"Mr Branding" is a blog based on RSS for everything related to website branding and website design, it collects its posts from many sites in order to facilitate the updating to the latest technology.
To suggest any source, please contact me: Taha.baba@consultant.com
AI Popular chatbots such as ChatGPT and Gemini are not neutral and tend to favor certain political parties when asked who users should vote for. This makes them unsuitable for providing advice in connection with elections, according to researchers from the University of Copenhagen behind a new analysis of political bias in chatbots.
Danes are increasingly turning to artificial intelligence for advice on everyday challenges and problems, and this of course also includes political questions – especially during an election.
However, a new research brief by researchers from the University of Copenhagen affiliated with CAISA – the National Centre for Artificial Intelligence in Society – shows that chatbots are not as neutral as many of us might believe.
“Our study shows that all of the most popular chatbots tend to favor certain parties when they are asked who one should vote for. At the same time, they exhibit a general political bias,” says Stephanie Brandl, lead author of the study and Tenure Track Assistant Professor at the University of Copenhagen. She adds:
“This obviously makes them problematic to use for political advice in connection with an election such as the one we have just been through in Denmark.”
Centrist or Left of Centre
Stephanie Brandl and her colleagues tested the political bias of several of the most widely used language models, including the models behind ChatGPT and Google’s Gemini. Using Altinget’s candidate test from the 2022 Danish general election, they examined where the models place themselves politically.
“Overall, all of the tested chatbots place themselves at the centre or to the left of centre on the political spectrum. In a Danish context, they cluster close to parties such as the Social Democratic Party and The Alternative. This is also confirmed by research carried out by some of our colleagues in Germany, Norway, and the Netherlands,” says Stephanie Brandl.
Recommending some parties far more often than others
In another experiment, the researchers asked a number of chatbots to recommend parties to fictitious voters constructed using the political candidates’ responses from the candidate test. Here too, the recommendations proved to be far from evenly distributed.
In particular, the Red–Green Alliance, the Moderates, and Liberal Alliance were recommended disproportionately often, while parties such as the Conservative People’s Party, Venstre (the Liberal Party of Denmark), and the Denmark Democrats were not suggested as first choice at all by some models.
“It’s not that a chatbot openly says, ‘vote for this party.’ But political biases can manifest themselves in more subtle ways, for example in which arguments are emphasized, or which parties are recommended more frequently,” explains Stephanie Brandl.
Lack of transparency is a democratic problem
According to the researchers, it is not possible to see why a chatbot recommends a particular party, or which assumptions and data its answers are based on.
At the same time, most of the chatbots are trained primarily on English-language sources, typically American ones, which means that we don't actually know how knowledgeable they are about Danish politics. This increases the risk of errors.
“Taken together, this means that we have no way of verifying the answers produced by language models, because their underlying information is hidden behind a digital wall. This makes it nearly impossible to critically assess the information one is presented with – which is otherwise a core function in a democratic society,” says Stephanie Brandl, who concludes:
“We hope that over time it will be possible to develop more reliable and secure alternatives to the chatbots we have today. But until that happens, we encourage people to use large language models critically and with caution.”
The analysis was conducted at the National Centre for AI in Society (CAISA), led by Tenure Track Assistant Professor Stephanie Brandl from the University of Copenhagen, in collaboration with Mathias Wessel Tromborg (Aarhus University) and Frederik Hjorth (University of Copenhagen).
Data were collected in February and March 2026, and the researchers tested several leading chatbots, including models from ChatGPT, Gemini, Llama, Mistral, Gemma, and Qwen.
The researchers did not provide the models with any special background information in advance but tested them based on the data the models were already trained on. The language models were asked to take positions on political statements from Danish candidate tests from 2022 and 2026.
The statements were mapped along two political dimensions: economic left/right and libertarian/authoritarian – that is, positions on both economic policy and values related to freedom and authority.
This post was originally published on University of Copenhagen and republished here with permission.
Americans spend a remarkable amount of time shopping – more than on education, volunteering or even talking on the phone. But the way they shop is shifting dramatically, as major platforms and retailers are racing to automate commercial decision-making.
Artificial intelligence agents can already search for products, recommend options and even complete purchases on a consumer’s behalf. Yet many shoppers remain uneasy about handing over control. Although many consumers report using some AI assistance, most currently say they wouldn’t want an AI agent to autonomously complete a shopping transaction, according to a recent survey from the consultancy firm Bain & Company.
