Who Checks AI When It Does the Work?

1

OpenAI Cuts the Cost of Computer Use. Who Benefits?

OpenAI’s new model is designed to complete tasks across software applications at a lower cost. Its arrival could make workplace automation more accessible, even as questions remain about reliability and employment.

OpenAI has introduced GPT-6.1 Terra, a model built to interpret visual information and operate computer applications. Announced on 26 October, it can work with documents, forms, technical drawings and software interfaces, allowing developers to automate tasks that previously required repeated human input.

According to OpenAI’s announcement, Terra’s standard API prices are 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol. The company says this could make high-volume tasks, such as processing invoices or transferring information between applications, considerably cheaper.

The model combines visual reasoning with computer use. For example, an agent could interpret a technical drawing, interact with design software and complete a sequence of operations instead of merely explaining how a person should perform them.

OpenAI reports improvements on several benchmarks. On Agents’ Last Exam, Terra scored 15 percentage points higher than GPT-5.6 Terra at medium reasoning effort, at approximately one-third of the cost. The company also reports faster task completion and improvements in visual reasoning.

These results come from OpenAI’s own evaluations. Performance in real workplaces may differ, particularly when software interfaces change or tasks require judgement beyond the model’s instructions.

OpenAI says Terra has undergone safety testing for computer use, cybersecurity and resistance to malicious instructions. The company classifies its cybersecurity capabilities at the Critical threshold under its Preparedness Framework and says it applies safeguards developed for more powerful models.

Terra is available through the OpenAI API and to eligible users of ChatGPT Work and Codex.

Source: OpenAI

2

Who Checks AI in Britain’s Immigration System?

An investigation describes automated tools operating across asylum administration and border surveillance. The consequences of mistakes are particularly serious when decisions concern someone’s right to remain in a country.

Artificial intelligence is being used across the UK’s immigration system, from summarising asylum interviews to identifying cases for investigation, according to Open Rights Group (ORG), a British digital rights organisation.

Its new report, Automated Discrimination: Unpacking the Digital Hostile Environment, examines how these systems are being introduced and what remains hidden from public scrutiny. The findings were reported on 27 October, ahead of the report’s scheduled public launch at King’s College London on 28 October.

ORG identifies automated systems used to summarise asylum interviews, search immigration policy, prioritise cases and flag marriages for investigation. It also describes technologies used in border enforcement, including facial recognition, drones and surveillance infrastructure.

The organisation says its research draws on freedom of information disclosures, government documents and interviews with specialists. Its central concern is that technology is being introduced into an already restrictive immigration system without sufficient transparency about how individual tools operate.

Some of the underlying concerns predate this report. In a separate investigation published on 12 October, ORG questioned the accuracy and oversight of AI tools used in asylum casework.

The new report broadens that examination to the wider immigration and border infrastructure. The report also raises concerns about the reliance on technology supplied by US, Israeli and other foreign defence and security companies.

The Home Office maintains that trained officials, rather than AI systems, make asylum decisions. It says the tools are intended to assist caseworkers and are subject to performance monitoring and quality assurance.

That distinction is important, but it does not entirely resolve the concern. A human decision-maker may still rely on an inaccurate summary or a misleading automated assessment.

ORG is calling for stronger regulation, greater transparency and a public register of government AI systems.

Source: Open Rights Group

3

Can AI Design the Chips Powering Future AI?

Cadence has introduced specialised models for autonomous semiconductor design. The technology connects directly to a wider debate: what happens when increasingly capable AI systems help build the infrastructure for their successors?

Artificial intelligence is beginning to participate in the design of the computer chips on which future AI systems will run.

On 26 October, semiconductor design company Cadence introduced Maestro, a portfolio of specialised AI models intended to work with its existing engineering tools and autonomous design agents.

According to Cadence’s announcement, Maestro combines models trained for semiconductor engineering with electronic design automation software and reinforcement learning.

Rather than simply answering engineering questions, the system is designed to help perform complex tasks within chip and electronic-system design workflows. Its agents can work with specialised software, evaluate results and refine solutions.

Cadence reports initial improvements of up to 30% in design quality and five times greater token efficiency compared with selected general-purpose foundation models. These are company-reported results from specific workflows, not evidence that AI can independently design complete advanced processors.

