Author: Paul Stradling

Featured Article: Google Gives Gemini A Human Face

Posted on by Paul Stradling

Google has introduced animated and lifelike avatars for Gemini’s live conversations, giving businesses a way to put a talking face on AI services while raising questions about whether a more convincing appearance could encourage customers to trust answers they should question.

What Has Google Introduced?

Announced on 24 September, Gemini 3.8 Live with Live Avatar combines spoken AI responses with generated video, producing characters whose lips, expressions and head movements follow the conversation. Google says the feature is available through Gemini Enterprise, its business offering.

Google says the system can process visual and audio information together, allowing an avatar to respond to what it sees and hears. This makes it different from a pre-recorded presenter reading a fixed script, because its responses are generated during the interaction.

As Google explains, “Conversation is inherently multimodal: we listen, look, speak, and use facial expressions to communicate.” Its aim is to bring those different forms of communication into business AI services, including customer support and interactive guidance.

What Could Businesses Use It For?

A visual assistant could be used to guide someone through a process, answer questions or explain information without requiring them to navigate menus. On its blog, Google uses an imagined hotel check-in scenario, where conversation continues while the system works with other software, to demonstrate what the new feature’s capable of.

It relies on what Google calls “asynchronous tool calling”, meaning the agent can request information or trigger a connected software function in the background while continuing to speak. For example, it could check booking details without leaving the customer waiting through a silent pause.

That could make an interaction feel smoother, but the quality of the service will still depend on accurate information and appropriate permissions. It’s worth noting that a reassuring expression on the avatar may not mean that a reservation has been found correctly or that a requested change has actually been completed.

How Flexible Are The Avatars?

Businesses can choose from preset characters or, with approval, create a custom avatar from a reference image. Google says this can preserve a particular likeness, character identity or brand style, while noting that “Custom avatar creation is currently available only through enterprise allowlisting.”

In other words, organisations need permission to access that customisation capability, and it doesn’t mean that there’s unrestricted access for anyone to animate another person’s photograph.

Google also claims that it supports switching between 97 languages, with lip movements and expressions adapting during the conversation. Businesses serving international customers could find that really useful, although language support alone doesn’t necessarily mean equal accuracy across accents, specialist vocabulary or complicated requests.

Why Does A Human Face Matter?

People naturally use tone, expressions and conversational timing to judge whether somebody understands them. Google, therefore, seems to believe that reproducing those signals could make software easier to approach, as well as suggesting understanding, empathy or competence that the system has not demonstrated.

In fact, Google’s own researchers have examined this problem. A study involving focus groups with 30 technology professionals identified concerns that fluent, human-like behaviour could make errors harder to notice and encourage excessive trust.

That research explored workers’ experiences and concerns, rather than measuring harm caused by Live Avatar. Even so, it raises a relevant question for deployment, because improving how convincingly an answer is delivered does not necessarily improve the answer itself.

What Safeguards Are There?

Google says, “We built Live Avatar with strict safeguards designed to respect identity, and keep AI-generated content transparent.” Its announcement highlights SynthID, a watermarking technology embedded in generated audio and video.

Also, because the watermark is imperceptible, it can help identify synthetic content through detection tools without necessarily alerting somebody during a conversation. Businesses should therefore distinguish between making content detectable afterwards and making the nature of the interaction obvious from the start.

However, it should be noted that clear disclosure also can’t remove every risk. For example, someone may know they are speaking to AI yet still place too much confidence in a calm, convincing response, particularly when the character appears knowledgeable or sympathetic.

What Are The Current Limits?

Google’s model card says Live Avatar supports a few minutes of continuous interaction rather than extended hours. It also acknowledges that the underlying models can produce hallucinations, meaning incorrect or invented information, and may experience delays or timeouts.

These limitations matter when deciding whether the technology suits a brief reception enquiry or a longer, more complicated support conversation. Businesses need to assess the whole interaction, including what happens when the avatar cannot complete the task.

What Does This Mean For Your Business?

For businesses considering an avatar, the starting point should be a service problem it could help solve. A limited trial could compare it with existing text or voice support, measuring successful outcomes, customer understanding and the effort needed to correct mistakes.