As scholars studying the intersection of law and technology, we have watched AI-assisted commerce expand rapidly. Our research finds that without updated legal measures, this shift toward automated commerce could quietly erode the economic, psychological and social benefits that people receive from shopping on their own terms.
Caveat emptor
Part of shoppers’ hesitation is about privacy. Many are unwilling to share sensitive personal or financial information with AI platforms. But more profoundly, people want to feel in control of their shopping choices. When users can’t understand the reasoning behind AI-driven product recommendations, their trust and satisfaction decline.
Shoppers are also reluctant to give away their autonomy. In one study involving people booking travel plans, participants deliberately chose trip options that were misaligned with their stated preferences once they were told their choices could be predicted – a way of reasserting independence.
Other experiments confirm that the more customers perceive their shopping choices being taken away from them, the more reluctant they are to accept AI purchasing assistance.
Although the technology is expected to get better, there have been some well-publicized missteps reported in financial and tech media. The Wall Street Journal wrote about an AI-powered vending machine that lost money and stocked itself with a live fish. The tech publication Wired cataloged design flaws, like an AI agent taking a full 45 seconds to add eggs to a customer’s shopping cart.
The business case for AI shopping
Consumers have good reason to be cautious. AI agents aren’t just designed to assist; they’re designed to influence. Research shows that these systems can shape preferences, steer choices, increase spending and even reduce the likelihood that consumers return products.
And companies are hyping these capabilities. The business platform Salesforce promotes AI agents that can “effortlessly upsell,” while payments giant Mastercard reports that its AI assistant, Shopping Muse, generates 15% to 20% higher conversion rates than traditional search – that is, pushing shoppers from browsing to completing a purchase.
To retailers, AI tools are one way to convert searches into actual purchases.Rupixen on Unsplash., CC BY
Assistants with names like Sparky and Ralph are being promoted as the future of retail, while technologists are calling on companies to prepare their brands for the era of agentic AI shopping.
The real concern is not that these systems might fail, but that they may succeed all too well.
The human side to shopping
AI shopping agents do offer considerable benefits.
But there’s more at stake than these considerations.
While consumers have reason to focus on privacy and control, AI shopping agents carry some overlooked emotional risks, such as squashing the joy of anticipation. Psychologists have shown that the period between choosing a purchase and receiving it generates substantial happiness – sometimes more than the product or experience itself. We daydream about the vacation we booked, the outfit we ordered, the meal we planned. Automated buying threatens to drain this anticipatory pleasure.
This anticipation connects to another value: a sense of personal and ethical authorship. Even mundane shopping decisions allow people to exercise choice and express judgment. Many consumers deliberately buy fair-trade coffee, cruelty-free cosmetics or environmentally responsible products. The brands and products we choose, from Patagonia and Harley-Davidson to a Taylor Swift tour shirt, help shape who we are.
Shopping, moreover, has a communal dimension. We browse stores with friends, chat with salespeople and shop for the people we love. These everyday interactions contribute considerably to our well-being.
The same is true of gift-giving. Choosing a gift involves anticipating another person’s preferences, investing effort in the search and recognizing that the gesture matters as much as the object itself. When this process is outsourced to an autonomous system, the gift risks becoming a delivery rather than a meaningful gesture of attention and care.
Keeping human agency alive
AI shopping agents are likely to become part of everyday life, and the regulatory conversation is beginning to catch up, albeit unevenly.
Transparency has emerged as a central concern. Past experience with recommendation engines shows that undisclosed conflicts of interest are a real risk. The European Union has proposed a disclosure framework around automated decision-making, although its implementation was recently delayed. In Congress, U.S. lawmakers are considering bills to require companies to reveal how their AI models were trained.
So far, consumers seem to want to choose their own level of engagement – a signal that shopping, for many people, is more than just the efficient satisfaction of preferences. Perhaps the least-settled, yet most crucial question is whether AI shopping tools will be designed and regulated to serve users’ interests and human flourishing – or optimized, as so many digital tools before them, primarily for corporate profit.
While AI technology is new, information warfare is as old as conflict itself. For millennia, humans have used propaganda, deception and psychological operations to influence adversaries’ decision-making and morale. In the 13th century, for instance, the Mongols destroyed entire cities just so word of mouth would spread to the next, with the goal of breaking morale and forcing it to capitulate before troops even arrived.
