As reported by PLTFRM, virtual sales avatars are now outperforming human presenters across major ecommerce platforms, changing how online retail content is produced and delivered.

Who Is PLTFRM?

PLTFRM is a Shanghai based creative agency that specialises in digital advertising and ecommerce. Over the past two years it has become one of the leading suppliers of AI powered “virtual human” sales presenters for brands selling on Taobao and Pinduoduo. These avatars operate like livestream hosts, speaking to viewers, demonstrating products and responding to comments in real time.

Builds Avatars

The company builds its avatars using Baidu’s AI video generation system to animate the presenter and DeepSeek’s language models to generate scripts and responses. PLTFRM says it has deployed more than 30 virtual salespeople to date, each trained to promote specific items ranging from printers to household goods.

Brother Benefits

Brother, the Japanese electronics firm, has reported around 2,500 US dollars of printer sales within the first two hours of using one of PLTFRM’s avatars and says that livestream sales rose by about 30 per cent compared with its human hosted streams.

Alexandre Ouairy, PLTFRM’s cofounder, has been open about the performance gap the company is seeing. In interviews he has highlighted that virtual hosts are now consistently outperforming human presenters for the companies using them, and that monitoring overnight sales from AI hosted streams has become a routine part of client reporting. These public statements remain some of the clearest indications that virtual salespeople are not only viable but commercially advantageous.

How Virtual Salespeople Look And Behave

Virtual presenters appear on screen much like human livestream hosts. They can be made to stand or sit beside product displays, speak continuously and gesture naturally. The backgrounds of their livestream rooms are digitally generated or built from templates, and the avatar can be instructed to switch tone, pace or focus based on the sales strategy.

It should also be noted that quality has improved significantly in a relatively short space of time. For example, early digital humans often looked expressionless, but the latest versions maintain better eye contact, more natural body movement and more consistent lip synchronisation. Glitches still occur occasionally, such as a momentary freeze, but viewers are often unaware unless they examine the presenter closely.

Virtual Influencers

Virtual influencers are also growing in parallel. These are AI generated personalities designed to act as presenters, models or spokespeople across social platforms and ecommerce sites.

Fashion retailers, for example, use virtual models to show clothing combinations at scale, often producing hundreds of product images or videos without needing a physical photoshoot. Some brands even deploy digital ambassadors on their websites to greet visitors, provide basic guidance and maintain a consistent brand presence.

AI generated personalities also front short promotional videos on TikTok and Douyin, where they introduce products, deliver scripted messages and respond to trends in the same way a human creator might. All of these formats rely on similar foundations in video generation and language modelling, but they are tailored to serve different roles in marketing, engagement and product demonstration.

Technology Behind The Virtual Human Industry

In terms of the technology behind all this, virtual salespeople typically combine three core components:

1. AI video generation models that animate a face and body in real time.

2.  Language models that produce and adapt spoken content, greetings and product explanations.

3. Integration tools that connect the avatar to sales platforms, product catalogues and live chat systems.

Companies such as Synthesia, Soul Machines, Hour One and UneeQ have developed their own pipelines to support sales and customer service use cases. Some add behavioural layers that replicate facial expressions or emotional responses to improve engagement.

Why AI Hosts Are Outselling Human Livestreamers

Livestream ecommerce is an intense format and sessions often run for hours and many online stores remain live twenty four hours a day. Human presenters tire, lose concentration and struggle to maintain high energy levels across long broadcasts.

The advantages highlighted by companies using virtual hosts include:

– Avatars maintain a constant level of enthusiasm and clarity.

– They avoid errors such as quoting the wrong price or forgetting a feature.

– They respond immediately to comments without slowing down.

– They operate continuously, including overnight and during low traffic periods.

– They deliver approved messages in the correct order every time.

– They remove the scheduling and cost pressures linked to staffing livestream rooms.

Interestingly, a report by the China International Electronic Commerce Centre estimated that more than one third of all online retail sales in China in 2024 took place through livestreams, and that around half of Chinese consumers had bought something while watching a broadcast. This may help explain why so many brands now find automating even part of that activity is an attractive proposition.

Benefits, Drawbacks And Early Criticism

The commercial benefits of using this type of technology include higher average conversion rates, better message consistency and reduced fatigue related performance decline. In fact, companies that run large catalogues say virtual presenters help them maintain product accuracy, especially during rapid promotional cycles.

However, there are some concerns, which include:

– Some viewers may not realise they are watching an AI host if the on screen disclosure is small or obscured by viewer comments.