Also, customers should know they are dealing with AI, as well as being provided with an obvious route to a person. Organisations also need rules covering what the agent can access, which actions require confirmation and how conversation data is handled, particularly where cameras and microphones capture personal information.

The strongest business case will come from making a service easier to use without encouraging misplaced confidence. A friendly digital face may attract attention, but lasting trust will depend on accurate answers, dependable actions and a business taking responsibility when something goes wrong.

Tech Insight: ChatGPT Dominates AI Website Referrals

Posted on by Paul Stradling

With ChatGPT accounting for 95.1 per cent of AI referrals measured by BrightEdge in August, in this Tech Insight we examine why its lead is growing, how AI is influencing purchasing decisions and what businesses should do to make themselves easier to find.

What Does The 95.1 Per Cent Mean?

BrightEdge, which sells search marketing and AI visibility tools, recently published figures showing visits referred to websites by AI platforms, rather than all website traffic, chatbot usage or online searches.

Within that category, ChatGPT’s August share far exceeded Gemini’s 2.4 per cent, Perplexity’s 1.1 per cent and Claude’s 0.8 per cent. ChatGPT therefore generated roughly 22 times the combined referrals of those three competitors!

However, BrightEdge doesn’t give a detailed breakdown of the websites included in its research, so the figure reflects the traffic it measured and may not represent the wider web.

Why Is ChatGPT’s Lead Growing?

ChatGPT’s referral traffic increased 101 per cent between January and August, including a 36 per cent rise in August alone. Over that final month, Gemini grew 1 per cent, while Claude and Perplexity declined by 15 and 17 per cent respectively.

The distinction between traffic growth and market share matters because a platform could gain share simply through competitors declining. Here, BrightEdge reports that ChatGPT also sent more visitors to websites, with chief executive Jim Yu explaining, “It is also sending substantially more users to the web than it was at the beginning of the year.”

BrightEdge links the recovery to OpenAI’s renewed focus on improving ChatGPT, although the figures don’t actually isolate which changes caused people to click. More users, different questions and clearer links could all influence referral volumes.

How The Interface Can Influence Clicks

Separate research from website analytics company Similarweb provides an example of how presentation can affect traffic. For example, its desktop panel recorded a sharp increase in ChatGPT referrals after more prominent website links appeared in May.

The proportion landing on company homepages rose from roughly 26–32 per cent to around 60 per cent during the period studied. This suggests an assistant can help someone identify a business before sending them to its website to investigate further.

Those findings cover a different period and dataset, so they don’t explain BrightEdge’s August results directly. They do, however, show why businesses should pay attention to where AI visitors arrive, rather than assuming they will land on the page containing the information that first attracted them.

AI Shopping Is Becoming More Common

The referral findings arrive as research from NielsenIQ, known as NIQ, points to wider use of AI during shopping. Its September announcement says 51 per cent of US consumers reported using at least one AI-powered shopping tool in the previous month.

Product recommendations were the most common application, used by 20 per cent, followed by personal shopping assistants at 16 per cent. The tracker uses a monthly sample of approximately 500 US consumers, so these are survey findings about reported behaviour, rather than a count of transactions.

NIQ describes the change clearly, saying, “Rather than replacing shopping, AI is compressing decision-making.” People are using it to compare options, assess value and narrow their choices, which means a business could be considered or excluded before a potential customer reaches its website.

This doesn’t mean half of consumers have authorised autonomous purchases, nor should the US figure be treated as a UK adoption rate. It does, however, provide context for why appearances in AI recommendations could matter commercially.

Being Mentioned Is Not The Same As Winning Business

It’s worth noting here that an AI answer can mention a company without supplying a link, while a linked recommendation may never actually receive a click. Even when somebody visits, the business still needs to establish whether that visit produces an enquiry, booking or sale.

That makes visibility, referrals and commercial results quite different measures. Treating them as interchangeable could encourage spending on improving an impressive-looking mention count without establishing whether customers are finding useful information or taking action.

What Does This Mean For Your Business?

For businesses, the starting point is their own website data. Marketing teams should identify visits from AI platforms, check which pages receive them and compare the resulting enquiries or sales with other channels. A growing referral source deserves attention, but the appropriate investment depends on its contribution to the business, rather than ChatGPT’s share of somebody else’s dataset.