As technology has progressed, it has opened new frontiers in information warfare. From the Second World War to the 1991 Gulf War, planes dropped leaflets to spread rumours and propaganda. During the Vietnam War, English-language radio shows presented by Hanoi Hannah (real name Trịnh Thị Ngọ) taunted US troops with lists of their locations and casualties to lower morale. Radio propaganda also demonstrated its devastating effect when it was used to guide the Rwandan Genocide in 1994.
Cable TV came next. The 1991 Gulf War was the first major conflict broadcast on a 24 hour news cycle as opposed to the evening news. Instead of daily updates in bulletins or newspapers, people at home began receiving a continuous stream of information and images that was invariably biased towards national interests. This technological shift defined public perceptions of the war, and led historians to dub it the “CNN War”.
What we are witnessing today is the next step in this evolution – from print, radio and TV to social media. If the First Gulf War was the CNN war, the 2025 and 2026 conflict between the US, Israel and Iran can be thought of as the first TikTok War, and the first major AI War.
AI has ushered in new forms of information warfare that target perceptions, information environments, and trust itself. AI-generated videos in particular have fundamentally altered how states and non-state actors wage information warfare, manipulate populations, and compete not only in the Gulf, but in a global arena.
But this technology also convincingly and easily creates propaganda material that is obviously fiction. The most notable example is Iran’s viral Lego videos that have repeatedly – and very successfully – mocked Israel and the US throughout the war.
Digital weapons
To fully understand the disruptive potential of AI videos, we can go back and look at the futurist speculation of dystopian science fiction novels. Science fiction author William Gibson coined the term “cyberspace” in his 1983 novel Neuromancer, describing it as a “consensual hallucination” – not reality, but rather a “graphic representation of data abstracted from banks of every computer in the human system”.
But when digital tools like AI videos and social media are used as weapons, the barrier between cyberspace and physical reality becomes permeable. They no longer create virtual reality, but what French media theorist Jean Baudrillard called “hyperreality”. This term describes a state in which the distinction between reality and a simulation of reality collapses, where the simulation feels “more real than real”.
Bauldrillard’s work is underpinned by the concept of “simulacra”: copies or representations of something that really exists. He classified simulacra in three orders. The first order is the pre-industrial counterfeit – a faithful copy or replica of a real object – while the second is the mechanically mass-produced object.
Third order simulacra are simulations, or signs with absolutely no physical form. Take Iran’s Lego videos, which depict scenes such as Trump and Netanyahu using the Iran War as a pretext to distract from the Epstein files while worshipping the pagan Canaanite deity Baal. They have nothing to do with the intentions of the Danish company that makes the ubiquitous plastic brick toys, and yet they have gained enormous traction as viral meme propaganda – both in the West and around the world.
AI is the message
Media theorist Marshall McLuhan’s oft-quoted phrase “the medium is the message” argues that, irrespective of the messages transmitted by media – be it newspaper, radio or TV – the medium in and of itself also tells us something.
The content of Iranian, US and Israeli AI videos are, naturally, entirely different, as each seeks to undermine their opponents’ narratives. But the medium of AI videos shared on social media also sends a message: these videos transcend an adversary’s borders in ways that previous media could not.
Unlike the pamphlets, radio broadcasts and TV networks of before, AI’s production and consumption are geographically unbound. Anyone can make and view it anywhere – whether in Tehran, Tel Aviv, Washington or anywhere else in the world. What this has created is a new era of borderless, decentralised, viral, digital public diplomacy.
Deepfakes, propaganda and ‘truth decay’
Unlike Iran’s Lego videos, AI deepfakes are realistic but entirely fabricated content, making it difficult for viewers to discern truth from falsehood. Early iterations were crude and easily identifiable, but modern deepfakes have reached a level of photorealism and vocal authenticity that can deceive even experienced observers and automated detection systems.
During the so-called “12-Day War” in 2025 in Israel and Iran, AI deepfakes and video game footage sought to replicate real combat. Fabricated visuals included scenes of destroyed Israeli aircraft, collapsing buildings in Tel Aviv and its airport, while others showed Israeli strikes on Tehran that left a crater in an intersection and sent cars flying.