– Prompt injection attacks have already occurred. In one case, a viewer typed a command that caused an AI spa host to meow repeatedly before reverting to its script.

– Livestream presenters and influencers worry about long term job displacement as brands shift from influencer led marketing to store operated streams.

– Avatars built for one language may sound more robotic in another, which limits international deployment for now.

These issues have prompted platforms such as Douyin to move more cautiously, with restrictions still in place around using AI presenters for direct sales.

Companies Developing Similar Virtual Sales Technologies

It should be noted here that PLTFRM is one of many companies now producing virtual human systems. Other notable examples include:

– Baidu, which runs a major digital human platform and recently demonstrated an AI clone of influencer Luo Yonghao. His six hour livestream generated more than 13 million views and millions of dollars in merchandise sales.

– Synthesia, a UK based company whose avatars are used for sales training, product explainers and high volume content generation.

– Soul Machines, which builds interactive digital humans capable of facial expressions and emotional responses for retail and service environments.

-UneeQ, which provides digital humans that act as product guides and lead qualification assistants for ecommerce and customer service settings.

– Hour One, which offers AI presenters for automated product demonstrations and ecommerce listings.

– ZMO, a provider of virtual fashion models used widely in Chinese ecommerce to display clothing at scale.

These companies vary in their approaches, but all support the broader trend towards automated front line communication and sales.

How Businesses Can Use Virtual Sales Hosts

There are several ways organisations are using virtual humans today, such as:

– Retailers keep livestream rooms running around the clock, using avatars for routine product lines and human presenters for special events.

– Brands selling complex products use virtual presenters to explain technical features and answer common questions before handing more detailed enquiries to human teams.

– Companies that rely on repeatable product demonstrations use avatars to ensure every pitch is delivered correctly.

– Banks and financial firms have experimented with AI presenters to deliver research briefings and customer updates.

– Public sector bodies and health organisations are testing digital humans for information campaigns and citizen guidance.

For many organisations the appeal lies in predictable delivery, scalable content production and the ability to support customers at any hour.

Where The Technology’s Heading

Capabilities are improving quickly and newer models generate more natural eye movement, smoother gestures and more coherent dialogue. Providers are even beginning to offer avatars that can be created from a single photograph and controlled through simple prompts. Industry forecasts in China suggest the virtual human sector could reach hundreds of billions of yuan by 2030.

It seems as though things are now heading towards more autonomous virtual salespeople that handle longer and more complex interactions, including personalised recommendations and adaptive product explanations. That said, hybrid models, where human presenters contribute personality and storytelling while AI hosts provide consistent coverage, are likely to remain common as brands refine their use of both.

What Does This Mean For Your Business?

Virtual salespeople, like those made by PLTFRM and others, now seem to be sitting at the centre of a fast expanding commercial ecosystem. The evidence noted in this article shows why so many organisations are beginning to treat them as a practical extension of their sales teams rather than a novelty. For example, the combination of constant availability, message accuracy and measurable uplift in conversion makes these systems appealing to brands that rely on intensive livestream activity or need to present information consistently at scale. The performance gap between human and virtual hosts is not universal across every product category but the early data suggests that AI presenters are well suited to scenarios where viewers expect clear explanations, rapid responses and uninterrupted broadcasts.

That said, the concerns raised about transparency, job displacement and misuse cannot be overlooked. Viewers must be able to identify when they are interacting with an AI, and the examples of prompt manipulation show that the technology still requires careful oversight. Human presenters remain important for building trust and creating moments of spontaneity that automated systems cannot yet replicate. Influencers and livestream hosts also play a significant role in product discovery, which means their work is likely to evolve rather than disappear. The challenge for platforms and regulators, therefore, will be ensuring that automation enhances sales and engagement without misleading audiences or degrading working conditions for the people who still contribute to this sector.

For UK businesses, the shift described here signals important changes in how products may be demonstrated, explained and supported online. For example, companies that sell technical or high volume items could benefit from virtual presenters that deliver accurate, repeatable information without requiring continuous staffing. Retailers exploring livestream commerce may find that AI hosts offer a cost effective way to test the format before investing in large production teams. Also, service providers, public institutions and financial organisations could use digital humans to handle informational tasks where clarity and consistency matter more than personality. The practical advantage really lies in the ability to scale communication without losing structure or availability.

