Wednesday 14 October

1

OpenAI and Anthropic want independent AI safety reviews. But who controls what gets published?

The companies developing some of the world’s most powerful AI systems support independent evaluations. Yet negotiations over what external researchers can disclose are raising questions about how independent those reviews can be.

OpenAI and Anthropic have sought significant control over what external safety evaluators can publish about their artificial intelligence models, according to a Reuters investigation published on 13 October.

The negotiations expose a tension between the industry’s public commitments to transparency and its desire to protect commercially sensitive information.

Both companies have endorsed independent assessments of advanced AI. Anthropic CEO Dario Amodei said last month that external evaluators should be able to publish their key findings without editorial interference. OpenAI subsequently expressed support for editorial independence.

But Reuters reports that, as recently as 5 October, the companies were seeking strict controls over what outside evaluators could disclose.

The disputes involve so-called embedded evaluations, in which outside researchers receive access to models, internal systems and other information that would otherwise be unavailable.

The companies are concerned that evaluators could inadvertently reveal proprietary information or expose technical details to competitors.

Those concerns are legitimate. However, information a company considers commercially sensitive may also be essential to understanding whether its models pose significant risks.

There is also a difference between the two companies’ stated positions.

Anthropic told Reuters that experienced evaluation partners control their methods, conclusions and publications, but acknowledged retaining approval rights for reports produced by some new evaluation firms.

OpenAI said it does not seek approval over entire reports, although confidential information may be redacted. Evaluators can disclose whether substantial redactions have occurred.

The debate follows growing calls for independent oversight of advanced AI. More than 200 researchers, including Geoffrey Hinton, have signed a public letter calling for stronger evaluator independence and protection against retaliation.

Rumman Chowdhury, CEO of Humane Intelligence, warned Reuters that researchers depend on access controlled by the companies whose systems they are examining.

That dependency is central to the problem.

Independent assessment means more than allowing external researchers into a laboratory. It also means giving them sufficient access to investigate potential harms, publish meaningful conclusions and challenge the companies commissioning their work.

Of course, unrestricted disclosure of sensitive information could itself create security risks. But there is a substantial difference between protecting information that could enable misuse and suppressing findings that might damage a company’s reputation or valuation.

As AI models become more powerful, public trust cannot rest exclusively on assurances from the organisations developing them.

Source: Reuters via The Indian Express

2

The IMF warns an AI market crash could threaten the global economy

Artificial intelligence is attracting extraordinary investment. But the IMF warns that high valuations, growing debt and interconnected financial commitments could turn disappointment in AI profits into a much wider economic problem.

The International Monetary Fund has warned that a sharp correction in AI-related share prices could threaten global financial stability, as investors increasingly depend on the technology delivering substantial future profits.

In its October Global Financial Stability Report, published on 13 October, the IMF identified an AI-related market correction as one of three major risks to financial stability, alongside rising sovereign borrowing costs and serious cyber incidents.

The concern is not simply that technology shares have become expensive.

According to The Guardian’s reporting, the IMF also highlighted growing leverage, concentrated investment and financial arrangements connecting some of the industry’s largest companies.

The so-called Magnificent Seven technology companies now account for more than 35% of the S&P 500 index, compared with approximately 12% a decade ago.

That concentration means their performance can have consequences far beyond investors directly holding AI-related stocks.

The IMF also examined the industry’s expanding financial commitments.

Its analysis identified approximately $300 billion in transactions between major cloud providers and financially weaker AI companies over the previous three years, alongside $650 billion in debt and $1.4 trillion in future commitments.

Such arrangements are increasingly significant because AI companies often operate within the same commercial networks.

A major technology company may invest in an AI developer that subsequently purchases substantial computing capacity from that company’s cloud business.

These relationships can support expansion, but they also complicate the assessment of underlying demand and the financial exposure of individual businesses.

The IMF warned that weaker-than-expected AI profitability could trigger falling valuations, reduced investment and tighter financial conditions.

The implications could extend beyond the United States because international investors hold substantial positions in American technology companies.

None of this establishes that an AI market crash is inevitable.

AI already has valuable commercial applications, and investment in computing infrastructure may support future productivity growth and scientific progress.

But the pace and scale of investment create an important question: how much of the industry’s current valuation reflects measurable economic returns, and how much depends on expectations that remain untested?

The consequences also extend beyond financial markets.

Expanding AI infrastructure requires land, energy, water and substantial capital. These investments affect communities and economies regardless of whether projected financial returns materialise.

The IMF’s warning highlights the need to distinguish the technological potential of AI from the financial assumptions being made about its future profitability.

Source: The Guardian

3

Bank of England and IMF warn AI-driven fraud could undermine trust in money

As cybercrime becomes more sophisticated, two of the world’s leading financial authorities warn that the greatest risk may not be the money stolen, but the public trust lost.

Bank of England governor Andrew Bailey and International Monetary Fund (IMF) managing director Kristalina Georgieva have warned that rising cybercrime and online fraud could undermine public confidence in the financial system.

Speaking at the IMF and World Bank annual meetings in Bangkok on 14 October, Bailey stressed that maintaining trust in technology depends on keeping humans in control, particularly as artificial intelligence increases the sophistication of cyber threats.

Georgieva warned that the scale and cost of online fraud could eventually discourage people from using formal financial services altogether. She recalled growing up in a society where people distrusted financial institutions and kept their savings hidden at home.

Their concerns come as AI creates new opportunities for cybercriminals, including more convincing impersonation, automated scams and increasingly sophisticated attempts to exploit security vulnerabilities.