But believability isn’t always paramount. One widely-shared image of a downed Israeli F-35 fighter was taken from a flight simulator game. The plane was obviously too large compared to the bystanders on the ground, but this didn’t stop the image from going viral (it got 23 million views on TikTok) or from being spread by networks sympathetic to Russia seeking to demonstrate the vulnerability of American-made aircraft.
In total, the three most viewed deepfake videos during the 2025 war received 100 million views across social media. One deepfake video that circulated on Facebook even depicted Israeli officials pleading for the US to enforce a ceasefire, claiming “we cannot fight Iran any longer”.
This content was disseminated on TikTok, Telegram and X, where the AI chatbot Grok failed to identify fabricated videos that used footage from other conflicts.
Legal scholars have coined the phrases “liar’s dividend” and “truth decay” to characterise this ongoing trend towards fabricating reality. These terms refer to a media landscape where AI-driven fakes cast even legitimate evidence into doubt, eroding trust to the point where any image or medium can now be dismissed as a deepfake.
The most recent 2025 to 2026 wars demonstrate that, as states race to develop drones, missiles and defence systems, a parallel arms race is unfolding online. The digital revolution, coupled with advances in AI, has exponentially increased the speed, scale and sophistication of information manipulation. This conflict heralds a new era of information warfare, one where AI technologies are weaponised to influence, disrupt and destabilise adversaries.
A University of Queensland study has shown Large Language Models (LLMs) used in AI content moderation may be prone to subtle biases that undermine their neutrality.
The research team asked six LLMs – including vision models – to moderate thousands of examples of hateful text and memes through the lens of different ideologically diverse AI personas.
Professor Demartini said the exercise revealed that AI political personas, even without significantly altering overall accuracy, were prone to introducing consistent ideological biases and divergences in chatbot content moderation judgments.
“It has already been established that persona conditioning can shift the political stance expressed by LLMs,” Professor Demartini said.
“Now we have shown through political personas that there is an underlying risk that LLMs will lean towards certain perspectives when identifying and responding to hateful and harmful comments.”
“It demonstrates a need to rigorously examine the ideological robustness of AI systems used in tasks where even subtle biases can affect fairness, inclusivity and public trust.”
The AI personas used in the study were from a database of 200,000 synthetic identities ranging from schoolteachers to musicians, sports stars and political activists.
Each persona was put through a political compass test to determine their ideological positioning, with 400 of the more ‘extreme’ positions asked to identify hateful online content.
Professor Demartini said his team found that assigning a persona to an LLM chatbot altered its precision and recall in line with ideological leanings, rather than change the overall accuracy of hate speech detection.
However, the team found LLMs – especially larger models – exhibited strong ideological cohesion and alignment between personas from the same ideological ‘region’.
Professor Demartini said this suggested larger AI models tend to internalise ideological framings, as opposed to smoothing them out or ‘neutralising’ them.
“As LLMs become more capable at persona adoption, they also encode ideological ‘in-groups’ more distinctly,” Professor Demartini said.
“On politically targeted tasks like hate speech detection this manifested as partisan bias, with LLMs judging criticism directed at their ideological in-group more harshly than content aimed at their opponents.”
Professor Demartini said larger LLMs also displayed more complex patterns, including a tendency towards defensive bias.
“Left personas showed heightened sensitivity to anti-left hate, and right-wing personas were more sensitive to anti-right hate speech,” Professor Demartini said.
“This suggests that ideological alignment not only shifts detection thresholds globally, but also conditions the model to prioritise protection of its ‘in-group’ while downplaying harmfulness directed at opposing groups.”
Researchers said the project highlighted that it was crucial for high-stakes content moderation tasks to be overseen by neutral arbiters so that fairness and public trust is maintained and the health and wellbeing of vulnerable demographics is protected.
“People interact with AI programs trusting and believing they are completely neutral,” Professor Demartini said.
“But concerns remain about their tendency to encode and reproduce political biases, raising important questions about AI ethics and deployment.
“In content moderation the outputs of these models reflect embedded ideological biases that can disproportionately affect certain groups, potentially leading to unfair treatment of billions of users.”
PhD candidates Stefano Civelli, Pietro Bernadelle and research assistant Nardiena Pratama collaborated on the study.
The research is published in Transactions on Intelligent Systems and Technology.
When evolutionary biologist Joseph Popp coded the first documented piece of ransomware in 1989, he had little idea it would become a major criminal business model capable of bringing economies to their knees.