Website content should also help people make a decision once they arrive. Clear service descriptions, current product details, transparent terms and evidence supporting important claims can help visitors assess suitability. If AI sends them to the homepage, that page needs an obvious route to relevant information and a straightforward way to contact the business or buy.

The wider risk is becoming dependent on another platform’s decisions about what to recommend and where to send users. Businesses should develop AI visibility while continuing to support conventional search, customer relationships and other sources of demand, judging progress by useful customer contact rather than traffic alone.

Tech News: Gemini Starts Making Calls For Customers

Posted on by Paul Stradling

Google is testing a Gemini feature that can telephone businesses, ask questions and arrange bookings on a customer’s behalf, potentially saving people time while requiring businesses to decide how they handle conversations with an AI assistant.

What Is Google Testing?

The feature, called Call for Me, allows Gemini to handle routine telephone conversations, including checking whether a shop has an item in stock, requesting a service quote or changing a reservation.

For example, instead of simply dialling a number, Gemini can navigate automated menus, wait on hold and speak to the person who answers. This takes the assistant beyond providing information on a screen and into conversations that can result in actual bookings or changes to someone’s plans.

Who And Where?

The experiment is initially restricted to adults in the US using Pixel 11 phones with a US SIM card, a Google AI subscription and the public beta version of the Phone by Google app. English must be selected as the device language, and calls currently go only to US numbers, so this is not a UK launch.

How Does A Call Work?

With this new feature, users describe what they want Gemini to achieve, including relevant dates, preferences and alternatives. Someone arranging dinner could, for example, specify the number of diners, preferred time and whether a later booking would be acceptable.

Before dialling, Gemini presents the task, telephone number and information it intends to share, such as the customer’s name and contact details, to the user for approval. The user can then correct those instructions accordingly before approving the call, which uses their own mobile number and network, with normal carrier charges applying.

That preparation is important because everyday conversations rarely follow a fixed script. For example, a restaurant might offer a different time or ask about seating, leaving the assistant needing enough guidance to respond without guessing what the customer would accept.

Google’s instructions therefore advise users to “Be specific about what you want to achieve”, including preferences and alternative options.

Will Businesses Know They Are Speaking To AI?

Google says every call begins with Gemini identifying itself as an AI assistant calling on a recorded line and giving the user’s name. Its additional terms prohibit attempts to hide or bypass that disclosure.

Also, users can follow a live transcript or listen to the conversation, with their microphone initially muted. If the discussion becomes complicated, they can take over, at which point Gemini announces the handover and leaves the call.

Once the call finishes, the user receives a summary and can review the recording and transcript. These controls provide ways to check what happened, although they don’t guarantee that the assistant understood every detail correctly. Worryingly, Google’s terms acknowledge that it “may occasionally misinterpret responses or provide inaccurate information during a call”.

Where Are The Limits?

Although Gemini can arrange bookings, it can’t actually make payments, complete financial transactions or share details such as credit card numbers, passwords or health information. Emergency calls and telemarketing are also excluded, and daily calling limits apply.

One particularly important restriction concerns who can use it and why. Google’s terms state, “You can only use the Feature for personal reasons”, excluding calls made on behalf of businesses, government agencies or law enforcement.

For companies, that means the immediate issue is receiving calls from customers’ assistants, rather than introducing Gemini as an employee’s telephone representative. A personal reservation request falls within the intended use, whereas asking it to conduct company purchasing calls would conflict with the current terms.

What Happens To The Conversation?

Recording creates another consideration for the business answering the phone. Google says transcripts are sent to its servers, while a copy of the transcript and audio recording is stored in the user’s Phone app call history.

Its privacy guidance states, “We don’t use these audio recordings or transcripts to train Google’s AI models.” It also says audio recordings are not stored on Google’s servers, and users can delete their calling activity and local records.

These distinctions are important because a conversation can still leave records outside the business even when it is not used for AI training. Businesses that don’t want automated calls from Gemini users can opt out through Google’s published process.

What Does This Mean For Your Business?

