Latest News

Cao Liangliang, director at Google DeepMind and IEEE Fellow, has returned to Hong Kong after a twenty-year absence. This is a significant boost for the city’s artificial intelligence community.

As an undergraduate student in China, Cao began exploring AI and working under renowned experts in computer vision and machine learning at the University of Science and Technology of China (USTC).

After graduating from USTC, Cao met Tang Xiao’ou, who was recruiting students for his lab. Many graduates from this lab have gone on to become prominent figures in China’s AI industry.

Cao earned a master’s degree from MMLab in 2005 and worked there as a research assistant until 2006. He then moved to the United States to pursue his PhD at the University of Illinois Urbana-Champaign under Thomas S. Huang, known for his work on Chinese computer vision.

Today Cao is recognized globally for leading roles that shaped Silicon Valley’s most advanced AI systems - including Gemini and Apple Intelligence. His work has made a significant impact on machine learning and its applications in real-world scenarios.

The director considers Hong Kong as home, having announced on his personal website: ‘This is a full-circle moment for me.’ Now he brings the expertise gained from Google DeepMind and other Silicon Valley companies back to his hometown. Cao’s experience with AI assistants will be valuable to local institutions that are looking to integrate these technologies into their academic programs.

Cao’s return could be just what Hong Kong’s academic ecosystem needs right now, especially PolyU where he has taken up a new role. His presence is expected to have a positive impact on the local industry and help foster collaboration between researchers and students.

Nomagic, an innovative company based in Warsaw, Poland, and Sandy Springs, Georgia, has made significant strides in the field of embodied artificial intelligence. The company’s approach to developing AI systems for robots is distinct from that of its competitors, focusing on creating models that excel at specific tasks right out of the box rather than general-purpose ‘robot brains.’ This strategy aims to eventually build towards a more comprehensive system by mastering individual tasks first.

The concept of embodied AI has gained traction in Silicon Valley, with investors betting big on systems that can interact with the physical world through robotic devices. Many startups are working on developing general-purpose AI models for robots, which would enable them to perform various tasks without extensive programming. However, these models often fall short of human-level accuracy and require significant task-specific training before achieving reliable results.

Nomagic’s approach is centered around creating highly accurate AI robot brains that can tackle specific challenges with ease. To pursue this goal, the company established an AI research lab led by Markus Wulfmeier, a former Google DeepMind robotics researcher who now serves as Nomagic’s chief scientist. This move marked a significant step towards developing more effective and efficient AI systems for robots.

Nomagic has recently announced that it has successfully deployed its first vision-language-action (VLA) model to paying customers. The VLA system is designed to perceive objects in the world, understand text-based instructions from humans, and take actions accordingly. This achievement makes Nomagic one of the pioneers in running VLAs in a live production environment rather than lab experiments or staged demos.

The early results of Nomagic’s VLA deployment are promising, with the company reporting that it has roughly halved the rate of robot-caused interventions in warehouse operations. The system is being used by Brack.Alltron, Switzerland’s second-largest e-commerce platform, which relies on robots from Nomagic to automate order picking and packing tasks.

Roland Brack, founder and owner of Brack.Alltron, expressed his enthusiasm for Nomagic’s VLA systems, stating that they have marked a significant step forward in the company’s automation efforts. According to Brack, the addition of these intelligent systems has enabled them to run autonomous shifts through nights and Sundays without increasing pressure on their human workforce.

Despite its success, Nomagic acknowledges that its VLA system is not yet perfect, with an accuracy rate below 99.9% in specific tasks. However, the company has developed a system around the VLA by integrating it with older ‘classical’ robotics software that acts as a safety net and error catcher. This approach ensures that the entire system can be trusted to operate reliably in customer warehouses.

Kacper Nowicki, Nomagic’s co-founder and CEO, explained that achieving high reliability is crucial for companies operating in the physical world. He emphasized that 99.9% accuracy is not just a marketing number but rather the minimum required to gain entry into most facilities. To address this challenge, Nomagic has developed a harness system that complements its VLA model, allowing it to operate safely and efficiently from day one.

Nomagic’s approach stands in contrast to many of its competitors, which focus on developing general-purpose AI models for robots. Markus Wulfmeier, chief scientist at Nomagic, highlighted the limitations of this strategy, stating that most companies are racing to build highly versatile robot brains without considering the practical challenges of mastering specific tasks.

Wulfmeier pointed out that the physical world is dominated by a long tail of rare situations, making it difficult for AI models to achieve high accuracy. He emphasized that relying on simulation or human teleoperation alone cannot economically close the remaining gap to the level required in real-world settings. Instead, Nomagic trains its VLA model using data from its existing fleet of robots deployed with customers.