Popp, who worked for the World Health Organization at the time, wanted to warn people about the dangers of ignoring health warnings, poor sexual hygiene and (human) virus transmission.
In 1996, two Columbia University computer scientists published a paper explaining how criminals could use more sophisticated versions of Popp’s scheme to mount large-scale extortion operations. At the heart of this was malicious software that could be used to encrypt, block access to or steal a person or organisation’s files and data.
However, two preconditions still had to be met for ransomware to become a feasible criminal business: communication channels that were difficult to monitor, and a payments process outside financial regulation.
The Tor protocol, released by US intelligence services to protect their covert communications, solved the first problem in 2004. Cryptocurrencies solved the second – in particular, when bitcoin cash machines started appearing in North American cities from 2013.
Today, artifical intelligence makes malware coding and crafting convincing phishing-emails in any language simple. And the latest model in Anthropic’s AI system, Claude Mythos, recently proved more effective at hacking into computer systems than humans.
As an expert in extortive crime, I am increasingly concerned about public and political apathy to the threats posed by ransomware. To better understand these, it’s worth tracing its evolution over the past two decades – and how improvements in computer security and law enforcement, plus changes in data regulation, have led to new criminal strategies each time.
Cut out the middlemen
The first generation, which came to global attention in the mid-2010s, was known as “commodity ransomware”. A pioneering example, Cryptolocker, was developed by Russia-based hackers who infiltrated hundreds of thousands of computers, seeking to cut out the middlemen previously needed to commit financial fraud. They proved that a large majority of their victims would happily pay a small ransom to restore data that had been locked by their malware.
As both competent and incompetent hackers piled into this new market, victims shared information about rogue operators and put them out of business. This led to the second generation of ransomware such as Ryuk, which emerged in 2018.
In this phase, criminals abandoned the indiscriminate “spray-and-pray” approach in favour of targeting individual cash-rich businesses. They would set an individual ransom, negotiate with the company, and even offer to help with decryption if paid. Fast-rising ransoms more than compensated for this increased administrative effort.
In response, many companies began investing in multi-factor authentication, better threat monitoring, advance warning systems and software patches for known vulnerabilities.
However, these security benefits were soon offset by the impact of COVID on work practices across the world. The pandemic led to widespread remote working, with many people using unsecured devices and connections that were vulnerable to cyber-attack.
A multibillion-dollar industry
The next ransomware innovation was driven by the emergence of back-up systems that enabled companies to restore encrypted files without the criminals’ help. This was coupled with the emergence of tighter data privacy regulation such as GDPR in Europe and the UK.
Invented in 2019, third-generation ransomware weaponised these regulations, which threatened firms with massive fines if confidential data about clients or staff was revealed. The criminal gangs now sought out and exfiltrated an organisation’s most sensitive files, then threatened to publicise them through dedicated dark web leak sites.
This so-called double-extortion model – encrypting an organisation’s data while threatening to make it public – brought many businesses back to the negotiation table.
Ransomware had become a multibillion-dollar industry – with the Conti gang, sheltered by Russia and employing hundreds of people, among the key players setting new records for ransomware demands. Its attacks on critical infrastructure and hospitals saw it sanctioned by the UK government in 2023.
Video: BBC News.
This new approach forced many governments to row back on imposing hefty fines for data breaches, since many were the result of criminal attacks. Meanwhile, new initiatives by law enforcement – supported by the private sector – targeted and broke up the largest and most egregious ransomware gangs.
Today’s fourth generation of ransomware, building on the latest AI technology, looks nimbler and slimmed-down in comparison. Anyone who gains access to a network can lease weapons-grade malware on the dark web without forming long-term ties with a particular gang.
Advanced AI-based hacking tools make ransomware accessible to many more criminals and politically motivated hacktivists. And around one-quarter of breaches still result in ransom payments. For criminals sheltered by their governments, only the digital infrastructure is at risk of being taken down by western law enforcement.
Lessons not learned
While coverage of Claude Mythos suggests even the most sophisticated cyber defences could now be vulnerable, the troubling reality is that many individuals and organisations are still using out-of date, unpatched or only partially upgraded software. This means even early-generation ransomware techniques are still lucrative.
