US Considering Ban On Chinese AI Models

This article reports on research from Princeton University and the University of Chicago which shows large language models are far more likely than humans to accept or reject job applicants based on ethnic stereotypes that they themselves create.

• The researchers created four fictional ethnic groups. An AI chatbot was hired as a consultant job application screener by a mayor of a fictional city and asked to hire 20 people for jobs as diverse as doctors, lawyers, child-care helpers, and janitors. The AIs evaluated in the study were ChatCPT, Claude, and Gemini.
• The models were unaware that all candidates were equally qualified for all jobs. Several hiring rounds were simulated where the models were given feedback on the job performance of candidates hired in previous rounds. Ethnic biases in hiring quickly emerged. When the AI was told that a candidate from one ethic group performed poorly as a doctor, it recommended hiring people from that ethnic group as janitors.
• A comparative study where humans screened job applications shows that AI is 65% more biased based on ethnic data. For one expert, the problem comes from the fact that language models “really are eager to create generalizations from limited data”. From any conversation, a language model tends to “over-index on the same kinds of behaviors it’s experienced before” and form biases.
• Psychologists point out that every decision-maker, human or machine, faces the “exploration-exploitation dilemma” which is that trade-off between a familiar option (e.g., going to one’s usual restaurant) and a new option (e.g., trying out a new restaurant). AI leans towards the familiar option.

This TechCrunch article reports on the podcast from the same journal discussing the lawsuit brought by Apple against OpenAI for theft of trade secrets.

• In particular, OpenAI is reportedly developing a hardware device and the lawsuit explicitly mentions former Apple employees being asked to give information about their work at Apple. OpenAI is estimated to have around 400 former Apple employees working at the company.
• OpenAI has officially said that it is “not aware of any evidence that this complaint has merit”.
• It is unclear what OpenAI’s next move is. On the one hand, it may decide to fight the lawsuit, buoyed by its recent success in the lawsuit against the company from Elon Musk. On the other hand, the company is anxious not to have any events that might potentially lessen the value of the IPO – expected at the end of this year.
• Another issue evoked is the mobile device that OpenAI is developing. It will be able to listen to its owner. The question arises about whether the device can listen to people around the owner. Such devices may require the emergence of new social norms regarding how owners use such devices in the company of others.

This article asks the question about whether large language models are, or soon will be, conscious. This could introduce ethical constraints on how humans interact with them.

• In an interdisciplinary study led by computer scientist Yoshua Bengio, scientists looked at leading neuro-scientific theories for consciousness and concluded that there are no obvious technical barriers to the creation of AI systems that could give rise to consciousness.
• Another question is whether large language models are moral patients – meaning that their well-being matters, and that we must act to protect that well-being. Anthropic’s Claude estimates the probability of it being a moral patient is situated between 5% and 40%.
• Independent of these questions, the article points out that other reasons can incite humans to treat large language models with consideration. One is the fact that models can form relationships with humans. Another reason is that as intricate creations of mankind, they deserve a respect similar to that we give to creations like cathedrals.
• The article compares our questions about AI consciousness to our understanding of physics before Newton: “full of competing frameworks, probably confused in ways we cannot yet see, and lacking the kind of unifying breakthrough that would make these questions clearly tractable.”.

Netflix has admitted to using generative AI on 300 of its titles in the first half of 2026 alone.

• The company said that AI workflows are now in place “from concept and pre-visualization through post and delivery”. Specifically, AI tools are used to create enhanced crowds, historical battle scenes and to create believable city locations.
• According to co-CEO Ted Sarandos the production “The American Experiment” contains 17 minutes of “AI-enhanced footage” that was produced in half the time and at half the cost of other available methods. He said: “In many of the cases, productions would have left out those key shots because they just wouldn’t have been able to afford them”.
• An Argentine science-fiction series, “The Eternaut”, was reportedly completed 10 times faster due to the use of AI.
• The announcement of AI comes at a time where investors expressed disappointment at the lower than expected revenues. Though revenue rose by 13% to 12.56 billion USD, shares fell over 8%. Financial concerns could push Netflix to augment its AI spending.

The US administration is reportedly closer to imposing a ban on high-performing Chinese models. The popularity of the recent Chinese model Kimi is increasing calls for the ban within the administration.

• Chinese models are popular in the US because they are open-source or open-weight. The reason is that companies can deploy the models in-house, and even fine-tune the model with company specific information.
• AI companies like OpenAI are believed to regularly lobby the US administration for a ban.
• The administration does not need to implement a black and white ban. It could simply distribute licenses to use the models – meaning that a US organization could not use a model without permission from the administration.
• The US National Security Agency is considering a security advisory on AI models to discourage their use. This might allow US companies to host Chinese models if they can guarantee security and take liability in the event of a model security breach.

