Our weekly picks of digital trust-related articles

After many thought that LLMs and image-generators will remove jobs from writers and image artists, the pendulum swings back: clients realize that these tools only get you halfway to a useful result. So they turn to the ones they wanted to replace, and ask them to fix the half-baken results. I find i interesting how some of the artists don't want to do these fix-up work - because it pays bad, and because they prefer starting from scratch.

I follow the advances of quantum computers with great interest, mainly because I'm curious when, or if, they will ever be able to break current cryptography algorithms. The holy grail of the algorithms is called 'Shor's algorithm', which can factorize numbers quickly. Already in 2001, a quantum computer factorized 15! Yet since then, no quantum computer has managed to factorize larger numbers, like 21. This article explains why 15 was easy, and why 21 is much harder. Instead of 21 operations, 2,405 operations are needed, or 240'500 operations if error correction is included.

While it was to be expected that Anthropic will also use the users' chats for training, I think the way they're approaching this is not too bad. Perhaps the pop-up is not clear enough, but at least past chats will not get in the LLM training grinder. One of the big question will be of course the part where they'll try to remove any sensitive data from the chats. That one will most probably end badly.

I appreciate this second, more critical, take on the E-ID from dnip for moving beyond a purely technical discussion or a legal dissection, and instead exploring the broader societal impact of the E-ID.

I particularly enjoyed this article because it challenges today’s automation-at-all-costs mindset, urging us to prioritize human-AI collaboration over replacement, with the goal that AI plus human expertise exceeds what AI can achieve alone. Learning when to collaborate versus automate is vital for more trustworthy and effective outcomes.

With Elon Musk's vision of turning X into the 'everything app', this paper on the unprecedented scale of how Tencent's WeChat - one of the super-apps clearly inspiring Musk - is tracking its users, should be a heads-up for privacy advocates.

The author highlights a crucial point that is often neglected by technology experts, entrepreneurs and government authorities alike: how the public understands and feels about this emerging technology. This article captures genuine opinions—optimism, anxiety, and confusion—from people outside the immediate quantum research community.

A successful—and almost uplifting—example of collaboration across law enforcement, government agencies, and businesses against cybercrime is exemplified by the Cybercrime Atlas project. In a sweeping INTERPOL-coordinated operation, authorities across Africa arrested 1,209 cybercriminals who targeted nearly 88,000 victims. The crackdown recovered USD 97.4 million and dismantled 11,432 malicious infrastructures. This operation demonstrates how cross-border collaboration can alleviate the human cost of global cybercrime.

Traditional cybersecurity is by itself a nontrivial problem to address, and even big companies with dedicated budgets and teams are not immune to attacks. In an already precarious security landscape, adding AI to the mix has not only the potential to make a bad situation worse, but truly disastrous, as this excellent write-up of the current state of cybersecurity in agentic AI demonstrates.

Whether you're for or against the proposed E-ID, a public discussion is the healthiest if it is founded on factually correct arguments. While this piece is clearly opinionated, it also tries to examine the main arguments from the opposition as neutrally as possible, and provides a good explanation and discussion for each of them, pointing out errors, but also conceding where valid arguments are made.

The Israeli airstrike campaign against Iranian military and cyber infrastructure on 12 June had an 'interesting' side effect. Accounts that had previously been identified as allegedly being managed by the Iranian Revolutionary Guard Corps (IRGC) and that promoted Scottish independence fell silent following the strikes. This resulted in a 4% reduction in all discussion related to Scottish independence on X, once again proving the existence of sophisticated, state-sponsored influence operations that use fake personas to manipulate democratic discourse. It also highlights the infrastructure required to run such operations.

I found this article interesting because, rather than perpetuating fear-driven narratives, it provides a thorough analysis backed by demographic realities in the Western world. Labour shortages, it suggests, make it unlikely that AI will 'take all our jobs'. It emphasises how AI can increase access to specialist roles for a wider range of workers. The real challenge, then, lies in ensuring that enough people gain this 'middle' level of expertise, so that society can benefit fully from AI-augmented productivity without eroding true expertise.

Like many legislative measures intended to protect users in the digital world, age-verification laws have noble aims. However, their practical implementation is messy. As is often the case, the solution not only likely involves privacy-preserving tech, but also requires compliance, enforcement, and government investment in educational initiatives to be effective.