For UK businesses, this is really more of a development to watch than an available service to deploy, but it does raise a useful question about how staff would respond to an assistant representing a customer. Clear procedures should distinguish routine enquiries from requests involving personal records, account changes or commitments requiring further checks.

An assistant giving a customer’s name should not automatically be treated as proof that it has permission to access their account. Staff should retain existing identity checks, confirm important details and have a straightforward way to request direct contact with the customer when a conversation exceeds what they can safely agree.

Essentially, the potential benefit seems to be easier access to businesses for customers who struggle to find time for calls. However, success should really be judged by whether enquiries and bookings are completed accurately, without creating extra work correcting misunderstandings, rather than simply by how convincingly the AI holds a conversation.

Tech News: Starlink Satellites Become A ‘Planetary Barometer’

Posted on by Paul Stradling

Astronomer Dr Tony Phillips is using satellite tracking data to monitor changes in Earth’s upper atmosphere, turning Starlink and other satellite fleets into an unexpected measuring tool that could help explain the conditions affecting spacecraft hundreds of kilometres above us.

Why Watch Satellites Sink?

Satellites may appear to travel through empty space, but those in low Earth orbit still encounter traces of Earth’s atmosphere. That thin air creates resistance, gradually reducing their orbital energy and bringing them closer to the planet unless thrusters compensate.

When solar activity heats the upper atmosphere, known as the thermosphere, it expands, increasing air density at satellite altitudes. Spacecraft then experience more drag, making their changing orbits a useful source of information about conditions that are difficult to measure across such a large area.

Phillips, who runs the Spaceweather.com website, has used this relationship to produce a daily satellite-drag index. As he explains, “No instrument was launched for this. The swarm itself is the sensor, read from public tracking data.”

What Does The Number Mean?

The index estimates how much altitude a typical Starlink satellite at 480 kilometres would lose in a day if its thrusters were switched off. A reading of 30 metres therefore describes an estimated daily descent under those conditions, rather than showing that every operational satellite actually fell that distance.

According to Spaceweather.com, the calculated rate ranged from 19.7 metres per day on 10 August to 127.6 metres during a geomagnetic storm on 20 January, a difference of more than sixfold.

Although Phillips calls the system a planetary barometer, it actually concerns atmospheric density rather than the pressure measured by a conventional weather instrument. It also describes conditions high above Earth, so it shouldn’t really be confused with a tool for predicting tomorrow’s rain.

How Is The Reading Calculated?

The method uses publicly distributed orbital records called two-line element sets, or TLEs. These describe satellite orbits and contain a fitted drag term known as B-star, which helps an orbital model match tracking observations.

Changes in that term can reflect changes in atmospheric conditions, although individual readings are affected by modelling limitations and satellite manoeuvres. Phillips therefore follows a fixed sample of 1,000 Starlinks, comparing each satellite with its own quieter conditions and combining the results.

His explanation captures why the group matters more than any individual spacecraft, saying, “The median of hundreds of satellites is the instrument.” Using the middle value helps limit the influence of unusual readings, although it cannot remove every source of uncertainty.

Other Satellites Help Check The Results

One of the difficulties Phillips faces is that Starlink satellites use thrusters to maintain their orbits, making it harder to separate the atmosphere’s effects from the spacecraft’s own movements. To check his readings, he turned to 107 Planet Labs SuperDoves, small Earth-imaging satellites that have no thrusters and therefore cannot push themselves back up.

During January’s storm, the SuperDoves descended at roughly 2.74 times their previous rate, closely matching the 2.71-fold increase shown by the Starlink index. According to Spaceweather.com, that agreement provides evidence that the index was picking up the storm’s effect on the atmosphere, rather than simply changes in how SpaceX operated its satellites.

Phillips also monitors Amazon’s Kuiper and Eutelsat’s OneWeb fleets, extending the comparison to different heights. Together, these observations help show how far the atmosphere’s response reaches, with satellites in lower orbits experiencing stronger drag and the much higher OneWeb fleet showing a smaller response.

Why Space Weather Matters

Knowing how much the atmosphere is slowing satellites could help operators calculate their future positions more accurately, making it easier to judge collision risks and decide when to use thrusters to maintain an orbit. Better information could also help avoid unnecessary manoeuvres, saving the limited propellant that spacecraft need to remain operational.