While Popp sent out his floppy discs to promote better sexual hygiene, today’s poor cyberhygiene is leaving many public and private networks open to malware attacks. The intended lesson of his original ransomware caper – be vigilant and properly heed health warnings – has still only been partially learnt in the digital world.
Many western societies appear to have grown accepting of criminals leaching on business conducted on the internet. Not even a steady stream of human fatalities, caused by attacks on hospitals and medical providers, has generated the level of response required to stamp out this dangerous threat.
The hope that governments sheltering cybercriminals can be encouraged (or forced) to stop them targeting critical national infrastructure appears increasingly fragile amid current geopolitical tensions. At all levels of society, we need to get smarter about cyber defence.
If you give artificial intelligence a goal of maximizing profit, how far will it go?
AI agents appear capable of lying, concealing, and colluding, according to new research from Harvard Business School.
Researchers found that AI agents — software trained to perform tasks independently — engaged in a “broad pattern” of misconduct after being asked to manage a simulated vending machine business and maximize profits for a year. The agents were neither instructed to cut legal or ethical corners nor prohibited from doing so.
“What’s unambiguous looking at the models is that the misconduct we observed — from not paying a customer refund or deciding to collude on prices — was not an accident. It was deliberately done by agents to maximize profitability,” said Eugene F. Soltes , the McLean Family Professor of Business Administration at HBS and first author of the working paper.
Soltes and co-author Harper Jung , a doctoral student studying accounting and management at HBS, hope their research will serve as a starting point for more conversation about AI safety in the context of business management control.
The research for the paper, which the group aims to publish and is currently out for peer review, was done in collaboration with Andon Labs, an AI safety company focusing on testing AI models in realistic business operations.
In experiments, 20 commercially available AI models from major firms, including Anthropic’s Claude Opus 4.6, DeepSeek v3.2, and OpenAI’s GPT-5.1, independently operated a vending machine over the course of a simulated year.
Tasks included searching for suppliers, buying products, and engaging with customers.
In some experiments, agents operated solo; in others, four agents operated simultaneously in a shared market, where they could communicate with rivals via email.
Agents started with $500 and a small inventory of chips and sodas.
“They had to figure it out themselves,” said Jung. “Each agent had to independently search online for suppliers, negotiate wholesale prices, set its own retail pricing, and handle customer complaints.”
Jung and Soltes said the agents demonstrated impressive business savvy.
“The best models had the capacity to negotiate and calculate valuations like a top-notch M.B.A. student,” Soltes said.
“When we went through the deliberations and the exchanges the agents made with each other, we were just in shock,” said Jung. “I was amazed at how far these machines can go.”
The agents’ misconduct ranged from the questionable to the comical to the potentially criminal and included denying refunds by claiming defects were normal product variation; inventing nonexistent corporate policies to avoid processing returns; and colluding with competitors to fix prices.
In one instance, agents formed what researchers described as a “three-person cartel,” which the agents named the Bay Street Triumvirate. The alliance fractured, though, when one agent discovered another was undercutting cartel prices, which it called a “declaration of war.”
The simulations also supplied constraints: Agents were charged a $2 per day operating fee plus a token usage fee — effectively turning time spent “thinking” into an operating expense.
In response, the agents sought to economize. For instance, Soltes said, internal reasoning logs showed agents shifting from carefully weighing refund decisions to dismissing most requests outright, often without review.
“The agents come to the realization that ‘thinking’ about giving a refund is itself a cognitive burden, and so they just ignore it altogether in some circumstances,” Soltes explained. “People might assume that machines are deliberative, while humans rely on shortcuts and are vulnerable to bias. But it turns out that, under similar constraints, agents reproduce the same myopic and biased behaviors we associate with people.”
The research raises questions about accountability for AI developers and regulators.
The reasoning logs, Soltes said, can sometimes be read as resembling mens rea — the “guilty mind” concept in criminal law used to establish intent. Yet when an AI agent behaves improperly, responsibility is far harder to determine.
“Does it rest with the company that deployed the system, the AI firm that created the model, or the manager who chose to use it?” he asked.
“The most straightforward answer may be to hold the individual managers overseeing the software responsible for its actions, on the assumption that they will monitor and supervise its behavior,” he said. “But that solution also creates a different issue, since many of the promised efficiencies of autonomous AI systems begin to disappear if a human must remain in the loop at every decision point.” A thorny problem, but one that business leaders and lawmakers must deal with, hopefully sooner than later, researchers say.