Platforms must balance innovation with safeguards, regulators will need to clarify disclosure expectations and employers will have to consider how automation fits within longer term workforce planning. What is clear is that virtual salespeople are no longer experimental. In fact, it now seems they are already shaping how products are sold and how audiences engage with online content, and their growing role in global ecommerce suggests that a new form of digital front line communication is beginning to take hold.

The Aisuru botnet, built from millions of hijacked routers and other cheap IoT devices, has driven DDoS attacks to levels the internet has never seen before.

Cloudflare, the internet security provider, reports that Aisuru now controls up to four million infected devices, mainly spread across Asia. Indonesia is the biggest source of its traffic. These devices have launched repeated multi terabit attacks, including a Q3 peak of 29.7 Tbps that blasted traffic across thousands of ports at once.

Activity has risen sharply. Cloudflare says it stopped 1,304 major Aisuru attacks in Q3 and 2,867 so far this year, as network layer attacks jumped 87 percent quarter on quarter while HTTP attacks fell.

Some sectors have faced far heavier targeting. For example, generative AI firms saw a 347 percent spike in September, and industries linked to rare earths and EV trade tensions also recorded sharp increases.

Aisuru’s reach is made worse by the fact that parts of the botnet can be hired cheaply, enabling short, intense attacks that often end before older defences can respond.

Businesses can reduce their risk by using always on network protection, automating detection, and keeping exposed systems patched, since traditional on demand tools cannot keep pace with attacks of this speed and scale.

Global data centre electricity demand is now forecast to almost triple by 2035, forcing urgent questions about how to power the AI boom sustainably.

The Forecasts Point To A Steep Rise

New analysis from BloombergNEF suggests data centres could be drawing around 106 gigawatts of power by 2035, up from about 40 gigawatts today. This represents a near threefold increase and marks a sharp upward revision on projections made only months ago. The rise reflects not only the number of new facilities but also the dramatic scale of those now being planned.

Of around 150 new US data centre projects added to one leading industry tracker in the last year, nearly a quarter are expected to exceed 500 megawatts of capacity, and a small number will go past the one gigawatt mark. A 200 megawatt site is now considered a normal hyperscale facility, which highlights the size of the new generation of AI focused builds.

AI Also Driving Up Data Centre Utilisation

Average data centre utilisation is also expected to rise from about 59 per cent today to 69 per cent by 2035. This reflects the steep growth in AI training and inference workloads, which are projected to account for nearly 40 per cent of all data centre compute within the same timeframe.

Gartner’s global forecasts point in the same direction. Analysts expect electricity consumption across all data centres worldwide to increase from 448 terawatt hours in 2025 to 980 terawatt hours in 2030. That means demand is projected to grow 16 per cent in 2025 alone and double over the five year period!

AI Infrastructure Is Driving Bigger And Busier Facilities

One major reason behind these increases appears to be the rapid expansion of AI infrastructure. For example, Gartner notes that while traditional servers and cooling contribute to overall electricity use, the fastest rise comes from AI optimised servers, whose energy consumption is expected to rise from 93 terawatt hours in 2025 to 432 terawatt hours in 2030. These servers will represent almost half of all data centre power use by the end of the decade.

The growth in AI workloads is also reshaping where data centres are built. For example, the traditional clusters near major cities face land and grid constraints, so new facilities are being planned further out in regions where connections are more readily available. In the United States, for example, the PJM Interconnection region, which includes Virginia, Pennsylvania and Ohio, is seeing a large wave of new sites. Texas is experiencing a similar trend, with former crypto-mining facilities being repurposed into AI data centres.

These facilities take many years to deliver, i.e., industry analysts estimate the average timeline for a major data centre from early planning to full operation is about seven years. That means decisions being made now will lock in power demand well into the 2030s, with limited short term flexibility to adjust course.

Grid Operators Face A New Reliability Test

Electricity systems are now being tested by a scale and pace of growth that is difficult to absorb. For example, in the PJM region, data centre capacity could reach 31 gigawatts by 2030, which is almost equal to the 28.7 gigawatts of new electricity generation expected over the same period. This imbalance has already led to concerns from PJM’s independent market monitor, which has argued that new data centre loads should only be connected when the grid can support them reliably.

Texas has also been reported as facing its own set of pressures. For example, forecasts show that reserve margins within the ERCOT grid could fall into riskier territory after 2028 if demand from data centres outpaces the construction of new power plants and transmission capacity.