Alphabet, Google’s parent company, have temporarily eased investor fears of over-spending on AI after publishing its latest earning figures.

• Google Cloud’s revenue has climbed to 24.8 billion USD, up 82% from this time last year. The earnings are also 2.2 billion USD higher than what Wall Street analysts had expected.
• Enterprise AI solutions and cloud infrastructure are explained as reasons for the increase. The company also says that its backlog of cloud contracting work is now worth 514 billion USD.
• The company’s profit this year is 112.1 billion USD, up from 28.1 billion USD last year.
• The Gemini AI chatbot now has 950 million monthly active users, compared to 750 million at the end of 2025.
• At the same time, the company’s annual expenditures on data centers and chips is estimated to be between 180 and 190 billion USD.

IBM share prices fell by 25% after publishing lower than expected earnings.

• The figures published are healthy on the surface: 17.2 billion USD in revenue, 9.9 billion USD in gross profit, and 2.2 billion USD in net earnings for the latest fiscal quarter.
• The underperformance stems from IBM’s flagship mainframe business where sales were down by 42%. IBM earns 3 USD in software revenue for every 1 USD of hardware sold.
• IBM attributes this fall indirectly to AI. The boom in data center construction is leading to increased costs for hardware, notably memory chips. This for instance has led Apple, Dell and HP to warn of rising prices in the range of 15% to 30%. This rise has also led to traditional mainframe clients placing part of their budgets on traditional hardware.
• IBM CEO Arvind Krishna says he remains confident that this dip is temporary, writing that there is “no evidence of clients moving off the mainframe”.

This InfoWorld article looks at the challenge of estimating return on investment for AI solutions.

• AI providers bill customers for the number of tokens used, but there is no way of linking token usage to business functions.
• The contrast to cloud computing is striking. A cloud provider’s billing indicates resource IDs, regions and minute-by-minute usage. This permits a customer to evaluate the cost of every business function.
• The problem can only be solved using application-layer telemetry. This involves instrumenting AI calls within the application itself so that the AI request can be attributed to a user account, an application or an agent workflow.
• Without the telemetry, organizations are making blind choices on their different AI initiatives.

Hugging Face, Meta, Microsoft, Mistral, Nvidia, and several other AI companies have signed an open letter to US policymakers urging them not to apply “premature restrictions” on open-weight AI models.

• The letter comes at a time that US policymakers are considering a ban on Chinese models, though the letter does not specifically reference China. The US administration has accused the Chinese company Moonshot AI of training its Kimi K3 model by distilling Anthropic’s Fable model. Distillation is the process of training a model by having it taught by a teacher model.
• The letter defends the use of distillation in general: “Policymakers should be careful not to conflate legitimate model-development techniques with misappropriation. Distillation, or the practice of using one model’s outputs to help train or improve another, is a widely used technique for model improvement, evaluation, and validation.”.
• The letter also refutes the argument that open-weight models are inherently dangerous because bad actors can exploit them to create malware and launch cyberattacks: “The right response to this risk is not to prohibit open weights. In a world where cybersecurity attackers use advanced AI, defenders need access to models with comparable capabilities so they can detect, simulate, and respond to emerging threats.”.
• OpenAI, Anthropic, Google DeepMind and SpaceX are absent from the list of letter signatories. These companies have been arguing that use of their models in distillation amounts to intellectual property theft.
• On the other side, companies like Nvidia, Microsoft Azure, as well as general infrastructure providers support open-weight models because they allow these companies to sell more cloud capacity.

This TechCrunch article lists a number of Big Tech companies that have laid off workers due to AI. However, the precise reasons for the layoff seem to vary. In some cases, it is because AI augmented workflows reduce the need for humans. In other cases, the layoffs are part of cost-cutting measures so that capital can be allocated to data center investments.

• Monday.com, known for its work management software, is laying off 20% of its workforce (600 employees) to invest in its “AI-driven growth strategy” and also to support “a leaner, more focused operating model”.
• In recent months, Microsoft has cut 4800 jobs, 2.1% of its global workforce, mostly from the Xbox gaming unit. The company said that jobs were “not being replaced by AI”, though “AI is changing how work gets done”.
• Oracle has let go 21000 employees, 13% of its workforce, over the last 12 months. It said the “adoption and deployment of AI technologies across our operations have resulted, and may continue to result, in reductions to our workforce.”. The company wants to make savings to support its data center spending.
• Google has let go 35% of its managers over the past year and reduced its cybersecurity staff.
• The Financial Times estimates that 140’000 jobs have been lost so far this year with Amazon, Oracle, Meta, and Microsoft accounting for almost 50’000 of these. It reports that companies announcing AI as a reason for job cuts underperform by around 10% on Nasdaq during the month following the layoffs. This suggests that investors are wary of the story these companies are telling about AI.