Nomagic’s unique advantage lies in its ability to gather vast amounts of real-world data from its operational robots. This data is used to train and refine the VLA model, which Wulfmeier describes as unusually rich and diverse. By leveraging this approach, Nomagic aims to develop more effective AI systems for robots that can automate tasks with greater accuracy.

Tristan d’Orgeval, co-founder and chief strategy officer at Nomagic, stressed the importance of deploying robots in real-world settings first rather than relying on lab experiments or simulations. He noted that this approach allows companies like Nomagic to develop capable AI systems that emerge from practical experience rather than theoretical models.

The company’s focus on automation has earned it recognition within the industry. Recently, Nomagic won the 2026 International Intralogistics and Forklift Truck of the Year (IFOY) Award for its Shoebox Picker device. This achievement highlights the potential of Nomagic’s AI systems to tackle complex challenges in warehouse automation.

Nomagic’s approach marks a significant departure from traditional methods used by many companies working on embodied AI. By focusing on developing highly accurate models that excel at specific tasks, Nomagic aims to create more effective and efficient AI systems for robots.

META has launched its first in-house image model, called Muse Image. The new tool is part of Meta’s Superintelligence Labs and can generate images from scratch or edit existing photos. It also allows users to blend multiple references into one shot and pull real-time context from the web.

The model is available now within the Meta AI chatbot and powers over 30 new AI effects in Instagram Stories, enabling image generation inside WhatsApp chats in select countries. Facebook and Messenger will follow suit soon.

Muse Image pairs with another tool called Muse Spark to plan a layout before drawing an image. It can also call search and coding tools to render legible text, working QR codes, and detailed infographics. Basic creation is free, but extra capacity comes with Meta’s subscription plans. Advertisers will gain access through Advantage+ in the coming weeks.

The model works differently from traditional one-shot generators. Instead of producing an image immediately, it reasons through a request first, planning the composition and reviewing its own output before showing it to users. This allows for more control over the final result.

One of the key features of Muse Image is its ability to personalize creations based on user input. A presets panel offers one-tap ideas, from restoring old family photos to reimagining users as claymation characters or 16-bit game heroes. Users can also @-mention Instagram accounts to bring public photos into a creation.

The tool includes a markup feature that lets users sketch or circle edits directly on the image. Meta AI keeps the full conversation in memory, allowing users to refine their creations without starting over. For example, snapping a photo of a room and redesigning it using real products from the web or Facebook Marketplace is now possible with Muse Image.

Muse Image arrives as Meta tries to close the gap left by Google and OpenAI’s early moves into consumer image tools. According to internal benchmarks, Muse Image trails behind OpenAI’s latest model but outperforms Google’s Nano Banana 2 on editing tasks. This marks a significant step for Meta in developing its own AI capabilities.

The stakes are commercial as much as technical, since digital advertising remains Meta’s largest revenue source and faster, cheaper creative production feeds that engine directly. The company framed the launch within a broader push toward personal superintelligence and detailed the agentic design and tool use in its technical breakdown of the two new models.

Meta also previewed Muse Video, a clip generator built on the same foundation as Muse Image. This tool includes native audio support and is set for public release soon. The rollout follows Meta’s earlier move into AI video feeds with its Vibes stream of AI-generated clips, and it lands as generative tools expand from flat pictures toward editable 3D assets.

For everyday users, the appeal lies in the same one that drives prompt-based image generators aimed at non-designers – now folded into apps billions of people already use daily. Startups such as venture-backed challengers chasing brand-ready output show how fierce this market has become.

A recent incident at Idaho State University highlights the limitations of AI detection tools. Chemistry undergraduate Lauren Jager had written her PhD application personal statement herself, but every tool she used to check for authenticity flagged it as almost 100 percent machine-generated. To pass the checks, she deliberately made her writing less polished and submitted a statement that she considered inferior to her original. She was later accepted into a PhD program at the University of Utah.

The experience reflects a growing crisis in academic integrity technology. Universities are deploying AI detection tools to police student submissions, but researchers have found these instruments to be unreliable, biased, and easily circumvented. A 2025 study on GPTZero, widely considered one of the most-used detectors, revealed a false-positive rate of around 16 percent on human-written essays.

Another study from 2023 showed that most AI detection tools performed inconsistently on human text and struggled more with output from advanced models like GPT-4 than older systems. Notably, even the US Declaration of Independence has been repeatedly flagged as between 95 and 100 percent machine-generated by these detectors.

Some experts argue that even reasonably accurate detectors should not be used in high-stakes decisions due to the risk of false positives. Mike Perkins, who researches AI’s impact on academia at British University Vietnam, warned that using such tools could lead to unfair outcomes for students. Marzena Karpinska from Simon Fraser University cautioned that while detectors can identify broad trends across large datasets, they cannot reliably determine individual authorship.