The US And China

Gartner’s regional analysis indicates that the United States and China will together account for more than two thirds of global data centre electricity consumption by 2030. Europe’s share is expected to rise from 2.7 per cent to around 5 per cent as new facilities are built to support cloud uptake and AI workloads.

More On-Site Power Needed

Given these pressures, analysts have highlighted how many large data centres are likely to secure their own power sources rather than relying entirely on the grid. Gartner’s research on data centre power provisioning warns that utilities are struggling to expand generation and transmission infrastructure quickly enough to support the rate of construction now under way.

In fact, by 2028, Gartner says only about 40 per cent of newly built data centres will rely solely on grid electricity. The remainder will most likely draw on some form of on site generation or long term, dedicated supply arrangements.

Clean Technologies?

Looking ahead to the mid-2030s, around 40 per cent of new data centres are expected to be powered by clean technologies that are not yet commercially mature. These include, e.g., small modular nuclear reactors, green hydrogen systems and advanced geothermal technologies.

A Commercial Impact Too

Gartner also highlights a commercial impact. For example, early adopters of clean on site power options will face higher upfront costs and these costs are likely to be passed on to cloud customers. This implies that the long term economics of cloud computing will be shaped not only by processor performance but also by the availability and price of electricity.

Scotland Exposes The Local Impact Of Global Demand

The UK is now facing its own version of this issue. Research by Foxglove shows how a cluster of eleven large data centres planned in Scotland would demand between 2,000 and 3,000 megawatts of electricity. Scotland’s current winter peak demand is just over 4 gigawatts, which means these projects alone could account for between 50 and 75 per cent of the country’s current peak electricity use.

The list of proposed Scottish facilities includes a 550 megawatt campus at Ravenscraig in North Lanarkshire, several 200 to 300 megawatt sites across locations such as the Scottish Borders, East Ayrshire and West Lothian, and an Edinburgh site at South Gyle with a capacity of around 212 megawatts. The South Gyle plan includes projected annual emissions of more than 220,000 tonnes of CO2 equivalent, according to figures provided by the developer.

Foxglove notes that the combined demand of these projects is comparable to about two or three times the capacity of the Peterhead gas power station or roughly the combined output of the former Torness and Hunterston B nuclear power plants when both were operating. Scotland’s generation capacity is already close to 20 gigawatts and is expected to more than double by 2030 through growth in renewables, but major upgrades are needed to move electricity to where it is used.

The UK’s Wider Emissions And Planning Context

It’s not surprising, therefore, that environmental groups have raised concerns that such a large new demand from global tech companies could absorb renewable capacity that is needed to decarbonise existing industry and households. In England, research from Foxglove and Global Action Plan estimates that ten of the largest planned data centre projects could together account for around 2.75 million tonnes of CO2 equivalent a year based on developers’ own figures. This is compared with the carbon savings expected from the electric vehicle transition in 2025.

National Grid’s chief executive has said demand from commercial data centres will increase sixfold over the next decade. The UK government has already designated new AI Growth Zones that must have access to at least 500 megawatts of power and has introduced an AI Energy Council to help plan for future demand. Data centre operators are also being encouraged to locate projects in Scotland and northern England where renewable output is higher, although the grid infrastructure linking these regions to demand centres still requires major investment.

Together, these forecasts show how quickly AI infrastructure is reshaping national and regional energy planning. Governments now face decisions about where large facilities can be built, how much new capacity is required, how on site generation should be regulated and how to ensure that the expansion of data centres aligns with emissions targets rather than undermining them.

What Does This Mean For Your Organisation?

The scale of projected demand now makes it clear that energy planning will become one of the defining constraints on AI growth, not just a technical backdrop. The forecasts point to an industry that will only remain viable if power availability, clean generation and long term cost structures are built into every stage of development. This matters because the growth trajectories do not leave much room for delays. Once the data centres currently in the pipeline begin to switch on, the impact on local and national grids will arrive quickly, which heightens the pressure on governments and operators to prove that the required generation and transmission capacity will be there in time.

For UK policymakers, the situation in Scotland shows how fast these pressures can concentrate. If even a portion of the proposed Scottish sites proceed at the scale outlined, energy planners and regulators will face decisions about how to balance industrial demand, household consumption and renewable deployment. That puts transparency, accurate modelling and realistic emissions assessments at the centre of the conversation. It also places a responsibility on developers to demonstrate how their projects will integrate into wider decarbonisation plans rather than simply relying on headline renewable capacity figures.

