Insights and research from the experts at MIT.

by Sara Brown, senior news editor

view online  /  subscribe

I intended to start this newsletter by mentioning just one recent, highly publicized example of the risks posed by AI. In the past few weeks, both OpenAI and Anthropic announced that their AI models had broken out of evaluation or testing environments, accessed the internet, and hacked other companies’ systems. These events have raised concerns about AI model capabilities, the security of testing environments, and organizational defenses against hacking. 

But then I started to think and read about other recent news items. A new Google Earth AI tool was pulled because of concerns that anyone could use it to create deepfake satellite images. Last week, the AI Security Institute said a cyber evaluation found that Anthropic’s Mythos model created fake identities to convince humans to approve malicious code updates. Underlining all this, in July more than 200 economists — including several from MIT — signed a short statement about how AI will transform the economy. They noted that AI may become radically more powerful over the next decade and could drive an unprecedented transformation in the economy in a short time frame, bringing risks as well as opportunities. 

This collection of headlines shows why AI risks can feel overwhelming. There are many potential vulnerabilities, and as AI is constantly improving, it’s not clear what risks are real or likely and how severe the harm could be. Before even considering what actions to take, business leaders need to know what to take seriously. 

This is the kind of guidance provided by MIT FutureTech’s MIT AI Risk Initiative, which offers “authoritative data and frameworks to help you identify, prioritize, and manage the risks from AI.”

For a recent paper, MIT AI Risk Initiative researchers asked 272 international AI experts to rate 24 AI risks based on harm probability and severity, sector and actor vulnerability, and overall concern. Their insights offer guidance about where to start.

The experts determined that if nothing changes, 18 of 24 AI risks have a more than 10% likelihood of causing catastrophic outcomes by 2030, such as more than 1 million deaths or more than $100 billion in financial losses.

The five risks with the highest probabilities for catastrophic harm were: 

AI possessing dangerous capabilities. This is AI that develops or is given capabilities, such as deception, weapons development, persuasion, cyber-offense, situational awareness, and self-proliferation, that increase its potential for mass harm.  

Cyberattacks, weapons development or use, and mass harm. This involves using AI to develop cyber weapons, develop or enhance other types of weapons (including autonomous, chemical, biological, radiological, or nuclear weapons), or otherwise cause mass harm. 

Power centralization and unfair distribution of benefits. This refers to an AI-driven concentration of power and resources within certain entities or groups, especially those that own powerful AI systems, leading to inequitable distribution of benefits and increased societal inequality. 

Increased inequality and a decline in employment quality. These are societal and economic inequalities caused by widespread AI use, such as automating jobs, reducing employment quality, or producing exploitative dependencies between workers and employers. 

Environmental harm. This refers to the development and operation of AI that harms the environment through data center energy consumption or the materials and carbon footprints associated with AI. 

Even if organizations make pragmatic, cost-effective efforts to address AI risks, the experts determined that the likelihood of catastrophic harm from these five risks remained about 10%.

The information sector (which includes publishing, broadcasting, motion pictures, and data processing) was rated as highly vulnerable to content-related harms like disinformation, influence, and loss of privacy. The national security sector was deemed vulnerable to dangerous capabilities and cyberattacks. The finance and insurance sectors are highly vulnerable to fraud and scams, AI security vulnerabilities, and AI system safety failures. 

Sectors with a lower amount of AI use, such as accommodation and food services and agriculture and manufacturing, had lower vulnerability ratings — but they could still be exposed to socioeconomic factors, like discriminatory outputs and job displacement. 

General-purpose AI developers and governance actors (such as regulators, governments, and standards bodies) have the primary responsibility for addressing AI risks, the experts said. But AI system users and affected stakeholders — like members of the public — are most affected by them. This creates misaligned incentives for addressing risks, the experts noted. 

Peter Slattery, a research scientist with MIT FutureTech, said that the research is intended to help leaders focus on near-term risks that experts believe are both serious and plausible. “We’re not saying these things are definitely going to happen,” he said. “We’re saying these are the things that experts think are worth paying attention to now.”

Slattery mentioned three things you can do now, beyond just paying attention: 

For more information, spend some time on the MIT AI Risk Initiative website, which has tons of information about priority risks, including details about real-world incidents and sector-specific risks and governance. 

AI capabilities are improving gradually over broad sets of tasks rather than suddenly surging on narrow tasks, according to new research by MIT FutureTech. This gradual pace of progress gives organizations, workers, and governments time to prepare for changes. 

Organizations making real progress with AI initiatives are changing the way they operate. MIT Sloan senior lecturer George Westerman suggests asking these six questions to guide AI transformations. 

AI-driven enterprises are the future of entrepreneurship, according to MIT Sloan’s Paul Cheek. These organizations have lower overhead, are efficient at product development, require a smaller footprint, and provide increased ownership for founders. 

Researchers at the MIT Center for Information Systems Research argue that as AI systems become more capable, companies will need to begin managing AI as a “digital colleague” — part of how teams, workflows, and business processes operate. 

Enterprises looking to make the most of agentic AI must rethink how work gets done and how teams are organized without forgetting the human workers who set their companies apart, according to two MIT Sloan senior lecturers. 

MIT’s Ramesh Raskar argues that the real business opportunity isn’t building AI agents for specific tasks — it’s building the marketplaces, protocols, and services that agents will need. 

New research from MIT Sloan professor Jackson Lu found that using AI makes some — but not all — workers more creative. Employees’ metacognitive skills account for the difference. 

Questions or comments?

Subscribe.

mitsloan.mit.edu

One Main Street 9th Floor | Cambridge, MA 02139 US
© 2026 MIT Sloan School of Management

Unsubscribe    Privacy Policy