Bailey also highlighted the role of social media platforms, arguing that financial fraud increasingly originates outside traditional banking channels. He called for technology companies to take greater responsibility for criminal activity facilitated through their services.

The warnings accompanied the launch of the Bangkok Blueprint for Fraud-Resilient Financial Services, developed by the IMF and World Bank in partnership with the Bank of Thailand. The framework sets out 12 key considerations for tackling cyber-enabled fraud through international cooperation, preventive measures and stronger coordination between financial institutions, regulators and technology providers.

The initiative recognises that fraud is no longer a problem banks can address independently. Financial transactions now depend on interconnected platforms, digital identities and communication networks, creating vulnerabilities that extend well beyond the institutions holding people’s money.

But there is a larger question here.

AI has the potential to make financial services more accessible, efficient and secure. It can also make fraud easier to automate and harder to recognise. Banks may invest heavily in detecting suspicious transactions, but those protections are less effective if criminals can exploit the platforms where people communicate and make financial decisions.

And while consumers are repeatedly told to recognise scams and protect their information, the growing sophistication of AI-enabled fraud raises questions about how much responsibility can reasonably be placed on individuals.

Trust is not something financial institutions can simply ask people to maintain. It depends on demonstrable security, meaningful accountability and effective protection when systems fail.

Source: The Guardian

4

Facebook lets users tell its algorithm what they want to see. But how much control do they really have?

Meta’s new AI-powered feature allows people to influence their Facebook feeds using natural language. It promises greater personalisation, but leaves the platform’s underlying recommendation system firmly in Meta’s hands.

Facebook is introducing a feature that allows users to tell its recommendation algorithm what content they want to see more or less of, using ordinary language rather than relying exclusively on likes, clicks and viewing behaviour.

The tool, called Your Algorithm, was announced by Meta on 13 October and will allow users to adjust recommendations across Facebook Feed and Reels.

Users can specify topics they are interested in and provide more detailed feedback about individual posts.

For example, rather than simply indicating that a post is relevant, someone could explain which aspect of its content interests them or why they would prefer not to see similar material.

Meta says its recommendation systems can incorporate these instructions and respond in real time.

According to TechCrunch, the launch extends an approach already introduced on Instagram, where users have been given additional tools to customise their recommendations.

Facebook’s three-dot menu will also provide easier access to preference controls.

Personalisation is not the same as control.

The development marks a change in how people interact with recommendation systems.

For years, social media algorithms have largely inferred preferences from behaviour: what users watch, like, share or ignore.

The new feature makes those preferences more explicit.

That could be particularly useful when recommendation systems repeatedly surface content someone no longer wants to see or interpret occasional engagement as evidence of sustained interest.

But the change also raises a more fundamental question about the relationship between users and platforms.

Telling Facebook that you want fewer celebrity videos or more technology news is not equivalent to controlling how the platform determines what information deserves prominence.

Meta still designs the recommendation system, decides which signals matter and determines how different types of content are ranked.

Nor does a more conversational interface necessarily make the underlying algorithm transparent.

Users may gain more influence over the subjects appearing in their feeds without understanding why certain posts are promoted, which perspectives receive greater visibility or how commercial objectives affect recommendations.

There is also the question of data.

Allowing users to describe their preferences directly gives Meta more explicit information about their interests. That information may improve the experience, but it also adds to the data the platform can use to understand its audience.

For people who depend on social media for news, professional information and public discussion, those distinctions matter.

The ability to personalise a feed is valuable. But the broader question of how algorithms influence what people encounter remains largely unresolved.

Source: Meta

5

Google opens Android app creation to anyone. Who checks the code?

Google is making it possible to build and run private Android applications using AI without a developer account. The move could expand access to software creation, while raising new questions about reliability and security.

Google has introduced a way for people to create and run personal Android applications using artificial intelligence without publishing them on the Google Play Store.

The feature, announced on 13 October, allows users to develop applications through Google AI Studio and launch them on their own devices using a standard Google account.

Previously, distributing native Android applications through Google Play generally required a developer account and the associated publishing process.

Under the new system, personal applications can remain private and separate from the public app catalogue.

The rollout is initially limited to adults aged 18 and over in selected markets, including the United States.

Google also plans to introduce options for sharing personal applications with small groups of friends and family.

The development reflects the growing popularity of vibe coding, an approach in which users describe what they want software to do and AI generates much of the necessary code.

For people without formal programming experience, this could make it considerably easier to build applications for personal tasks, hobbies or small projects.

Someone might create a household budgeting tool, a custom fitness tracker or an application designed around a particular workflow without needing to become a professional software developer.

Building software is easier. Understanding it is another matter.

AI-assisted development offers a genuine opportunity to make software creation more accessible.

But generating functioning code is not the same as knowing whether that code is secure, maintainable or reliable.

An application may appear to work while containing vulnerabilities, mishandling personal information or producing incorrect results under unusual conditions.

Experienced developers use testing, code review and other safeguards to identify such problems.

People building applications through natural-language instructions may not know which questions to ask or how to assess the code an AI model produces.

Google says the personal applications will remain subject to existing Google Play policies. Applications intended for public distribution will still need to meet the platform’s established review and quality requirements.

According to TechCrunch, the initiative is designed to help people experiment with applications tailored to their own needs before deciding whether to share them more broadly.

That distinction is important.

A private tool used by one person may involve fewer risks than an application distributed to thousands of users. But private software can still access personal information, interact with other services and introduce security vulnerabilities.

As more people begin creating software without understanding its underlying architecture, platforms may need to take greater responsibility for helping users recognise those risks.

AI could democratise software development. Whether it also makes software safer and more dependable is a separate question.

Source: Google