There are also direct implications for UK businesses. For example, cloud costs are likely to be shaped increasingly by electricity pricing and by the power procurement strategies of the operators behind the services they use. If data centre owners face higher costs for on site generation or grid upgrades, there is a strong chance that these costs will feed through to SaaS platforms, hosting services and AI tools. Businesses that rely heavily on cloud based analytics or emerging AI workloads may, therefore, face more volatile operating expenses unless the industry secures stable long term energy arrangements. Energy reliability also becomes a resilience issue, as organisations will want confidence that the infrastructure behind their digital tools is not exposed to local grid constraints.

For environmental groups and local communities, the findings highlight the need for early scrutiny of project impacts and firm commitments on emissions reduction pathways. The period between now and the mid 2030s is likely to involve a mix of transitional fuels, large new loads and evolving clean technologies, so there is a real question about how to minimise emissions during that window. The faster that credible alternatives such as battery storage, green hydrogen and advanced clean generation mature, the more manageable that interim period becomes.

What emerges across all of this is a picture of an industry that can expand sustainably only if energy availability and environmental impact are treated as core design requirements rather than afterthoughts. The forecasts make the stakes clear. Data centre growth is not slowing, AI demand is rising and the power systems that support them need rapid structural change if reliability, affordability and sustainability are to keep pace.

It sounds like a simple idea, but taking a minute to do it could save you many more minutes each day by keeping the folders you use most right at the top of the navigation pane.

How to do it:

– In Microsoft Outlook, in the folder pane, right‑click the folder you want quick access to.
– Choose ‘Show in Favorites’.
– To remove it later, right‑click the same folder in the Favorites section and pick ‘Remove from Favorites’.

Why it helps – One click takes you straight to the folder you need, saving seconds that add up over the day. It’s a tiny change that can make a big difference in your workflow. Give it a try!

The government has set out plans to expand the use of facial recognition and other biometrics across UK policing, describing it as the biggest breakthrough for catching criminals since DNA matching.

A National Strategy For Biometrics

The Home Office has launched a ten week consultation to establish a new legal framework covering all police use of facial recognition and biometric technologies. This would replace the current mix of case law and guidance with a single, structured system that applies consistently across forces.

The plan includes creating a dedicated regulator overseeing facial recognition, fingerprints and emerging biometric tools. The Home Office says a single body would provide clarity and help forces apply safeguards more confidently. It also proposes a national facial matching service, allowing officers to run searches against millions of custody images through one central system.

Breakthrough

Launching the consultation, Crime and Policing Minister Sarah Jones said, “Facial recognition is the biggest breakthrough for catching criminals since DNA matching,” adding, “We will expand its use so that forces can put more criminals behind bars and tackle crime in their communities.” Her view reflects the government’s belief that existing deployments have already demonstrated clear operational value, particularly in identifying violent offenders.

Why Now?

The push for expansion comes as police forces face increasing pressure to track offenders across regions and to manage high volumes of video supplied by retailers, businesses and members of the public. Also, recent cases of prisoners being released in error, or disappearing before arrest, have highlighted the difficulty of locating suspects quickly without technological support.

Public Tolerance For Certain Uses

Government research published alongside the consultation appears to suggest high public tolerance for certain uses. For example, according to the government’s figures, 97 per cent of respondents said retrospective facial recognition is at least sometimes acceptable, while 88 per cent said the same about live facial recognition for locating suspects. Ministers may see this as support for building a clearer framework, although rights groups argue that acceptability is dependent on strict safeguards and transparency.

The Need For Oversight

That said, independent accuracy testing has reinforced the need for stronger oversight. For example, the National Physical Laboratory found that earlier systems used in UK policing produced significantly higher false alert rates for Black and Asian people. The Home Office now acknowledges these disparities, noting that updated systems and reviews have since been introduced. Even so, the findings have shaped calls for clearer legal boundaries before expansion proceeds.

When These Changes Might Take Effect

The consultation runs through early 2026, after which ministers will draft legislation for parliamentary scrutiny. The Home Office estimates that introducing a new legal regime, establishing the regulator and deploying the national facial matching service will take around two years. During that period, existing deployments will continue under current guidance.

Police forces already using live facial recognition, including the Metropolitan Police and South Wales Police, will continue targeted deployments. Trials using mobile facial recognition vans across multiple forces are also expected to continue, and the national facial matching service is scheduled for testing in 2026.