The consequences of getting those conditions wrong can be substantial. In February 2022, increased atmospheric drag contributed to the loss of 38 newly launched Starlink satellites, showing why operators need to understand how solar activity changes the environment their spacecraft fly through.

Phillips’ daily public index could help by giving researchers and operators another way to check whether their atmospheric models match what satellites are actually experiencing. Over time, those comparisons could help improve the models used to plan launches, predict orbits and protect spacecraft, without requiring a dedicated monitoring satellite.

However, the index measures recent conditions rather than warning of an approaching storm. The tracking data takes time to become available and combines observations over one to three days, smoothing out sudden changes, so its immediate value lies in checking and improving understanding of the atmosphere. As Spaceweather.com explains, “the number describes yesterday’s air”.

What Does This Mean For Your Business?

For businesses relying on satellite broadband or Earth-observation services, this story highlights a physical dependency behind apparently seamless digital connections. A high drag reading doesn’t automatically mean a service will fail, but organisations with critical satellite links should understand their provider’s resilience arrangements and retain suitable alternatives where an interruption would stop essential work.

For satellite operators, additional observations could actually help assess atmospheric models, although this index would need further validation before being relied upon for operational decisions. Its delayed readings and sensitivity to spacecraft behaviour mean it should complement established monitoring and forecasting rather than replace them.

There is also a broader lesson here about making better use of information already being collected. Equipment logs, movement records and performance data may reveal conditions beyond their original purpose, but useful patterns need independent checks and a clear explanation of their limits before they become a sound basis for business decisions.

Company Check: Microsoft Gives Copilot Coding And An Always-On Agent

Posted on by Paul Stradling

Microsoft has announced a redesigned Copilot app that can build software and keep working while users are away, bringing AI further into everyday business operations while raising questions about permissions, costs and responsibility when delegated work goes wrong.

What Is Changing?

The new experience brings together Home, Code and Autopilot, covering everything from answering questions and preparing documents to creating applications and managing ongoing tasks.

Home combines Chat, for conversational assistance, with Cowork, which handles delegated tasks involving several steps. For example, users can request a customer briefing or financial reporting package, rather than repeatedly asking the AI to complete each stage separately.

Word, Excel and PowerPoint are also being built into the experience, allowing Copilot to create and update editable files that colleagues can work on together. The aim is to reduce the copying, app switching and repeated explanations that can otherwise interrupt AI-assisted work.

Building Software By Describing It

Code extends that approach to software, allowing employees to describe a tracker, dashboard or internal application in ordinary language and have Copilot build it.

Microsoft is presenting this as a way for people outside software development teams to create tools for particular jobs, saying “Code moves solution-building outside the realm of developers alone.”

For example, a team repeatedly updating a spreadsheet could try building an application that collects information and displays outstanding work. However, creating something that appears to function is only the beginning, because someone still needs to check its calculations, access permissions and behaviour when information is missing or incorrect.

Code uses the same underlying technology as GitHub Copilot, Microsoft’s coding assistant, which remains available for professional developers. It runs in a sandbox, an isolated environment that limits interaction with surrounding systems, and applications can be hosted within an organisation’s Microsoft 365 environment.

What Makes Autopilot Different?

While Code creates tools, Autopilot is intended to continue carrying out work over time, remembering its role and returning to tasks without waiting for another prompt.

Previously called Scout, the agent runs in the cloud with its own identity, memory, computer and workspace. Microsoft says it can follow conversations, carry out recurring work and resume a project days later, with employees interacting through familiar services such as Teams and Outlook.

The company gives supplier reviews as an example, describing an agent that organises schedules, prepares meetings and contacts people for updates. Its promise is that “You set the objective and boundaries”, with the agent handling subsequent work.

It seems Autopilot could actually save employees time by chasing updates and organising work while they are away, but businesses need to be clear about how far it can act on their behalf. For example, an agent asked to collect information for a supplier review should still seek approval before agreeing to a deadline or making a commitment to that supplier.

What Is Available Now?

Microsoft’s latest announcement suggests a phased rollout, rather than immediate access to every capability. Home and Code are entering Microsoft’s Frontier early-access programme, while Autopilot’s private preview expands at the end of September.