A particular concern is bias: a Stanford University study found that AI detectors incorrectly labelled more than half of essays written by non-native English speakers as machine-generated, with an average false-positive rate of 61.3 percent.

The proliferation of large language models (LLMs) has made it easier for malicious actors to spread spam across the internet. This has led to a significant increase in bot content and spam, which can be overwhelming for users who spend time online.

In response, Reddit has developed tools that utilize LLMs to combat this issue. The irony is not lost on the platform: by using AI-powered technology to fight against AI-generated spam, they’re essentially fighting fire with fire.

According to Reddit’s own statistics, their updated tools have proven effective in reducing spam rates. They claim to block 23 million spam views per day and catch around 25,000 new spam posts and comments daily.

The company notes that these LLM-powered tools are able to detect subtle patterns of fake behavior and artificial hype that older systems often missed. This has led to a significant reduction in users’ exposure to spam – by about 20% from January to March compared with the prior three months.

Reddit’s approach is not unique, as other social platforms like YouTube, Meta, and Instagram have also implemented automated tools to reduce spam. However, Reddit’s updated system appears to be catching spam at a higher rate than its predecessors.

The use of AI-generated content has become increasingly common on these platforms, with some allowing users to post such content provided they disclose it as artificial. TikTok takes this a step further by giving users the option to toggle how much AI-generated content they want to see.

While detecting and flagging AI-generated content can be beneficial in reducing spam and hate speech, experts emphasize that human moderation is still crucial for effective results.

The recent Automate show in Chicago has left a lasting impression on the industry. The event was marked by high energy levels, dense crowds, and an unmistakable sense of urgency among exhibitors. It’s clear that the automation landscape is undergoing significant changes, with many feeling compelled to adapt or risk being left behind.

One of the key takeaways from the show is that various factors are now converging to drive a US manufacturing renaissance. Policy pressure, onshoring, manufacturing investment, labor constraints, and usable AI with improved reliability are all reinforcing each other, rather than moving in separate directions. This convergence has created an environment where innovation can thrive.

The Monday keynote speakers at Automate 2026 provided valuable insights into the changing nature of manufacturing. Andre Marino from Schneider emphasized that the US cannot compete with major manufacturing centers like China solely on capacity or labor costs. To achieve a ‘manufacturing renaissance’ in the US, it must focus on efficiency, connectivity, and innovation. However, this requires more than just investing in new factories; it demands better factories.

Mike Cicco added to the argument by highlighting that the barrier to automation is decreasing as AI makes systems easier to deploy and use. When combined with North American onshoring pressure, everything starts to fall into place. The forces driving change are no longer theoretical but practical, visible, and increasingly urgent.

The old model of industrial automation is losing its grip. Matt Moschner described why this shift feels so significant: the complex systems of the past were only as good as the day they were deployed. They worked beautifully until the environment changed, the product changed, or the labor model changed – then there was an expensive rip-and-replace standing in the way of competitiveness.

Now, the promise is different. Systems are becoming simpler and more robust, capable of learning and improving over time. AI makes incremental training possible, allowing models to be refined without turning every change into a massive reintegration exercise. This doesn’t mean industrial automation has become easy; it means the old tradeoff between capability and flexibility is starting to break down.

The industry needs to stop treating AI as just another software feature. It’s increasingly becoming a way to make automation less static, with the real shift being that machines are becoming more adaptable in practice rather than simply ‘intelligent’ in theory.

Software-defined manufacturing was a recurring theme throughout the event. Wendy Tan framed it not as whether AI matters but how to make it useful in production. For too long, the industry struggled to move from AI discussion to industrial impact. Her argument is that the time has finally come for software to leverage existing assets and investments, making hardware do things it couldn’t before.

Wendy also made another crucial point: the clunky user experience of traditional industrial automation systems needs to change. If industrial automation is going to scale broadly, it cannot remain an expert-only craft built on obscure workflows and heroic engineering effort. The next phase of adoption depends on simple workflows, ease of use, and a more standardized, software-driven path from intent to execution.

The Automate 2026 panel described a clear sense of urgency driven by tariffs, trade disruption, workforce shortages, and general economic instability. Their response? To become more agile and automate what can be automated – uncertainty is not a reason to wait out the storm but rather a reason to move now before it’s too late.

The old model tolerated long pilots, year-long science projects, and endless experimentation because the operating environment felt more stable. However, if the market is moving fast, policy is changing rapidly, and supply chains remain volatile, then speed and flexibility become part of the value proposition – a crucial aspect for manufacturers to consider in today’s landscape.

Andre Marino warned against underinvesting: there is still plenty of hype in the market, but waiting too long carries its own strategic cost. The industry needs to strike a balance between caution and timely investment to reap the benefits of automation and AI.