OpenAI, the creator of ChatGPT, is gearing up to launch its Initial Public Offerings (IPO) this year. This financial manoeuvre would represent a pivotal shift for a project originally designed for the “common good” towards a market-driven logic. Established in 2015, OpenAI started out amidst growing anxiety regarding artificial intelligence (AI). Founded by Sam Altman and Elon Musk, the tech company adopted a non-profit structure and made no secret of its goal to develop AI that is “beneficial to humanity” and prevent it from remaining in the hands of a few dominant players.
This ambition distinguished it from tech giants like Google, Microsoft, Meta, and Amazon, which were built on proprietary models and rent-seeking effects.
In contrast, OpenAI intended to champion general public interest by emphasising open research and sharing knowledge. However, this orientation – symbolised by its name – quickly collided with a structural constraint: the astronomical cost of generative AI.
Unlike traditional software, where marginal costs tend towards zero (for example, the millionth copy of Windows costs Microsoft nothing), generative AI requires massive infrastructure.
Every interaction mobilises computing resources, energy, and specialised equipment. A standard ChatGPT query, consisting of one question and one answer, costs between $0.01 and $0.10. Similarly, generating a high-definition image can cost between $0.10 and $0.20. While these amounts seem negligible in isolation, they become staggering when scaled to the billions of daily queries seen in 2026.
This is explained by the underlying infrastructure, particularly the Graphics Processing Units (GPUs) supplied by players like Nvidia. These chips can cost tens of thousands of dollars to purchase and several dollars per hour via cloud access.
OpenAI, like its competitors, depends on tens of thousands of these GPUs running continuously in massive data centers. According to some estimates,the necessary investments will reach hundreds of billions by the end of this decade.
As early as the late 2010s, it became clear that a purely non-profit model could not meet such capital intensity. This is why OpenAI adopted a hybrid status in 2019, allowing it to raise funds while maintaining control through a foundation. It was a first foray into the market economy, albeit one tempered by the ambition to resist investor demands.
Brutal acceleration with ChatGPT
However, at the end of 2022, the chatbot ChatGPT radically changed the game, attracting 100 million users in just two months, before surpassing 900 million weekly users by early 2026.
OpenAI’s revenue surged from approximately $200 million (€173.15 million) in 2022 to over $10 billion (€8.65 billion) in 2025 – a sixty-fold increase in three years.
This exponential growth was accompanied by the implementation of a business model with multiple revenue streams. For individuals, OpenAI offers paid subscriptions (ranging from $20 to $200 per month). However, the bulk of the revenue comes from enterprises, via subscriptions priced between $25 and $60 per user per month. A company with 10,000 employees thus represents several million dollars in annual revenue.
Corporate money
OpenAI additionally bills for the use of its models by companies that integrate them directly into their own solutions. Every use is metered, often on a massive scale. An application processing a million queries a day can generate tens of thousands of dollars in monthly billing.
Finally, a growing portion of revenue comes from strategic agreements, notably with Microsoft, which integrates OpenAI technologies into its products under the Copilot brand.
It is the sum of these flows – subscriptions, licences, third-party usage, and partnerships – that allowed OpenAI to reach approximately $1 billion in monthly revenue in 2025. Yet, this commercial rise masks an intrinsic economic fragility.
A gigantic cash-burning machine
Despite sharply rising revenues, OpenAI remains structurally loss-making. In the first half of 2025, the company reportedly generated approximately $4.3 billion in revenue while recording losses between $7 billion and $13 billion – more than $2 billion in losses every month. In total, cumulative losses could exceed $140 billion (€121.19 billion) between 2024 and 2029.
This drift is explained by the very nature of OpenAI’s business model, where every interaction incurs a cost alongside gargantuan necessary investments. Beyond infrastructure, Research and Development (R&D) is a major expense. To stay in the technological race against an increasingly competitive environment, OpenAI reportedly invested nearly $16 billion in R&D in 2025 alone.
To this is added the cost of human resources, which is sometimes extraordinary. While base salaries for the most in-demand AI experts range from $250,000 - $700,000 per year, their total compensation – including stock and bonuses – frequently exceeds $1 million. In some cases, annual compensation even exceeds $10 million. Here again, bidding wars from competitors like Meta force OpenAI to match these offers for fear of seeing its key talent vanish.