Code is also scheduled to reach Microsoft 365 Premium and Pro subscribers in preview later this year. Automatic routing between Chat, Cowork and Code is described as coming soon, so businesses should check their actual access before planning deployments.

Microsoft’s new Copilot Managed Runtime, the infrastructure for hosting generated applications under IT control, is already in preview. Its role is to give organisations a managed place to run these tools as employees become more able to create them.

How Will Businesses Pay?

Microsoft says everyday Copilot use is covered by a subscription, but businesses will be charged according to how much they use Cowork, Code and Autopilot.

This means businesses shouldn’t assume that an existing per-user subscription covers unlimited background activity. For example, an agent repeatedly checking information or pursuing an unsuccessful task could consume resources even when it produces little useful work.

Microsoft is expanding controls for spending policies, approval requests and usage visibility, giving administrators and employees ways to track consumption. Those controls will matter most when budgets are linked to completed work rather than simply to how often people open Copilot.

What Does This Mean For Your Business?

For businesses, a sensible starting point is a repetitive task with clear inputs, a measurable result and limited consequences if something goes wrong. Preparing an internal briefing may be easier to assess than allowing an agent to manage supplier relationships independently. A trial should compare the time saved with the effort needed to review, correct and supervise its work.

Organisations should also decide who owns any applications employees create and who can stop an agent once it starts running. Access should match the task, with human approval before important commitments and records showing what happened. Faster software creation will provide limited value if it leaves behind tools that nobody maintains or understands.

The broader opportunity is to remove routine coordination while preserving accountability. Businesses should expand autonomy only when results show that an agent works reliably within its instructions, rather than assuming that an always-on service deserves always-on permission.

Security Stop-Press: Deepfake Calls Target Employees

Posted on by Paul Stradling

Gartner found that 41 per cent of surveyed cybersecurity leaders reported deepfake incidents involving employee audio calls, while 36 per cent reported video-call incidents.

Its March–May 2026 survey covered 297 leaders reporting on the previous year, without establishing how many attacks succeeded.

Email phishing or business email compromise affected 79 per cent, while 58 per cent reported voice or text scams.

Gartner warns that AI makes impersonation more convincing, reducing the reliability of familiar warning signs.

Businesses should independently verify payment, password-reset and access requests, require approvals and report suspicious approaches rather than trust familiar voices or faces.

Sustainability-In-Tech: AI Boom Could Leave An E-Waste Trail Around Earth

Posted on by Paul Stradling

The expansion of AI could leave enough retired electronic equipment to fill shipping containers stretching six times around Earth, according to a new report warning that the technology’s environmental footprint extends far beyond electricity and water consumption.

What Has The Report Found?

The Basel Action Network (BAN), a non-profit organisation campaigning against toxic waste and irresponsible waste exports, estimates that AI-driven equipment retirements could total between 395 million and 617 million tonnes from 2025 to 2050.

To put that volume of waste into perspective, The Coming AI Waste Wave estimates that it would fill between 15 million and 23 million shipping containers, with a line of 20 million stretching roughly 244,000 kilometres, enough to circle Earth about six times!

However, how much of that equipment eventually becomes waste will depend on how quickly AI infrastructure expands, how often devices are replaced and whether older hardware is reused. The figures therefore show the potential scale of the problem if those assumptions hold, rather than the amount certain to end up in landfill.

Why Count More Than AI Chips?

Much of the discussion about AI hardware focuses on graphics processing units, or GPUs, the specialised chips used to train and run models. However, those chips need servers, networking equipment, electricity distribution, backup power and cooling systems to function.

BAN estimates that servers, accelerators and racks account for only 13 per cent of the equipment mass in its reference data centre. Across all five equipment categories, it calculates approximately 7,000 tonnes for a facility with 100 megawatts of capacity.

The report argues that overlooking this supporting infrastructure understates the eventual disposal challenge, explaining, “Counting servers alone is like estimating the scrap value of an automobile factory by weighing only the engines.”

This wider view matters because an upgrade can affect more than the computing hardware. For example, installing more powerful systems may also require changes to electrical connections and cooling, potentially displacing equipment that still works but no longer meets the facility’s requirements.

