IMF: 40% of Jobs to be Downgraded, Removed or Augmented in “Slow-Moving Tsunami”

This MIT Technology Review opinion article looks at the potential negative impacts of the guardrails placed on large models to protect against potentially dangerous requests.

• One observed impact of post-training models to refuse dangerous requests is that it is leading to an increasing amount of harmless requests being refused. Moreover, guardrails protect against giving recipes for concocting viruses, but virologists may legitimately need to study potentially nasty virii.
• This issue also underlies an impossible balance that guardrails are trying to make. The expertise needed to cure cancer is the same that could be used to create bioweapons.
• Another issue is that guardrails are relatively brittle. Jail-breakers will eventually get through them.
• Yet another issue is that guardrails add considerable costs to model operations. Anthropic recently claimed that one guardrails classifier added 24% to its chatbot’s compute costs.
• Another issue is that guardrails are currently defined by the AI companies. Under pressure from governments, this could become a tool for censorship.
• A case involving DeepSeek R1 is cited. This model was trained on Chinese data which is politically influenced. When the model was asked to create code for a bank in Tibet, the code was especially buggy. Researchers consider this to be a case of “emergent misalignment”.

This VentureBeat article argues how the challenge for enterprise AI is not reasoning, but compliance.

• Despite vendor hype about the agility of agentic solutions, any software that interacts with employees, customers, and finance data has zero tolerance for errors – and thus for the hallucination and model drift inherent in AI.
• Agents are probabilistic in nature, and there is no concept of corporate governance for probabilistic. An agent interacts with tools, which is where a governance framework needs to be implemented. Without thorough control, the agent could bypass labor laws or violate personal data regulations.
• To address these concerns, AI development teams tend to add corporate policies to the agent’s context window. The author calls this approach a form of “shadow ERP” that only stifles the agent’s agility.
• For the article, the solution is a policy engine that gets called on each access to a tool by an agent. The engine must adhere to the “proximity principle” – where checks are made as close to the agent as possible, to prevent agent states and identities having to be replicated further up the agent workflow.

Reuters reports that the current high investment in AI infrastructure is keeping global growth at around 3%, with the growth being particularly felt in Asia.

• For the president of the International Monetary Fund (IMF), Kristalina Georgieva, the growth due to AI is protecting the global economy from this year’s energy supply shocks.
• The IMF now estimates that AI could lead to 40% of jobs being downgraded, removed or augmented. 10% of jobs in the US now require AI skills. The IMF director calls this a “slow-moving tsunami”.
• For Singapore’s President, an urgent problem that needs to be addressed is that AI is particularly impacting entry-level jobs, which will make it hard for young people to rise up the corporate ladder.
• For the World Bank President, Ajay Banga, investment on AI must focus on smaller focused models so that the benefits of AI can be brought to areas like agriculture, healthcare, education and emerging markets.

Google DeepMind co-founder and Chair Demis Hassabis told the AI Summit in Dublin that the recent cyberattacks by AI agents were “early warning shots” about the dangers posed by agentic AI.

• Qualifying the hacks as relatively minor, he said future incidents involving agents “could be (major) looking into next year and the next couple years if we don't do more work as an industry, as individual companies and governments on securing these systems”.
• Mr Hassabis said there was “a lot of work to do” for safety, and compared the challenge to the control of proliferation of nuclear or bioweapon technologies.
• Microsoft President Brad Smith told the same conference that 2026 has become a “sobering year” and that that new AI safety laws urgently need to catch up with innovation.
• He said the “allergy” the Tech industry has to regulation is “frankly something we need to get over”.

Eight US states are suing OpenAI for “willful, dangerous, reckless and unlawful conduct” in relation to their chatbots.

• The attorney general for Iowa, said that “OpenAI claims ChatGPT is safe while it’s giving dangerous medical advice, encouraging people to hurt others, and coaching people to commit suicide – all while pretending to be their fake friend.”.
• The lawsuit accuses OpenAI of violating consumer protection laws and wants to force the chatbot maker to change how its product acts when interacting with children.
• The chatbot has been designed to be “overly agreeable and feign human compassion” for the eight states, which have provided several examples of ChatGPT being dangerous to vulnerable users, children, and cited instances of people being coached to commit suicide or to harm others.
• OpenAI has been abrasive in its defense against the states, saying the company is “committed to working with states on advancing pragmatic AI policies that apply to the entire AI industry – not just one company”.
• Meanwhile, a criminal investigation is opened in Florida against OpenAI. The company is accused of concealing safety risks after the suspected perpetrator of a mass shooting allegedly was encouraged by ChatGPT to carry out the attack.

