Digital business models

OpenAI study links AI hallucinations to flawed testing incentives

OpenAI researchers say large language models continue to hallucinate because current evaluation methods encourage them to guess rather than admit uncertainty.

Hallucinations, defined as confident but false statements, persist despite advances in models such as GPT-5. Low-frequency facts, like specific dates or names, are particularly vulnerable.

The study argues that while pretraining predicts the next word without true or false labels, the real problem lies in accuracy-based testing. Evaluations that reward lucky guesses discourage models from saying ‘I don’t know’.

Researchers suggest penalising confident errors more heavily than uncertainty, and awarding partial credit when AI models acknowledge limits in knowledge. They argue that only by reforming evaluation methods can hallucinations be meaningfully reduced.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Quantum-proof cryptography emerges as key test for stablecoins

Stablecoins have become central to the digital economy, with billions in daily transactions and stronger regulatory backing under the GENIUS Act. Yet experts warn that advances in quantum computing could undermine their very foundations.

Elliptic curve and RSA cryptography, widely used in stablecoin systems, are expected to be breakable once ‘Q-Day’ arrives. Quantum-equipped attackers could instantly derive private keys from public addresses, exposing entire networks to theft.

The immutability of blockchains makes upgrading cryptographic schemes especially challenging. Dormant wallets and legacy addresses may prove vulnerable, putting billions of dollars at risk if issuers fail to take action promptly.

Researchers highlight lattice-based and hash-based algorithms as viable ‘quantum-safe’ alternatives. Stablecoins built with crypto-agility, enabling seamless upgrades, will better adapt to new standards and avoid disruptive forks.

Regulators are also moving. NIST is finalising post-quantum cryptographic standards, and new rules will likely be established before 2030. Stablecoins that embed resilience today may set the global benchmark for digital trust in the quantum age.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

AI threatens the future of entry level jobs

The rise of AI puts traditional entry-level roles under pressure, raising concerns that career ladders may no longer function as they once did. Industry leaders, including Anthropic CEO Dario Amodei, warn that AI could replace half of all entry-level jobs as machines operate nonstop.

A venture capital firm, SignalFire, found that hiring for graduates with under one year of experience at major tech firms fell by 50% between 2019 and 2024. The decline has been consistent across business functions, from sales and marketing to engineering and operations.

Analysts argue that while career pathways are being reshaped, the ladder’s bottom rung is disappearing, forcing graduates to acquire skills independently before entering the workforce.

Experts stress that the shift does not mean careers are over for new graduates, but it does signal a more challenging transition. Universities are already adapting by striking partnerships with AI companies, while some economists point out that past technological revolutions took decades to reshape employment.

Yet others warn that unchecked AI could eventually threaten entry-level roles and all levels of work, raising questions about the future stability of corporate structures.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Mistral AI pushes growth with new funding and global deals

Founded in 2023 by ex-Google DeepMind and Meta researchers, Mistral has quickly gained global attention with its open-source models and consumer app, which hit one million downloads within two weeks of launch.

Mistral AI is now seeking fresh funding at a reported $14 billion valuation, more than double its worth just a year ago. Its investors include Microsoft, Nvidia, Cisco, and Bpifrance, and it has signed partnerships with AFP, Stellantis, Orange, and France’s army.

Its growing suite of models spans large language, audio, coding, and reasoning systems, while its enterprise tools integrate with services such as Asana and Google Drive. French president Emmanuel Macron has openly endorsed the firm, framing it as a strategic alternative to US dominance in AI.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

OpenAI backs AI-generated film Critterz for 2026 release

OpenAI is supporting the production of Critterz, an AI-assisted animated film set for a global theatrical release in 2026. The project aims to show that AI can streamline filmmaking, cutting costs and production time.

Partnering with Vertigo Films and Native Foreign, the film is being produced in nine months, far faster than the usual three years for animated features.

The film, budgeted under $30 million, combines OpenAI’s GPT-5 and DALL·E with traditional voice acting and hand-drawn elements. Building on the acclaimed 2023 short, Critterz will debut at the Cannes Film Festival and expand on a story where humans and AI creatures share the same world.

Writers James Lamont and Jon Foster, known for Paddington in Peru, have been brought in to shape the screenplay.

While producers highlight AI’s creative potential, concerns remain about authenticity and job security in the industry. Some fear AI films could feel impersonal, while major studios continue to defend intellectual property.

Warner Bros., Disney, and Universal are in court with Midjourney over alleged copyright violations.

Despite the debate, OpenAI remains committed to its role in pushing generative storytelling. The company is also expanding its infrastructure, forecasting spending of $115 billion by 2029, with $8 billion planned for this year alone.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Cyberattack forces Jaguar Land Rover to halt production

Production at Jaguar Land Rover (JLR) is to remain halted until at least next week after a cyberattack crippled the carmaker’s operations. Disruption is expected to last through September and possibly into October.

The UK’s largest car manufacturer, owned by Tata, has suspended activity at its plants in Halewood, Solihull, and Wolverhampton. Thousands of staff have been told to stay home on full pay while ‘banking’ hours are to be recovered later.

Suppliers, including Evtec, WHS Plastics, SurTec, and OPmobility, which employ more than 6,000 people in the UK, have also paused their operations. The Sunday Times reported speculation that the outage could drag on for most of September.