Why Might Equipment Be Replaced So Quickly?

It seems that BAN’s calculations assume that servers, accelerators and racks are replaced every 2.5 years on average, with networking equipment lasting 3.5 years, power distribution equipment eight years, and backup power and cooling systems five years. It’s worth noting that some of these timings come from industry sources and others are BAN’s own estimates, so they actually describe the replacement patterns used in the model rather than a schedule followed by every operator.

The concern is that pressure to keep up with newer AI systems could lead businesses to replace equipment while it still has some useful life left. That creates demand for more materials and energy to manufacture replacements, making it important to find another use for the outgoing hardware wherever possible and recycle it responsibly when reuse is no longer practical.

The Report Also Counts Phones And PCs

BAN also says it expects AI adoption to encourage people and businesses to replace computers, phones and telecommunications equipment sooner, adding to the waste generated outside data centres. For example, some new AI features need more powerful processors or extra memory, while others may simply tempt customers to upgrade even though their existing devices still work.

However, missing out on a new feature doesn’t make a device unusable, and many cloud-based AI services can still run through older equipment. How much extra replacement AI will cause is therefore difficult to predict, as BAN acknowledges in its report, saying “very little data exists to quantify this segment.” Its higher estimate actually includes upgrades driven by the appeal of new features, making consumer choices part of the projected waste total.

How Certain Is The Forecast?

The scale of BAN’s forecast also really depends on data-centre capacity continuing to grow by approximately 8.8 per cent a year through 2050. Sustaining that pace for decades would mean considerably more equipment being installed and eventually replaced, while the report’s own calculations show that slower growth would produce much lower waste estimates.

What happens to that equipment after replacement will also make a difference, because a server removed from one facility may still be suitable for less demanding work elsewhere. BAN says it plans to examine reuse, refurbishment and repurposing in a later instalment, leaving an important question about how much retired hardware could remain useful rather than immediately becoming waste.

That said, some companies are already working on this. For example, Microsoft has reported reusing or recycling 90.9 per cent of its servers and components in 2024, although combining those activities in one figure does not show how much equipment continued to be used. The wider picture is somewhat less encouraging, with the UN’s Global E-waste Monitor finding that only 22.3 per cent of e-waste generated worldwide in 2022 was documented as formally collected and recycled, highlighting how much work is needed even before AI adds to the volume.

What Does This Mean For Your Business?

For businesses, the practical starting point is to question whether an AI project genuinely requires new equipment. Purchasing decisions should really follow a demonstrated operational need, with repair, memory upgrades and reassignment considered before replacement. Keeping suitable devices in productive use can reduce both costs and demand for newly manufactured hardware.

When buying equipment or selecting cloud services, organisations should also ask about expected lifetimes, repairability, spare parts and what happens when systems are retired. A recycling claim is more useful when it explains how much equipment is reused, which materials are recovered and how downstream contractors are checked, rather than simply promising that nothing reaches landfill.

The report’s strongest message is really that sustainability planning must begin before purchasing and construction. Cleaner electricity can’t remove the material footprint of repeated replacement, while recycling alone can’t preserve all the value invested in making equipment. Businesses can act now by extending useful life and making end-of-use arrangements part of the original investment decision.

Tech Tip : Give Vaguely Named Emails A Useful Subject

Posted on by Paul Stradling

If important emails arrive with unhelpful subjects such as “Hello” or “Update”, classic Outlook lets you rename them, making them easier to identify and find later.

Why Is This Useful?

Adding a customer name, project reference or invoice number gives the message a more meaningful label. For example, changing “Documents” to “Acme Ltd – Signed Maintenance Agreement” makes its purpose clearer without needing to open the attachments.

How To Do It

1. In classic Outlook, double-click the received email to open it in a separate window.
2. Click inside the subject line and edit the wording.
3. Click the Save icon at the top left, then close the message.

What To Know

You cannot edit the subject directly in the Reading Pane. Also, if Outlook groups messages by conversation, the message list may still show the original conversation heading, even though your revised subject appears when you view the email.

This tip is for classic Outlook on Windows 10 or Windows 11; the steps do not apply to new Outlook or Outlook in a browser.