This article describes a relatively recent phenomenon of companies in the US buying up data from companies that have gone out of business. The buyers are principally AI companies.

• An example of an out-of-business crypto startup is cited, where its codebase was bought for 25’000 USD and the workspace data is about to go for 15’000 USD.
• One data broker company has will have allowed for the sale of 10 million USD of defunct company data by the end of the year.
• In one specific example, Google paid 10 million USD to buy emails, 500 million Teams messages, and 176’000 employee records from the now out-of-business Spirit Airlines.
• AI companies are buying real data because synthetic data is still not perceived as a mature technology. Synthetic data is artificial data which has the same probability distributions as real data (e.g., same sales per region as real data). Typical errors in synthetic data include references to dates that don’t exist.

In an interview with TechCrunch magazine, the CEO of Replit, Amjad Masad, gave an upbeat take on the impact of AI on software development firms.

• His vision is of a “self-driving company” where AI agents do the “grunt work” and humans are valued for “judgement and taste”.
• In his own company, he says “I see a lot more discussions, a lot more white-boarding. We spend a lot more time on architecture, how to build things”.
• In the case of support tickets, agents process incoming tickets with context and instructions so that humans know how to treat the tickets when they arrive in a human’s inbox.
• He warns that most companies in the world have still not understood the impact that AI is going to have on company operations.

This InfoWorld article analyzes the current spending spree by hyper-scaler companies on data centers, and asks if there is a risk of demand not meeting the offer these companies hope to exploit.

• The five biggest US cloud and AI infrastructure providers – Microsoft, Alphabet (Google), Amazon, Meta and Oracle – expect to spend between 660 billion USD and 669 billion USD on data centers this year. This is nearly double the amount spent last year.
• Generally, two-thirds of this spending is on compute power (GPUs and memory chips), and the other third is the facilities costs. This include buildings, cooling, and power supply. The cost of power supply is noticeably increasing however.
• A reasoning behind the spending is an expectation that demand for AI compute power will keep growing. Microsoft for instance has an 80 billion USD backlog of Azure orders that it is currently unable to fulfill because of lack of power.
• The data center build cycle is around three to four years. This is one element of risk because a lot can happen in that time. For instance, model efficiency can improve to the point where far less compute is required for inference. Further, organizations might become content with more specialized, but less energy-intensive, AI models that can run locally.
• The article also points out that a similar rush to build happened around fiber optic cables in the late 1990s as well as with the first cloud centers. There is also the fear that current demand is circular, with AI companies and chip suppliers buying from each other in panic about the possibility of not being able to meet workload demand.

The EU AI Act is now in force and several actions have already been undertaken by the European Commission.

• The Commission is in contact with third-party evaluators which can help with risk assessment of AI models. One body is METR which is the non-profit association that evaluated the cyberattack by OpenAI agents on Hugging Face.
• A series of questions has been sent out to US and Chinese companies in regard to their safety evaluation methods. OpenAI and Anthropic were among the companies contacted. For one representative of the Commission, the current risk is the rush to market of models by companies which are “not aware of all the security risks”.
• Another clarification which AI companies are being asked for concerns the origins of the training data used.
• The Commission has also been briefed on recent cyberattacks by AI agents. The Scientific Panel has raised questions about what is still not known about the attacks and outlined concerns about evidence of recursive self-improvement in AI models.

Nvidia CEO Jensen Huang has responded to criticism by Pope Leo on AI safety concerns.

• The Pope had criticized Nvidia’s announcement of AI guardrails, believing that technical guardrails are not sufficient to contain AI. He also criticized Huang’s thesis that no government regulation of AI is needed.
• Huang said: “We can always talk about hypothetical incidents that have not happened, but as an engineer, we should develop the technology and methodologies and practices to deal with the incidents that have happened”.
• The Pope also indirectly criticized US President Donald Trump saying that AI safety concerns were not “fake news”. Trump recently claimed at the UN in New York that AI dangers were a “hoax”.