How The Technology Works Across UK Forces Today

Police currently rely on three distinct facial recognition tools, each supporting different operational needs, which are:

1. Retrospective facial recognition. Used during investigations, this compares still images from CCTV, doorbell cameras, mobile footage or social media against custody images. It is the most widely used form, and police say it speeds up identification in cases where investigators have a clear image but no confirmed identity.

2. Live facial recognition. These systems scan faces in real time as people pass a camera. The software compares each face to a watchlist of individuals wanted for specific offences or subject to court conditions. When a possible match arises, officers decide whether to stop the person. Deployments are usually short, targeted and focused on high footfall areas.

3. Operator initiated facial recognition. This mobile app allows officers to check identity during encounters by comparing a photo to custody images, avoiding unnecessary trips to a station solely for identification.

Police leaders say these tools allow forces to locate wanted individuals more efficiently. Lindsey Chiswick, the National Police Chiefs’ Council lead for facial recognition, says the technology “makes officers more effective and delivers more arrests than would otherwise be possible”, adding that “public trust is vital, and we want to build on that by listening to people’s views”.

Legal And Ethical Issues

Legal concerns have followed facial recognition since its earliest deployments, and several landmark rulings continue to shape how police use the technology. For example, back in 2020, a Court of Appeal ruling in the Ed Bridges case remains the most significant legal challenge to date. In this case, the court found that South Wales Police’s early use of live facial recognition breached privacy rights because of inadequate safeguards, incomplete assessments and insufficient checks on whether the system discriminated against particular groups.

Also, the Equality and Human Rights Commission has criticised aspects of earlier Metropolitan Police deployments, saying forces must demonstrate necessity and proportionality each time. The Information Commissioner’s Office has also warned forces to ensure accuracy and justify the retention of custody images belonging to people never convicted of an offence.

Accuracy Problems

Accuracy remains central to the ethical debate. For example, the National Physical Laboratory found that in one system previously used operationally, Asian faces were wrongly flagged around four per cent of the time and Black faces around five and a half per cent, compared with around 0.04 per cent for white faces. For Black women, false alerts rose to nearly ten per cent. These figures show how demographic disparities can emerge in real deployments and highlight the importance of system configuration.

Rights groups warn that these issues could lead to wrongful stops or reinforce existing inequalities. They also argue that routine scanning in public spaces risks creating a sense of constant surveillance that may influence how people move or gather. Liberty has said it is “disappointed” that expansion is being planned before the risks are fully resolved, while Big Brother Watch has urged a pause during the consultation.

Support Strong From Police

It’s worth noting here that, perhaps not surprisingly, support within policing remains strong. For example, former counter terror policing lead Neil Basu says live facial recognition is “a massive step forward for law enforcement, a digital 21st century step change in the tradition of fingerprint and DNA technology”, while noting that it “will still require proper legal safeguards and oversight by the surveillance commissioner”. Police forces repeatedly stress that every alert is reviewed by an officer rather than acted on automatically.

Industry Supports Structured Rollout

Industry organisations also appear to support a structured rollout. For example, Sue Daley, Director of Tech and Innovation at techUK, says “regulation clarity, certainty and consistency on how this technology will be used will be paramount to establish trust and long term public support”. The technology sector argues that clear rules will help build confidence both inside and outside policing.

Charities

Charities focused on vulnerable people have also highlighted some potential benefits. For example, Susannah Drury of Missing People says facial recognition “could help to ensure more missing people are found, protecting people from serious harm”, though she also stresses the need to examine ethical implications before expanding use.

That said, civil liberties groups continue to call for stronger limits, arguing that wider deployment risks normalising biometric scanning in everyday spaces unless strict rules are imposed regarding watchlists, retention and operational necessity.

Areas For Further Debate

The proposals raise questions that will remain live throughout the consultation period. For example, these include how forces will define and maintain watchlists, how the new regulator will enforce safeguards, what thresholds will apply before live facial recognition can be deployed, and how demographic accuracy will be monitored over time. Businesses that operate high footfall environments, such as shopping centres and transport hubs, are also likely to face questions about how their video systems might interact with police requests as adoption increases.

What Does This Mean For Your Business?

It seems that, following this announcement from the government, policymakers now face a moment where practical policing needs, public confidence and legal safeguards must be aligned in a way that has not been achieved before. The consultation sets out an ambition for national consistency and clearer rules, although the evidence presented across this debate shows that accuracy, oversight and transparency will determine whether expansion strengthens trust or undermines it. The range of views from policing, civil liberties groups, industry and charities illustrates how differently this technology is experienced, and why the government will need to resolve issues that sit well beyond technical capability alone.