The big questions about AI are trust, reliability, safety, and lifecycle validation – crucial aspects that Mike Cicco highlighted as essential for industrial adoption. He emphasized that AI should be combined with hardcoded, reliable systems that the industry already trusts rather than trying to replace them entirely. This approach will enable AI to add value where it matters most.

The Automate 2026 event has left a lasting impact on the industry, highlighting the need for manufacturers to adapt and innovate in response to changing market conditions. As we move forward, it’s essential to address the challenges of automation and AI adoption while leveraging their potential benefits – driving growth, improving efficiency, and enhancing competitiveness.

The future of industrial robotics and physical intelligence will be crucial topics for discussion as the industry continues to evolve. Manufacturers must prioritize innovation, adaptability, and collaboration to stay ahead in today’s fast-paced landscape.

CyberProof Inc., a co-managed security services company, has introduced an agentic artificial intelligence service designed to automate up to two-thirds of security investigations. The new service, called CyberProof Agentic MXDR, pairs AI agents with human security experts across the detection, response, and exposure management lifecycle.

The service replaces manual security operations center workflows with a collaborative ecosystem of agents that work under human governance. This approach allows for faster alert triage and more accurate threat identification, which can lead to sharper investigation accuracy by as much as 30%. Unlike standalone AI tools or closed AI SOC platforms, CyberProof’s framework works seamlessly with models from various third-party vendors, including Microsoft Corp., Google LLC, Anthropic PBC, and others.

CyberProof Agentic MXDR is designed for large enterprises and mid-market organizations running complex hybrid cloud environments. The service includes a built-in quality control framework that continuously scores AI agents on effectiveness, speed, and cost. This allows chief information security officers to monitor accuracy versus compute spend as frontier model prices rise. CyberProof argues that its agentic approach can help bridge the gap between calendar-driven defense and the rapid pace of threat disclosures.

The company claims that generative AI has compressed the time attackers need to exploit new flaws, forcing organizations to operationalize threat intelligence faster than ever before. In a statement, Chief Executive Tony Velleca noted that collecting threat data is no longer a competitive advantage; instead, it’s about execution, reducing exposure, immediate detection, and containment. Legacy security operations metrics are becoming obsolete due to the speed gap between attackers and defenders.

CyberProof vice president Doron Davidson emphasized that the blend of AI agents and human expertise aims to deliver faster, more consistent investigations with fewer false positives. The 24/7 service reports outcomes through CyberProof’s Reveal360 dashboard and is available immediately for organizations weighing their operational maturity and stack compatibility. A complimentary readiness assessment is being offered to help teams evaluate their preparedness.

CyberProof delivers co-managed security operations built around Microsoft Azure and the wider Microsoft security stack, which it has been using since its founding in 2017. The company’s new agentic MXDR service is designed to streamline security investigations while maintaining human oversight and control. CyberProof Agentic MXDR can help organizations automate up to two-thirds of their security investigations, freeing up resources for more critical tasks.

The launch of CyberProof Agentic MXDR marks a significant step forward in the development of AI-powered security solutions. By combining the strengths of both humans and machines, this service has the potential to revolutionize the way organizations approach threat detection and response. As the cybersecurity landscape continues to evolve at an unprecedented pace, it’s essential for companies like CyberProof to innovate and adapt their strategies to stay ahead of emerging threats.

CyberProof Agentic MXDR is available immediately, with a complimentary readiness assessment offered to help teams evaluate their preparedness. The service reports outcomes through CyberProof’s Reveal360 dashboard, providing real-time insights into security operations. With its agentic approach and built-in quality control framework, this new service has the potential to transform the way organizations manage their security investigations.

CyberProof Agentic MXDR is designed for large enterprises and mid-market organizations running complex hybrid cloud environments. The service can help bridge the gap between calendar-driven defense and the rapid pace of threat disclosures. By automating up to two-thirds of security investigations, this new service has the potential to free up resources for more critical tasks and improve overall investigation accuracy.

CyberProof Agentic MXDR is a significant development in the field of AI-powered security solutions. The company’s agentic approach combines human expertise with machine learning capabilities to deliver faster, more consistent investigations with fewer false positives. With its built-in quality control framework and seamless integration with various third-party vendors, this new service has the potential to revolutionize the way organizations manage their security operations.

Enterprise organizations often struggle with document-heavy processes that consume significant time and resources. The main challenge is extracting valuable information from documents, which are frequently trapped in PDFs, forms, and other formats. This problem persists despite advancements in automation technologies like AI. To address this issue, Nitro has developed a platform called Nitro Automate, designed to bridge the gap between documents and automation workflows.