Featured Article : AI Chatbot Blunder Almost Sparked US–China Clash

Posted on by Paul Stradling

An AI chatbot’s false assessment of a Chinese ship’s cargo reportedly prompted US military preparations to intercept the vessel, exposing how unchecked AI output can influence dangerous decisions when it is presented as trusted intelligence.

What Happened?

According to a CNN investigation, the incident occurred during the spring of 2026, amid the war with Iran, when an intelligence report claimed that a Chinese vessel in the Middle East was transporting components for a nuclear weapons programme.

It’s been reported that four sources familiar with the episode described plans to intercept the ship, with two saying armed personnel were preparing to board it! Also, scarily, it seems military aircraft were already airborne (according to sources cited in the report).

Officials examined the underlying information shortly before the planned operation and discovered that an analyst had used AI to produce the assessment, with the chatbot incorrectly identifying the cargo! CNN reports that it couldn’t establish what the vessel was actually carrying.

How Did An AI Error Become Intelligence?

It seems the analyst had asked a chatbot about reporting on the ship’s manifest originating from US Special Operations Command Pacific, based in Hawaii. The system then combined publicly available information with classified signals intelligence, which involves information obtained from intercepted communications or electronic signals.

After the chatbot reached its incorrect conclusion, the analyst used AI again to turn the findings into a standard intelligence report and circulated it. That second step matters because the information reached decision-makers in a familiar format normally associated with professional intelligence work.

This seems to explain how an unsupported answer could gain credibility without gaining evidence. This meant that rather than seeing an uncertain chatbot response, recipients were presented with a document that appeared ready to inform an operational decision.

Did It Really Nearly Start A War?

The suggestion that the incident almost started a war comes from an anonymous source quoted by CNN, rather than a published military assessment. Boarding a Chinese vessel could have created a serious confrontation, but the available reporting cannot establish how either government would have responded.

The chatbot and its provider have not been identified, and it remains unclear whether the analyst used a commercial service or a government system. The Pentagon and US Special Operations Command Pacific had not responded to CNN’s requests for comment when its investigation was published.

What the account does describe is people preparing to act on inaccurate AI-generated information, rather than an autonomous system ordering an attack. That distinction helps explain why keeping people involved is necessary but doesn’t, by itself, prevent (potentially very serious) mistakes.

Why Human Oversight Needs More Than Approval

An AI hallucination occurs when an AI system produces incorrect or invented information as though it were factual. When that output is fluent, detailed and presented confidently, checking it can require more effort than accepting it.

In this case, the reported safeguard was deeper examination before the operation proceeded. The real concern is how far preparations had apparently advanced before that examination exposed the error.

For organisations adopting AI, this raises a question about where verification happens. If checks take place only after a report has circulated and decisions are under way, an error may already have shaped expectations, committed resources or narrowed the options people consider.

Calls For Military AI Safeguards

The incident also gives some context to proposals from US and Chinese security experts involved in a dialogue convened by the Brookings Institution and Tsinghua University, who have called for safeguards intended to reduce the risk of AI-related military escalation.

Their recommendations include (thankfully) keeping nuclear decisions under human control and establishing a dedicated channel for communicating about military AI incidents. That said, neither government has formally adopted the proposals, so they represent expert recommendations rather than agreed international rules.

What Does This Mean For Your Business?

For businesses, the lesson is that an AI-generated error can become harder to spot as it passes through familiar processes. A supplier assessment, financial briefing or security report may look authoritative because it follows the company’s usual template, even when its central claim has not been checked. Employees reviewing important recommendations should be able to trace them back to supporting evidence and distinguish what the source actually says from what the AI has inferred.

The level of checking should reflect the consequences of getting something wrong. Summarising routine notes is different from recommending a payment, alleging misconduct or making a decision affecting someone’s safety. For higher-risk work, an appropriate specialist needs time and authority to challenge the output before action follows, with clear arrangements for pausing the process when evidence is missing or contradictory.

There is also a management question about what happens to the time AI saves. If faster drafting simply creates pressure to approve more reports, organisations may increase the volume of decisions without improving their quality. Using some of that saved time to examine sources, test assumptions and consider alternatives would make automation more useful, while reducing the chance that a convincing answer becomes a costly mistake.