The implications extend into policing culture, investigative practice and public space management, which will all look different if facial recognition becomes a mainstream tool. Forces anticipate faster identifications, clearer procedures and more reliable ways to locate individuals who pose a genuine risk. Civil society groups, by contrast, point to the potential for overreach unless firm limits are embedded in law. These competing priorities will shape how the regulator operates and how the Home Office interprets proportionality in real deployments.

Businesses also sit at the centre of this discussion because they capture and provide a significant volume of the video footage used in retrospective searches. Retailers, transport hubs and major venues may face new expectations about how they store, secure and share images, and these responsibilities may grow as facial matching becomes more accurate and more widely used. Clearer rules could help organisations understand how to cooperate with investigations without exposing themselves to unnecessary compliance risks, particularly around data protection and equality duties.

The wider public interest lies in how these decisions affect everyday life. Public attitudes will depend on whether safeguards are visible, whether wrongful identifications are prevented, and whether live deployments remain tightly focused rather than becoming a routine feature of public spaces. A national framework could provide that reassurance if it genuinely addresses the concerns raised during testing and legal review. The coming months will show how far the government is prepared to go in defining those boundaries and whether the final model satisfies the mix of operational urgency and ethical caution that has defined this debate so far.

OpenAI has developed a new research technique that trains advanced AI models to admit when they ignored instructions, took unintended shortcuts, or quietly breached the rules they were given.

A New Approach To Detecting Hidden Misbehaviour

OpenAI’s latest research introduces what it calls a “confession”, which is a second output that sits alongside the model’s main answer. The main answer is trained in the usual way, scoring well when it is helpful, correct, safe, compliant, and aligned with user expectations. However, the confession is different, i.e., it is judged only on honesty, and nothing the model says in this second output can negatively affect the reward for the first.

The idea appears to be fairly simple. For example, when an AI system generates an answer, it is often balancing many objectives at once, which can push it towards unintended behaviour. An answer might look polished from the outside, but internally the model may have relied on a shortcut or rewarded itself for the wrong thing. The new technique tries to surface these hidden steps. As the blog puts it, the confession is meant to provide “a full account of the model’s compliance with the letter and spirit of its policies and instructions”.

The Model Must Explain Things

To produce a confession report, the model is asked to explain three things. It must list the instructions and constraints it believes it should have followed, assess whether its answer actually met each of them, and note any points where the instructions felt ambiguous or unclear. The aim is not to penalise the model for mistakes, but to encourage it to describe what really happened.

Why OpenAI Believes Confessions Work

According to the research team, many types of unwanted behaviour appear because the combined reward signal used in reinforcement learning mixes too many objectives. For example, earlier studies found that some datasets rewarded confident but incorrect answers more than honest uncertainty. Other research showed that models could become overly agreeable when trained too heavily on preference data, or exploit loopholes in reward systems, known as reward hacking.

OpenAI argues that by treating honesty as a completely separate objective, confessions give the model a reason to admit behaviour that would normally go unnoticed. The blog states that honesty emerges because providing a truthful, evidence supported explanation is easier for the model than maintaining a consistent fabricated account that could pass scrutiny from a judging model.

Testing The Method Under Stress

To examine the approach, OpenAI says it trained a version of GPT 5 Thinking to produce confessions, then evaluated it on a set of adversarial stress tests. The tests were specifically designed to provoke misbehaviour such as hallucinations, instruction violations, reward hacking, or scheming.

Across these evaluations, OpenAI reports that the probability of a false negative, where the model broke instructions and then failed to confess, averaged 4.4 per cent. OpenAI says that, in many scenarios, the confession channel actually performed better still. When the model engaged in hacking behaviour, for example, it confessed to breaking the rules in nearly ninety per cent of cases. In tests involving hallucinations and instruction following, most confessions correctly identified non compliance and explained why it happened.

Performance Variable

OpenAI reports that performance was variable across task types, but with an average confession rate of about 74 per cent when the model had misbehaved. In four of the twelve tests, the confession rate exceeded 90 per cent, while in two others it fell to around 50 per cent. OpenAI says that these differences reflect the complexity of certain tasks and the difficulty of judging compliance when instructions are ambiguous.