The company also describes a system of continuous engineering improvement using feedback from design tools. It refers to recursive self-improvement, but in this context the term describes refinement within an engineered workflow, not an autonomous intelligence rewriting itself without restrictions.

Nevertheless, the development has wider implications.

AI increasingly assists with writing software, conducting research and optimising hardware. Agentic systems extend this process by allowing models to execute sequences of actions through specialised tools.

The combination resembles one element of a much-discussed possibility in AI research: systems contributing to the development of more capable successors.

Some researchers argue that increasingly autonomous AI could eventually accelerate technological development beyond the ability of human institutions to supervise it. Others consider predictions of uncontrolled self-improvement speculative, noting the practical constraints imposed by manufacturing, experimentation, computing resources and human expertise.

Cadence has not demonstrated that scenario. Its announcement concerns specialised engineering automation within an existing commercial platform.

Yet its technology illustrates why the debate is no longer confined to software. The machinery of AI development is itself becoming a target for AI automation.

Source: Cadence

4

CodeRabbit Uses AI to Check AI’s Code. Is That Enough?

CodeRabbit is using OpenAI’s GPT-6.1 Sol to investigate software vulnerabilities across connected repositories. The approach promises broader security reviews, but raises questions about how automated findings are independently verified.

Software security company CodeRabbit is using OpenAI’s GPT-6.1 Sol as the principal model in an agent designed to investigate vulnerabilities across complex codebases.

The partnership described on 26 October extends automated code review beyond individual changes. Rather than examining a single file in isolation, the system can analyse how modifications affect shared libraries, software dependencies and security mechanisms elsewhere in a project.

This matters because a seemingly harmless change can introduce a vulnerability in another component. Modern applications often depend on large collections of interconnected software, making such problems difficult to identify through isolated reviews.

CodeRabbit’s tools include Deep Scan, which examines relationships across a repository, and CodeRabbit Secure, which can investigate connected repositories. For selected findings, the system can attempt to verify whether a vulnerability is exploitable inside an isolated testing environment.

It can also draft a proposed correction for a human developer to review.

OpenAI says CodeRabbit has used Sol to identify security-related issues across thousands of repositories and organisations. However, the announcement does not provide an independently verified assessment of the system’s overall accuracy, false-positive rate or ability to detect previously unknown vulnerabilities.

The company also acknowledges that CodeRabbit combines Sol with other AI models and conventional security tools. The reported results therefore cannot be attributed exclusively to a single model.

A central distinction is that the reviewing agent operates separately from the agent that originally wrote the code. This may reduce some risks associated with software verifying its own output, although independence alone does not guarantee accuracy.

Human reviewers remain responsible for approving changes before deployment.

Source: OpenAI

5

BrainChip Brings AI Learning to Everyday Devices

BrainChip’s latest processor is designed to recognise and learn patterns directly on small electronic devices. Local processing could reduce energy use and data transmission, although it does not automatically guarantee privacy.

A new processor from BrainChip could allow everyday electronic devices to recognise sounds, detect changes and learn certain patterns without continuously sending information to remote servers.

The company announced its AKD1500 SPI chip on 22 October, extending its family of neuromorphic processors designed for low-power artificial intelligence.

According to BrainChip’s announcement, the processor can work alongside the microcontrollers already found in devices such as smartwatches, doorbells and factory sensors.

Microcontrollers are small computers responsible for operating electronic equipment. Adding conventional AI capabilities can require more powerful processors, greater electricity consumption and substantial redesign.

BrainChip’s approach uses a standard SPI connection to attach a specialised AI processor to an existing microcontroller. This allows the main controller to continue operating the device while the additional chip analyses sensor information.

The processor uses neuromorphic techniques inspired by aspects of biological neural processing. Rather than treating every incoming signal equally, such systems can concentrate computation on relevant events.

BrainChip says the chip can learn certain new sounds, words, users or patterns locally, without retraining a model in the cloud.

This could be useful for wearable health technologies, voice-controlled devices and equipment that must continue functioning without a reliable internet connection.

The company also points to applications in industrial monitoring, robotics and specialist medical devices.

However, the announcement does not establish that every product incorporating the processor will keep personal information private. Manufacturers still determine what information their devices collect, store or transmit.

The AKD1500 SPI is commercially available through BrainChip and distribution channels, according to the company.

Source: BrainChip