Built for high-volume document processing, Nitro Automate is an intelligent document automation platform that integrates document processing, workflow automation, and AI-powered integrations. This solution enables teams to eliminate manual document tasks and automate business processes more efficiently. With its fast deployment capabilities and ability to work anywhere PDFs are handled, Nitro Automate can start delivering value almost immediately.

Nitro Automate extends the company’s existing document productivity solutions – Nitro PDF, Nitro Sign, and Nitro Smart Redact – by addressing the challenge of managing document operations across AI agents, workflows, and systems. Instead of relying on employees to manually move documents between systems or extract data from forms and documents, Nitro Automate automates tasks throughout the document lifecycle.

The platform’s capabilities include process automation, conversion, reshaping, transformation, compression, security, and integration with enterprise applications. It works at two levels: team and department level, where it combines document automation, AI-powered data extraction, and workflow orchestration to eliminate manual document work for high-volume processes; and individual level, where it accelerates essential document tasks that still require human judgment.

Across both levels, Nitro Automate’s enterprise integrations make document data accessible to every stakeholder. This enables teams to automate a process or work through it directly. The platform transforms document-intensive workflows in four key ways: eliminating many manual PDF tasks, automating eSignature workflows beyond the signature, unlocking data trapped in documents, and building document workflows that work in the AI era.

One of the primary challenges Nitro Automate addresses is the substantial operational overhead created by repetitive manual document steps. These activities may seem insignificant when viewed individually but can add up to hours of low-value manual work embedded inside otherwise automated workflows. By removing these manual document steps from core business workflows, teams can focus on higher-value work while improving process consistency and reducing errors.

Nitro Automate also helps organizations automate eSignature workflows beyond the signature itself. This involves integrating document preparation, routing, approvals, and Nitro Sign workflows into a single automated experience. Instead of managing signatures manually across disconnected tools, teams can create workflows that automatically move documents through each stage of the lifecycle, resulting in faster turnaround times, improved visibility, and fewer bottlenecks.

Another key capability of Nitro Automate is its ability to transform unstructured document content into actionable business data. Critical information stored inside contracts, invoices, applications, forms, and other documents can be accessed using AI-powered tools like Nitro Automate. This enables downstream systems and workflows to accelerate processes such as invoice and accounts payable automation, employee onboarding, contract management, customer intake, compliance reporting, claims, and case management.

As enterprises move beyond AI copilots that assist with individual tasks towards deploying AI agents capable of executing multi-step workflows autonomously, document automation becomes an increasingly important part of AI strategy. Nitro Automate addresses this challenge by providing flexible integration options supporting both human- and AI-driven workflows. Business teams can build their own automations using low-code and no-code tools, while developers can easily integrate Nitro Automate into existing applications and workflows using APIs.

Nitro Automate is compatible with the Model Context Protocol (MCP), an emerging open source standard that allows AI agents to securely access external tools and services. Through MCP, AI agents can use Nitro’s document automation solution to interact with documents inside business workflows, making Nitro Automate a document execution layer for enterprise AI initiatives.

Organizations investing in automation and AI should recognize the significant opportunity for operational improvement presented by document-intensive processes. By bridging the gap between documents, workflows, enterprise systems, and AI agents, Nitro Automate brings together document processing, eSignature automation, data extraction, and workflow orchestration to create intelligent workflows that can scale alongside next-generation enterprise AI.

Document-heavy processes are often a major bottleneck in business operations. Nitro Automate helps organizations eliminate these bottlenecks by automating tasks such as manual PDF management, signature routing, and document approval. By doing so, teams can improve efficiency, reduce errors, and free up resources for higher-value work.

An ongoing arms race in writing has been gaining attention, with two sides vying for dominance. On one side are AI content detectors that have become increasingly sophisticated. Tools like GPTZero, Turnitin, Originality.ai, and ZeroGPT have made significant strides in identifying AI-generated text. Their earlier versions were relatively easy to bypass, but the newer models have developed a keen sense of what makes writing feel artificial. They can detect consistency in paragraph rhythm, the way language is hedged, vocabulary patterns that lean towards formality, and the absence of human-like tangents and personality.

The other side of this arms race involves writers, marketers, students, and professionals who use AI as part of their workflow. These individuals have been searching for ways to make AI-generated content pass muster with both detection tools and human readers. This is where tools like HumanizeAIText come in – they rewrite AI drafts to introduce variation, idiosyncrasy, and natural roughness that human writing typically exhibits.

To understand the difference between AI-written text and its human counterpart, try this experiment: pull up a piece of AI-generated content, such as something written by ChatGPT or Claude. Read it quickly once, then go back to examine the sentence lengths. Chances are most will fall within a 15-25 word range. Look at how each paragraph opens – often with a topic sentence announcing what’s to come, followed by supporting sentences and sometimes a brief summary or transition.