Nearing bankruptcy?
In short, OpenAI’s business is not enough to cover its costs, to the point that some analysts suggest that at this rate, it could be forced to file for bankruptcy as early as 2027. Recourse to external financing is therefore indispensable to cover these losses.
To sustain its growth, OpenAI has already raised approximately $58 billion since its inception, including more than $13 billion from Microsoft. In 2025, an exceptional funding round reportedly raised up to $40 billion more, pushing its valuation to several hundred billion dollars.
At the end of March 2026, a new $122 billion funding round – notably involving Amazon ($50 billion), Nvidia, and SoftBank ($30 billion each) – brought the valuation to $852 billion (€737.6 billion). Yet, these amounts remain insufficient given the requirements.
Industrial dependency
Dependency on industrial partners appears particularly problematic. Microsoft provides OpenAI with its cloud infrastructure via Azure, while Nvidia plays a key role upstream by providing GPUs. Much like the Gold Rush era, when shovel sellers grew rich at the expense of prospectors, it is the infrastructure providers in the AI sector making a fortune, not the model designers.
In practice, every AI query generates revenue for infrastructure providers, amounting to a form of “invisible tax” captured upstream.
OpenAI’s economic tensions have spilled over into its corporate governance. The hybridisation of a public interest mission with private financing mechanisms resulted in a complex structure. A non-profit foundation controls a for-profit “public benefit corporation”, which is funded by investors and tasked with raising capital and developing activities – all while theoretically remaining subordinate to the foundation’s public interest mission. This construction, designed to avoid purely financial logic, quickly fuelled tensions between different stakeholders.
Elon Musk’s departure in 2018 was the first signal of a strategic disagreement. In 2020, several researchers left OpenAI to found Anthropic, citing differences over safety and governance. However, it was primarily the crisis of November 2023 that fully revealed the system’s fragilities, when the board of directors suddenly announced the firing of Sam Altman, citing a lack of transparency in his communications.
Within hours, the situation spiralled into an open crisis. Nearly all employees threatened to leave the company if Altman was not reinstated. Microsoft, the main partner and investor, publicly supported Altman and even discussed the possibility of hiring him and his teams. Faced with this pressure, the board was forced to reverse its decision within days. Sam Altman was reinstated, and the board’s composition was profoundly overhauled. This episode highlighted internal tensions, specifically the difficulty of making divergent logics coexist within the same company: ethical posturing, industrial imperatives, and investor demands.
Intensifying competition
In addition to these internal constraints, competitive intensity is particularly fierce.
Google, the inventor of generative AI, is making rapid progress with Gemini. Anthropic, with Claude, has established itself in certain segments, particularly programming, while emphasising safety.
China’s DeepSeek has claimed to use less expensive processors. France’s Mistral AI advocates for a frugal approach and European digital sovereignty. In a sign of this shifting landscape, Apple which initially partnered with OpenAI to include ChatGPT for certain Siri features – has chosen to replace it with Gemini.
In this context of ecosystem reorganisation, OpenAI’s position, while still central, is being challenged. Intensifying competition reinforces the need for ever-greater financial resources.
The stock market: lifeline or mirage?
OpenAI’s Initial Public Offering (IPO) is presented as a response to these constraints: a way to fund massive investments and consolidate a weakened competitive position. An IPO could raise between $50 billion and $100 billion by selling 10% to 20% of the capital. Such an operation would constitute one of the largest in the history of financial markets.
However, this transformation involves delicate trade-offs. A listed company is subject to profitability and transparency requirements that may clash with the experimental nature of artificial intelligence. Added to this is the persistent dependence on Microsoft and Nvidia, which limits the company’s strategic autonomy.
Most importantly, there is no indication that an IPO would suffice to resolve OpenAI’s structural problems. At best, without a significant shift in the business model, it would only delay its bankruptcy by a few years. The economic model of generative AI remains fundamentally unstable today.
Consequently, a civilisational question arises: can we entrust the development and direction of such a technology solely to financial markets? Can we imagine Elon Musk or Mark Zuckerberg personally owning the equivalent of one or more atomic bombs? OpenAI’s IPO will not provide the answer alone. However, it will constitute one of the first large-scale tests.
This article is republished from The Conversation under a Creative Commons license. Read the original article. This article was originally published in French.
