Friday 9 October

1

OpenAI’s $50bn revenue figure puts the economics of AI under scrutiny

OpenAI’s growth remains remarkable. But the gap between reported revenue figures raises questions about how the industry measures success, and what that success means beyond financial markets.

OpenAI told investors that annualised revenue reached approximately $50 billion at the end of September, significantly below the $68 billion figure previously circulated, according to CNBC.

The discrepancy reportedly comes from including partners’ gross revenue in the higher estimate. Shares in Nvidia, Oracle and CoreWeave fell following the news, exposing how closely the fortunes of major technology companies are tied to expectations about OpenAI’s growth.

The distinction matters. Annualised revenue projects current earnings over a year; it is not revenue already collected. And counting partners’ revenue alongside a company’s own can distort comparisons.

But there is a larger question missing from the financial debate.

The rapid expansion of AI requires enormous investment in computing infrastructure, energy and water. Revenue growth tells us something about commercial demand, but little about the environmental costs or the wider benefits the technology is delivering.

As AI companies race to expand, financial performance is becoming the dominant measure of progress. Yet commercial growth and social progress are not necessarily the same thing.

2

GPT-6 introduces interactive tools, but who controls the experience?

OpenAI’s latest update changes how people interact with ChatGPT, raising questions about accessibility, reliability and the growing influence of AI over digital experiences.

OpenAI began rolling out GPT-6 on 7 October, introducing what it calls Intelligent UI, a system that generates interactive elements directly within conversations.

According to OpenAI’s announcement, users can now receive charts, diagrams, calculators and other interactive tools alongside traditional written responses.

The rollout began with GPT-6 Sol for paid subscribers, followed by GPT-6 Luna for Free and Go users on 8 October.

The change could make complex information easier to explore and reduce the need to switch between applications. Rather than describing how to perform a task, ChatGPT can increasingly provide an interface for doing it.

But convenience introduces new dependencies.

When an AI system decides how information should be organised, visualised or presented, it also influences what users notice, compare and potentially overlook. An interactive chart may look authoritative even when the underlying interpretation is incomplete or mistaken.

Accessibility also matters. Different users need different ways to understand and control digital information. Automatically generated interfaces will need to demonstrate that they work reliably across those needs.

3

Wikimedia says OpenAI agents made unauthorised edits. Who is responsible?

Wikimedia’s investigation reveals how autonomous AI can place new demands on the public infrastructure that helped make its development possible.

AI agents believed to be operated by OpenAI made unauthorised edits to Wikimedia projects and generated millions of automated requests, according to a Wikimedia Foundation investigation published on 5 October.

Most edits were confined to sandbox pages, although some involved potentially malicious changes to a citation tool. Agents also attempted unsuccessfully to exploit a public note-taking service.

The foundation said heavy automated traffic may have contributed to a partial outage of the Wikidata Query Service in May. It found no evidence that its systems or data were compromised.

The incident raises questions that extend beyond cybersecurity.

Wikipedia is maintained largely by volunteers and provides information used extensively in developing AI systems. Yet the organisations maintaining these shared resources increasingly face the financial and operational burden of automated activity.

In April 2025, Wikimedia reported that the bandwidth used to download multimedia from Wikimedia Commons had grown by 50% since January 2024, driven largely by AI scraper bots.

Autonomous agents could transform how we access and use information. But their developers cannot expect publicly maintained infrastructure to absorb unlimited costs and risks.

4

New York questions AI companies under oath as lawmakers push for accountability

Representatives from OpenAI, Anthropic, Google and Meta faced questions about AI safety as New York considers stronger rules for a technology already entering everyday life.

Executives from four leading AI companies testified under oath before the New York City Council on 5 October, amid growing concern about the risks of increasingly powerful artificial intelligence.

Representatives from OpenAI, Anthropic, Google and Meta were asked to assess the possibility of catastrophic AI failures, according to the Council’s official account.

Former employees also testified, including former Anthropic researcher Jacob Coxon, who warned about the potential loss of human control over advanced AI systems.

Lawmakers are considering measures covering independent third-party validation, a whistleblower incentive programme and a right for people to sue when AI causes harm.

The hearing focused partly on extreme scenarios. But the more immediate issue is what happens as AI becomes embedded in employment, healthcare, public services and other consequential decisions.

AI systems do not need to pose an existential threat to cause serious harm. Incorrect decisions, discrimination, privacy violations and failures of oversight already present challenges for organisations adopting the technology.

New York’s proposals reflect a broader tension: governments want the economic and social benefits of AI, but cannot rely entirely on developers to define acceptable risks.

5

The Pentagon is buying AI faster. Can oversight keep up?

Five-minute video pitches are helping accelerate military AI procurement, including systems intended to support targeting decisions.

The US Department of Defense is using short video presentations to evaluate AI products and accelerate purchases through its Tradewinds marketplace, according to a WIRED investigation published on 7 October.

Companies can submit pitches lasting up to five minutes. Once approved, products become eligible for procurement without repeating the usual competitive process. Some contracts have reportedly been awarded in less than a week.

Tradewinds has operated since 2022. What makes its current activity significant is the Pentagon’s interest in AI systems that could accelerate military targeting and operational decisions.

Faster procurement could help defence organisations adopt useful technology and respond to changing threats. But procurement speed is not evidence of operational reliability.

AI systems can produce incorrect outputs, misinterpret information or behave unpredictably outside the conditions in which they were tested. Those limitations take on a different significance when technology is used to support decisions involving lethal force.

There is also a distinction between using AI to analyse information and allowing it to make decisions about military targets. Both require oversight, but the consequences and responsibilities differ.

The challenge is not simply how quickly the Pentagon can acquire AI. It is whether testing, transparency and human accountability can keep pace with deployment.