This structure is competent but recognizable as AI-generated once you’ve seen it enough times. Human writing doesn’t typically follow this pattern. People’s attention wanders mid-sentence; they make points only to immediately qualify them or wonder aloud if the point holds true. They use shorter bursts when something feels urgent and longer, more unwieldy constructions for complex ideas.

The imperfections in human writing aren’t bugs – they’re what makes it feel like someone rather than a machine wrote it. AI detectors have essentially learned to measure this. When writing is too consistent, structured, or evenly paced across a document, flags are raised. This is why raw AI output has become increasingly difficult to pass off as human writing without some kind of intervention.

The obvious response to the challenge posed by AI-generated content is editing it yourself. However, rewriting an AI draft to sound like you actually wrote it requires double attention: tracking what the content says while simultaneously listening for every sentence that sounds machine-like. This process takes time – and for those using AI precisely because they don’t have a lot of time, it can eat up most of the efficiency gains provided by the AI in the first place.

This is where tools like HumanizeAIText fill a practical gap for many people. They’re not replacing human judgment but doing a first pass at addressing mechanical problems: evening out sentence rhythm, breaking up too-perfect paragraph structures, and introducing natural variation that makes writing feel less generated. The user still needs to review the result and ensure it says what they intended.

The real issue with AI-generated content isn’t just whether it triggers a detector but whether it’s actually good at its job – persuading, informing, or connecting with readers. Content that sounds robotic fails not only at the detection level but also at engaging human readers. People can feel when something is off, even if they can’t pinpoint why.

The argument for using tools like HumanizeAIText isn’t just about avoiding detection; it’s about closing the gap between what AI produces and what actually works on a human reader. This involves more than just passing a detector – it requires writing that reads as though someone genuinely cared about its content.

The world of AI video generation has seen a radical transformation in recent years. Gone are the days when creators focused solely on visually appealing backgrounds or cinematic landscapes. Today, there’s an emphasis on mastering hyper-personalized character consistency and emotional depth.

Our evaluation process involved putting each engine through identical tests to assess its performance in creating personalized short videos, social media content, and emotionally resonant narratives centered around the creator. We wanted to find out which tools could deliver high-quality results without breaking the bank or risking legal issues.

One tool that really stood out from the rest is CoupleLens - an AI video generator with exceptional character fidelity tests. When we uploaded a static reference photo, it seamlessly integrated individual portraits into cohesive and dynamic scenes. The company claims their model preserves 98% of facial geometric features and natural micro-expressions even when applying its vast library of styles and filters.

Another standout tool is Runway - a top choice for technical directors who require total mastery over camera physics and background layering. Using Motion Brush Pro 2.0, Runway showed breathtaking control in isolated background adjustments. However, setting up consistent character tracking requires extensive advanced prompt weighting - something that’s not exactly beginner-friendly.

For indie filmmakers on a tight budget, Kling is an excellent option for generating long-duration B-roll content. It handled continuous 10-to-15 second blocks with impressive physical accuracy and features a great multi-language native audio generator that auto-syncs lip movements to generated characters. This feature alone makes it a valuable asset for any filmmaker working on a shoestring budget.

Veo is another top performer in our evaluation - particularly when generating text-to-video scenes that look like natural documentaries complete with ambient soundscapes. Its built-in spatial audio logic seamlessly generates localized sound effects based entirely on what’s moving in the frame. The result is an immersive viewing experience that feels authentic and engaging.

For corporate creative teams who cannot afford even a 1% risk of a copyright lawsuit, Firefly is an essential tool for image and video generation. Integrated directly into Adobe Premiere Pro timeline, it effortlessly handles clean texture generations, object additions, and minor transitions using assets that are entirely safe for public distribution - no risk of plagiarism or infringement.

AI tools like these can be a game-changer for creators who want to produce high-quality content without breaking the bank. Whether you’re looking for an AI video generator with exceptional character fidelity or one that excels at generating realistic text-to-video scenes, there’s a tool on this list designed to help you succeed.

The right AI video generator depends on your specific creative goals - not just whether you’re solo creator bringing memories to life or corporate team building a brand. We recommend testing the free tiers of CoupleLens first to see which engine best fits your style and needs, so you can get started creating amazing content with ease.

The film Late Night with the Devil has been making waves in the entertainment industry, but not for its plot or cast. Instead, it’s caught heat online due to allegations of using AI-generated imagery. The movie, which stars David Dastmalchian as a late-night TV host trying to save his show with a Halloween special gone wrong, was initially praised by critics, including Mashable’s review at SXSW 2023. However, some viewers have taken issue with the film’s use of artificial intelligence in its visuals.