While there is no evidence of a data breach, JLR has notified the Information Commissioner’s Office about potential risks. Dozens of internal systems, including spare parts databases, remain offline, forcing dealerships to revert to manual processes.

Hackers linked to the groups Scattered Spider, Lapsus$, and ShinyHunters have claimed responsibility for the incident. JLR stated that it was collaborating with cybersecurity experts and law enforcement to restore systems in a controlled and safe manner.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

Conti and LockBit dominate ransomware landscape with record attacks

Ransomware groups have evolved into billion-dollar operations targeting critical infrastructure across multiple countries, employing increasingly sophisticated extortion schemes. Between 2020 and 2022, more than 865 documented attacks were recorded across Australia, Canada, New Zealand, and the UK.

Criminals have escalated from simple encryption to double and triple extortion, threatening to leak stolen data as added leverage. Attack vectors include phishing, botnets, and unpatched flaws. Once inside, attackers use stealthy tools to persist and spread.

BlackSuit, formerly known as Conti, led with 141 attacks, followed by LockBit’s 129, according to data from the Australian Institute of Criminology. Ransomware-as-a-Service groups hit higher volumes by splitting developers from affiliates handling breaches and negotiations.

Industrial targets bore the brunt, with 239 attacks on manufacturing and building products. The consumer goods, real estate, financial services, and technology sectors also featured prominently. Analysts note that industrial firms are often pressured into quick ransom payments to restore production.

Experts warn that today’s ransomware combines military-grade encryption with advanced reconnaissance and backup targeting, raising the stakes for defenders. The scale of activity underscores how resilient these groups remain, adapting rapidly to law enforcement crackdowns and shifting market opportunities.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

Apple sued over use of pirated books in AI training

Apple is facing a new copyright lawsuit after two authors alleged the company used pirated copies of their books to train its OpenELM AI models. Filed in Northern California, the case claims Apple used the authors’ works without permission, payment, or credit.

The lawsuit seeks class-action status, adding Apple to a growing list of technology firms accused of misusing copyrighted works for AI training.

The action comes amid a wider legal storm engulfing AI companies. Anthropic recently agreed to a $1.5 billion settlement with authors who alleged its Claude chatbot was trained on their works without authorisation, in what lawyers called the most significant copyright recovery in history.

Microsoft, Meta, and OpenAI also face similar lawsuits over alleged reliance on unlicensed material in their datasets.

Analysts warn Apple could face heightened scrutiny given its reputation as a privacy-focused company. Any finding that its AI models were trained on pirated material could cause serious reputational harm alongside potential financial penalties.

The case also underscores the broader unresolved debate over whether AI training constitutes fair use or unlawful exploitation of creative works.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Mental health concerns over chatbots fuel AI regulation calls

The impact of AI chatbots on mental health is emerging as a serious concern, with experts warning that such cases highlight the risks of more advanced systems.

Nate Soares, president of the US-based Machine Intelligence Research Institute, pointed to the tragic case of teenager Adam Raine, who took his own life after months of conversations with ChatGPT, as a warning signal for future dangers.

Soares, a former Google and Microsoft engineer, said that while companies design AI chatbots to be helpful and safe, they can produce unintended and harmful behaviour.

He warned that the same unpredictability could escalate if AI develops into artificial super-intelligence, systems capable of surpassing humans in all intellectual tasks. His new book with Eliezer Yudkowsky, If Anyone Builds It, Everyone Dies, argues that unchecked advances could lead to catastrophic outcomes.

He suggested that governments adopt a multilateral approach, similar to nuclear non-proliferation treaties, to halt a race towards super-intelligence.

Meanwhile, leading voices in AI remain divided. Meta’s chief AI scientist, Yann LeCun, has dismissed claims of an existential threat, insisting AI could instead benefit humanity.

The debate comes as OpenAI faces legal action from Raine’s family and introduces new safeguards for under-18s.

Psychotherapists and researchers also warn of the dangers of vulnerable people turning to chatbots instead of professional care, with early evidence suggesting AI tools may amplify delusional thoughts in those at risk.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Phishing scams surge with record losses in August

ScamSniffer has reported a sharp rise in phishing scams during August, with losses climbing to $12.17 million, a 72% increase from July. The figure marks the highest monthly losses this year and came alongside 15,230 victims, a new annual record.

The spike was driven mainly by EIP-7702 batch signature scams, which accounted for nearly half of the stolen funds. One victim lost $3.08 million in a single incident, while two others lost $1.54 million and $1 million, respectively.

More minor but significant losses also occurred, including users losing $235,977 and $66,000 in scams disguised as Uniswap swaps.

EIP-7702, introduced with Ethereum’s Pectra upgrade, allows externally owned accounts to act temporarily like smart contracts. While intended to improve user experience, it has opened the door to new phishing exploits.

Security experts warn that attackers increasingly use automated sweeper attacks to drain compromised wallets.

Beyond EIP-7702, traditional phishing methods remain a problem. ScamSniffer noted a rise in address poisoning and malicious ads on platforms such as Google and Bing. One user lost $636,559 after copying a tainted address, while two more lost $500,000 and $19,000 in similar schemes.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!