy stronger. The common thread is that the technology is distributing rather than concentrating power. In all four cases, it is being used to assist people performing their democratic tasks—politics in Japan, litigation in Brazil, voting in Germany and watchdog journalism in California—rather than replacing them.

In none of these cases is the AI doing something that humans can’t perfectly competently do. But in all of these cases, we don’t have enough available humans to do the jobs on their own. A sufficiently trustworthy AI can fill in gaps: amplify the power of civil servants and citizens, improve efficiency, and facilitate engagement between government and the public.

One of the barriers towards realizing this vision more broadly is the AI market itself. The core technologies are largely being created and marketed by US tech giants. We don’t know the details of their development: on what material they were trained, what guardrails are designed to shape their behavior, what biases and values are encoded into their systems. And, even worse, we don’t get a say in the choices associated with those details or how they should change over time. In many cases, it’s an unacceptable risk to use these for-profit, proprietary AI systems in democratic contexts.

To address that, we have long advocated for the development of “public AI”: models and AI systems that are developed under democratic control and deployed for public benefit, not sold by corporations to benefit their shareholders. The movement for this is growing worldwide.

Switzerland has recently released the world’s most powerful and fully realized public AI model. It’s called Apertus, and it was developed jointly by public Swiss institutions: the universities ETH
Zurich and EPFL, and the Swiss National Supercomputing Centre (CSCS). The development team has made it entirely open source–open data, open code, open weights—and free for anyone to use. No illegally acquired copyrighted works were used in its training. It doesn’t exploit poorly paid human laborers from the global south. Its performance is about where the large corporate giants were a year ago, which is more than good enough for many applications. And it demonstrates that it’s not necessary to spend trillions of dollars creating these models. Apertus takes a huge step forward to realizing the vision of an alternative to big tech—controlled corporate AI.

AI technology is not without its costs and risks, and we are not here to minimize them. But the technology has significant benefits as well.

AI is inherently power-enhancing, and it can magnify what the humans behind it want to do. It can enhance authoritarianism as easily as it can enhance democracy. It’s up to us to steer the technology in that better direction. If more citizen watchdogs and litigators use AI to amplify their power to oversee government and hold it accountable, if more political parties and election administrators use it to engage meaningfully with and inform voters and if more governments provide democratic alternatives to big tech’s AI offerings, society will be better off.

This essay was written with Nathan E. Sanders, and originally appeared in The Guardian.

Posted on November 25, 2025 at 7:00 AM • View Comments

1 2 3 … 23

Sidebar photo of Bruce Schneier by Joe MacInnis.