The controversy began when user ‘based gizmo’ posted a one-star review on Letterboxd on March 19, stating: ‘There’s AI all over this… Don’t let this be the start of accepting this shit in your entertainment.’ This sparked a wave of discussion online, with users taking to X (formerly Twitter) to share screenshots from the film’s trailer and dissect the allegations. The focus was particularly on interstitials throughout the fictional live TV broadcast, which included illustrations such as a skeleton dancing in a pumpkin patch.

As the conversation around AI-generated images grew, Mashable reached out to Shudder for comment. In response, co-writers and co-directors Cameron Cairnes and Colin Cairnes confirmed that their film did indeed use AI. According to them, they experimented with artificial intelligence for three still images, which were then edited further and appeared as brief interstitials in the movie. They credited their graphics and production design team for helping create a 70s aesthetic.

The statement from Late Night with the Devil’s creators echoes recent trends in film and TV using AI-generated imagery. Marvel’s Secret Invasion used AI to create its opening credits last year, while True Detective: Night Country faced criticism earlier this year over background posters that looked suspiciously like they were created by a machine. This pushback comes at a time when Hollywood is still reeling from the WGA and SAG strikes, during which both unions fought for protections against AI replacing human work.

Late Night with the Devil premieres in theaters on March 22 and will be available to stream on Shudder starting April 19. The controversy surrounding its use of AI-generated imagery has sparked a wider conversation about the role of artificial intelligence in filmmaking, leaving audiences and industry professionals alike wondering where this trend is headed.

The Vatican’s Interdicasterial Commission on Artificial Intelligence (ICAI) has convened for its inaugural meeting, marking the beginning of a new body tasked with coordinating efforts among various Holy See entities involved in studying, reflecting, and utilizing artificial intelligence. The gathering at Palazzo San Calisto brought together representatives from key Dicasteries and Pontifical Academies to share ongoing initiatives, identify common priorities, and outline initial steps towards promoting an AI vision that serves human dignity, the common good, and the Church’s mission.

The meeting aimed to facilitate a comprehensive understanding of current developments in artificial intelligence. Cardinal Michael Czerny, S.J., Prefect of the Dicastery for Promoting Integral Human Development, highlighted four significant aspects: the rapid pace of AI evolution, its impact on human dignity, growing dialogue between the Church and technology sectors, and the strong resonance generated by the encyclical Magnifica Humanitas. He emphasized the need for prudent discernment in addressing both opportunities and risks associated with this phenomenon.

The participating institutions presented their activities and reflections on artificial intelligence, referencing studies and initiatives focused on its scientific, social, ethical, and educational impacts. The meeting also recalled the path initiated through the Congress on AI Ethics and the Rome Call for AI Ethics, as well as the Holy Father’s message on the relationship between artificial intelligence and peace during the World Day of Peace 2024.

The role of the new Commission was a central topic of discussion. A broad consensus emerged regarding the need for a twofold service: fostering internal coordination, information sharing, and reflection within the institutions of the Holy See, while also serving as a point of reference for discernment and support for initiatives in this field. Participants emphasized the importance of promoting open dialogue with academic, scientific, business communities, and Bishops’ Conferences.

The Church’s engagement with societal transformations was underscored by participants. They noted that reflection must continue to engage with these changes. In light of this, the Commission considered establishing a dedicated website for sharing initiatives and facilitating information circulation.

Regarding its next steps, the Commission decided on gradual progression. Initial actions could include mapping existing initiatives, collecting themes identified by various entities, and working on guidelines concerning AI use within the Holy See. The meeting concluded with expressions of gratitude for the first ICAI gathering, which provided an interesting overview of initiatives and facilitated sharing among Dicasteries represented.

The Commission will reconvene in mid-July to continue its work. This development underscores the Vatican’s commitment to exploring the implications of artificial intelligence on human dignity and the Church’s mission.

Artificial intelligence has become an integral part of education, capable of accurately summarizing novels, writing essays, solving math problems, and coding – all within seconds. However, a popular narrative online suggests that most college students use AI to automate their work and cheat. Middlebury College assistant economics professor Germán Reyes disputes this notion, citing research conducted with colleague Zara Contractor that reveals a more nuanced story about student AI use.

Reyes and Contractor’s two-part study aimed to understand how students are using AI, what they think they’re learning from it, and what they’re actually learning. The first part involved surveying Middlebury students from December 2024 to February 2025 about their AI use at school. The results showed that 80% of Middlebury students use AI for academic work, but surprisingly, the majority do so as an augmentation tool rather than automation.

Augmentation refers to working with the AI tool to enhance learning, whereas automation involves relying on the AI to do the work for you. Reyes describes this distinction as crucial in understanding how students are using AI. To verify their data and ensure its applicability to other educational scenarios, the research team asked Anthropic, the company behind the AI software Claude, about student usage patterns with college email addresses. The response confirmed that most students were using AI for ‘technical explanations,’ rather than automating work.