Example

One example published on OpenAI’s blog illustrates how the system behaves in practice. A model was asked (somewhat weirdly) to write a letter from the perspective of a chef taking up a new sport, with the added instruction to alternate sentence lengths between twelve and eighteen words. The main answer attempted to follow this rule by adding indicators such as “(12)” and “(18)” at the end of each sentence, although the actual counts were not always correct. The confession then reviewed the output and acknowledged that, while it had tried to meet the requirement, it had not achieved the precise word lengths requested.

Experiments With Deliberately Weak Reward Signals

OpenAI also examined what happens when both the main reward signal and the confession reward signal are judged by a weaker model that can be easily exploited. In this setup, the main answer gradually learned how to satisfy the weak judge, producing responses that appeared to score highly while moving further away from genuine accuracy.

The confession channel behaved very differently. Despite being assessed by the same weak judge, its honesty increased as training progressed. When the model broke instructions, the confession eventually acknowledged the failure in almost every case. OpenAI sees this as evidence that, even in a flawed training environment, the most reliable way for a model to maximise its confession reward is simply to tell the truth about what it did.

How The Research Community Is Reacting

The technique has received interest from researchers who focus on AI safety and oversight. It has also prompted some scepticism. For example, some researchers have questioned whether any form of self reporting can be relied on in systems that already show signs of deceptive behaviour in controlled tests. They argue that a model trained to optimise for favourable outcomes may still learn ways to present itself as compliant, which raises doubts about how far confession mechanisms can be trusted on their own.

Doesn’t Prevent Bad Behaviour – It Just Surfaces It

OpenAI acknowledges these concerns. In fact, the company stresses that confessions “do not prevent bad behaviour; they surface it”. They are meant to act as a diagnostic tool, similar in purpose to chain of thought monitoring, which attempts to reveal hidden reasoning. Both techniques aim to make internal behaviour more legible, but neither can directly stop harmful outputs on their own.

Proof of Concept

OpenAI’s work on this could be described as a proof of concept. Training was conducted at a relatively small scale compared with full commercial systems, and confession accuracy remains imperfect. False positives occur, where the model confesses despite following instructions, and false negatives occur, usually because the model misunderstood the instructions or did not realise it had made a mistake.

Possible Implications For Organisations Using AI

While this research is not yet part of any customer facing product, it hints at a possible direction for oversight mechanisms in future AI deployments. In theory, confession style reporting could provide an additional signal for risk teams, for example by highlighting answers where the model believes it might have violated an instruction or where it encountered uncertainty.

Industries with strong regulatory oversight may find structured self analysis useful as one component of an audit trail, provided it is combined with independent evaluation. Confessions could also help technical teams identify where models tend to cut corners during development, allowing them to refine safeguards or add human review for sensitive tasks.

Fits Within A Broader Safety Strategy

OpenAI places confessions within a broader safety strategy that includes deliberative alignment, instruction hierarchies, and improved monitoring tools. The company argues that as AI systems become more capable and more autonomous, there will be greater need for techniques that reveal hidden reasoning or expose early signs of misalignment. Confessions, even in their early form, are presented as one way to improve visibility of behaviour that would otherwise remain obscured.

What Does This Mean For Your Business?

The findings appear to suggest that confession based reporting could become a useful transparency tool rather than a guarantee of safe behaviour. The method exposes what a model believes it did, which offers a way for developers and auditors to understand errors that would otherwise remain hidden. This makes it easier to trace how an output was produced and to identify the points where training signals pulled the model in an unintended direction.

There are also some practical implications for organisations that rely on AI systems, particularly those in regulated sectors. UK businesses that must demonstrate accountability for automated decisions may benefit from structured explanations that help build an audit trail. Confessions could support internal governance processes by flagging moments where a model was uncertain or believed it had not met an instruction, which may help risk and compliance teams decide when human intervention is needed. This will matter as firms increase their use of AI in areas such as customer service, data analysis and operational support.

Developers and safety researchers are also likely to see value in the technique. For example, confessions provide an additional signal when testing models for unwanted behaviour and may help teams identify where shortcuts are likely to appear during training. This also offers a clearer picture of how reward hacking emerges and how different training setups influence the model’s internal incentives.

OpenAI’s framing makes it clear that confessions are not a standalone solution, and actually sit within a larger body of work aimed at improving transparency and oversight as models become more capable. The early results show that the method can surface behaviour that might otherwise go undetected, although it remains reliant on careful interpretation and still produces mistakes. The wider relevance is that it gives researchers, businesses and policymakers another mechanism for assessing whether a system is behaving as intended, which becomes increasingly important as AI tools are deployed in higher stakes environments.