As the speed that articles can be generated has increased with the rise of LLMs, Wikipedia just increased the speed it can delete them. If an article has obvious signs of having been created by an LLM, it can be deleted quickly with only two approvals. I think this is a good way of making sure that the articles on Wikipedia are high quality: remove the slop (bad quality text) as fast as possible.

Scandalous examples such as the fake video of Obama being arrested are a bleak reminder of not only what is possible today, but also that these technologies are becoming more and more commonplace in our daily lives. So what are the current state-of-the-art techniques for detecting deepfakes? On the occasion of this year's DEFCON, this article aims to provide a bird's eye view.

This article prompts reflection on what we mean by 'trust' when we talk about 'trustworthy' AI. There are many dimensions to trust, and the author helpfully breaks them down. In human-AI interactions, misalignments can occur when stakeholders interpret 'trust' differently. For example, companies might emphasize the epistemic aspect—reliance on knowledge and its acquisition—of trust, while users may view AI outputs through a moral, value-based lens. To develop genuinely trustworthy AI, it is essential to establish a common understanding among stakeholders of all the layers of trust

It often appears as if disinformation was spread by a large number of social media users. However, research suggests that it is a comparatively small percentage of the users that is primarily responsible for creating and widely sharing divisive content, with these voices being amplified by the platforms’ algorithms. As bleak as this may be, I find a silver lining in the experiments that showed that unfollowing such accounts lead to a substantial reduction in animosity towards other political groups.

An impressively large line-up of AI leaders and experts are advocating for more research into 'chain-of-thought' (CoT) monitoring of reasoning models. This technique, as the name implies, aims to understand how AI reasoning models work. It could become a key method for understanding how AI agents think and what their goals are, and could enhance AI safety by tracking alignment and flagging harmful interactions. Presently, visibility into these models’ 'thought processes' exists but, apparently, is not guaranteed, hence the call to arms before this window is closed.

With all the hype around agentic AI, the industry is rushing to embrace it. However, alarm bells have been sounded again and again concerning misaligned behaviour of LLMs and Large Reasoning Models (LRMs), ranging from 'harmless' misinformation to deliberately malicious actions. This raises serious questions whether the current technology is really mature enough to be entrusted with any level of autonomy, let alone replace human workers

From a cryptographer's point of view, the big breakthrough in quantum computing would be if it can successfully factorize numbers in the 1000-digit range. As it turns out, this is actually quite difficult. The record from 2012 of factorizing the number 21 is still unbeaten! And all reports of factorizing bigger numbers used very, very specific numbers, so specific that even a dog can factorize them! Progress in quantum computers is advancing, but we may have to wait one or more decades before we really start to reap their promise.

Severe floods in Texas sparked a wave of conspiracy theories, with claims circulating online that the disaster was caused by geoengineering or weather weapons. This highlights a growing tension between the speed at which formal institutions can communicate accurate information and the rapid spread of AI-fueled disinformation. The resulting vandalism of radar infrastructure and threats against meteorologists disrupted emergency services, showing just how vulnerable necessary communication channels are in times of emergencies.

A lot of cryptographic proofs rely on something called the 'random oracle model' and the 'Fiat-Shamir transformation'. Together, they can create a mathematical proof of the security of a specific zero knowledge protocol. However, the random oracle model is never used – in real algorithms, it is replaced by a hash function. What can go wrong? It turns out, a lot. You can create a zero knowledge proof of a lie, e.g., prove that you have the cryptocurrency funds necessary for a certain transaction, when in reality you have not!

As a software engineer, I'm looking at LLMs both as a tool for, but potentially also a danger to, my job: will it replace me one day? In this study, they measured the time that seasoned software needed to fix a bug, both with and without the aid of LLMs. The outcome in this specific study was surprising – it took longer to solve a problem with an LLM than without! There are a lot of caveats in the study, but it's interesting to see that it's not very clear if LLMs are even a help, let alone a danger to my job!

Am Beispiel deutscher Medien wird die Entwicklung des Online-Journalismus seit den 90ern bis heute nachvollzogen. Besonders interessant finde ich die Diskussion der jeweiligen Problematiken denen die Verlage im Web 1.0, in den sozialen Medien und heute mit generativer AI konfrontiert sind/waren.