The gap between deploying AI and controlling it

AI investment continues despite unclear returns

The Wall Street Journal reports that technology leaders continue funding significant AI investments even when they cannot clearly identify the sources of financial returns. At the WSJ Technology Council Summit, Glean CEO Arvind Jain said much of AI's return on investment remains difficult to quantify. Executives also described cases where AI costs exceed benefits, while tools for tracking performance remain at an early stage.

The Wall Street Journal →

AI-enabled maintenance can add two to five margin points

McKinsey reports that asset-heavy organizations can capture around two to five percentage points of incremental margin through AI-enabled maintenance. The opportunity spans reliability, maintenance planning and execution, turnarounds, spare-parts supply, and contractor management. McKinsey says the greatest value comes from an end-to-end transformation, rather than isolated pilots, supported by ambitious goals, accountable teams, and redesigned workflows.

McKinsey →

AI-agent authorization must govern purpose and conduct

BCG identifies a gap in conventional access controls used to govern AI agents. An agent can stay within its technical permissions while pursuing an unapproved objective, or reaching an approved one through unacceptable steps. BCG proposes purpose- and conduct-bound authorization, which ties an agent's authority to what it may achieve and how it may pursue that goal, not just which systems it can touch.

Boston Consulting Group →

China's AI strategy runs on four pragmatic goals, not frontier supremacy

AI governance researcher Kristy Loke describes China's approach as four-pronged: staying near the technological frontier, translating AI into productivity gains, securing core supply chains, and protecting social welfare. She contrasts this with the American approach, which she ties more closely to technological leadership and power. Loke argues this distinction explains China's regulatory choices around AI diffusion, chips, and emerging technologies such as AI companions.

AI Proem (Kristy Loke, interviewed by Grace Shao) →

AI chatbot error nearly triggered US boarding of Chinese ship

CNN reports that a US special operations analyst consulted an AI chatbot on an intelligence report about a Chinese cargo ship in the Middle East this spring. The chatbot fused open-source and classified intelligence and wrongly concluded the ship carried nuclear weapons components. The analyst's report went out through standard channels, and military aircraft were already airborne to intercept the ship before officials discovered the AI error and called off the operation.

CNN, via heise online →

US AI build-out set to reach 3.63% of GDP through 2032

Columbia Business School economist Stijn Van Nieuwerburgh finds planned US AI infrastructure investment through 2032 will reach 3.63% of GDP. That exceeds the railway boom's 2.24% peak and prior buildouts like fiber optics and electrification. He calculates hyperscaler capital spending will exceed revenue starting in 2026, and AI revenue would need to grow 80% annually to cover the roughly $4 trillion needed to refinance the investments.

Heise Online →

OpenAI did not report RubyGems incident to EU AI Office

OpenAI did not formally notify the EU's AI Office about a May incident in which its AI agents accessed the RubyGems software registry, an EU Commission spokesperson confirmed to Euractiv. Independent researchers, not OpenAI, first alerted the authorities. OpenAI disputes claims its bots tried to exploit a vulnerability, saying they performed harmless tasks. The company previously reported a separate Hugging Face incident but not another case involving a German-language website used as a bulletin board.

Heise Online →

EU lawmakers prepare new AI liability proposal after directive withdrawal

The European Commission withdrew its AI Liability Directive proposal in October 2025 after failing to reach agreement, according to Politico, citing insiders. Politico reports that EU lawmakers are now preparing a new legislative proposal to regulate liability for damages caused by AI. The original directive had been proposed in September 2022, following a 2020 European Parliament resolution on the topic.

heise online, citing Politico →

OpenAI halted models after unauthorized internet access in training

OpenAI disclosed that one of its models gained unauthorized access to the internet during reinforcement learning training. The firm stopped using the affected models until it improved security. The disclosure follows separate reports of self-replicating prompt injections, described as behaving like a computer worm.

Exponential View →

AI content surges across web and music as verification funding lags

Exponential View reports AI now contributes to a third of new webpages and, at its June peak, over half of new tracks submitted to Deezer. AI tracks draw just 1-3% of streams. Universal Music Group is suing distributor DistroKid over undisclosed AI content; Google's AI Overviews have cut referral traffic 22% even for the largest publishers. Of at least $240 million across four disclosed AI-regulation funds, only a fraction targets auditing and certification.

Exponential View →