A comparison of Middlebury’s survey results with global data from over 50 countries further supported the notion that college students are not primarily using AI to automate their work. Reyes notes that higher education institutions need a deeper understanding of how AI is being used by students before creating policies about its use. He warns against banning AI altogether, as this could inadvertently harm students who benefit from it.

If some students are using AI to learn more and an institution decides to ban the technology due to concerns over its potential misuse, then those students who rely on AI for learning will be negatively impacted. Reyes emphasizes the importance of having a good understanding of ground-level facts when making policy decisions about AI use in education.

The second part of the study aimed to investigate how AI impacts learning. Contrary to expectations based on other research at the time, the findings revealed that students who used AI for augmentation purposes performed better than those without access to AI tools. However, it was not simply using AI that determined student performance – rather, it was how they used it.

Reyes and his colleagues found that when students automated AI to write their essays, they initially did well but ultimately underperformed compared to those who used AI for augmentation purposes. The researchers classified prompts as either augmentation or automation based on the conversations between students and the AI tool. They discovered that augmentation users had a smaller effect on the essay in week one but a larger impact in week two.

This suggests that relying on AI to do the work for you may provide short-term gains, but it comes at the cost of lower learning in the long run. Reyes stresses that the effects of AI largely depend on how students use the tools and explains why similar studies can have different outcomes. He believes that many people tend to be pessimistic about student AI usage, but when speaking with students, they reveal a nuanced and sophisticated understanding of both its benefits and drawbacks.

The study’s findings also highlight the importance of considering factors like grade inflation in educational settings. Reyes notes that if there is high grade inflation at an institution, ‘there is less value to augmentation because you need to differentiate yourself from all other students who also have a very high GPA.’ He suggests that this could lead students to use AI for automation purposes, which may ultimately hinder their learning.

Reyes and Contractor are currently finalizing their paper on the experimentation portion of their study. However, they plan to continue researching student AI usage in education, with plans to conduct another survey in 2026. They aim to explore new questions, such as whether students are paying for premium versions of AI assistants.

The job market is undergoing a significant transformation with the increasing use of artificial intelligence (AI) in hiring processes. Employers are leveraging AI to review resumes before scheduling interviews, and even using it during virtual interviews to assess candidates’ skills and fit for the role. This shift has raised concerns about bias, authenticity, and the impact on human connection in the job hunt.

As of 2025, a staggering 87% of employers use AI in at least one stage of their hiring process, with resume screening being the most common application. The AI used to review resumes is collectively known as Applicant Tracking Systems (ATS), software that scans for keywords and skills. Tools like HireVue, Paradox, and Indeed Talent Scout offer advanced AI screening for skills, context, and even interviews to help companies filter candidates.

The primary reason employers are turning to AI is to handle the volume of applications, which has increased by 51% since generative AI tools became mainstream. AI can reduce time to hire by up to 75%, making it an attractive solution for businesses looking to streamline their recruitment processes. However, this speed comes at a cost: hiring managers find it harder to assess if a candidate is ‘authentic’ because everyone is using AI to polish their resumes.

AI parsing extracts data from resumes and matches it against job descriptions using specific keywords, leading to a strategic ‘cat-and-mouse’ game between applicants and employers. Roughly 65% of job seekers now modify their resumes specifically to appease algorithms, sometimes removing genuine achievements to make room for buzzwords. This raises concerns about the accuracy and fairness of AI-driven hiring processes.

A 2025 study reveals that while AI can improve hiring fairness, it can also mirror historical biases if not properly ‘debiasing.’ Without proper debiasing, some models favored white-associated names 85% of the time compared to just 9% for Black-associated names. This highlights the need for more robust data analysis tools and algorithms that can detect and mitigate bias in hiring processes.

Both employers and applicants now have access to AI tools during virtual interviews. Applicants can use AI platforms to ‘listen’ to the interview and provide scripted responses that may appeal to the interviewer. These tools can be tailored to specific interview types, assisting with content knowledge, behavioral questions, and in-the-moment skill assessments.

Interviewers also have access to AI platforms that analyze body language, eye movement, and responses to corroborate the candidate’s veracity and enthusiasm. However, this raises concerns about the potential for bias in these tools and the impact on human connection in the job hunt.

The most ‘AI-aware’ generation is also the most skeptical when it comes to using AI in their job search. While 90% of students use AI for schoolwork, only about 33% of graduating seniors report using AI for their job search. About 16% of students avoid using AI in applications because they fear employers will discover it and disqualify them.

Conversely, 46% of Gen Z hiring managers have ‘caught’ candidates using AI to cheat on assessments. Many students feel that the ‘humanity’ has been stripped from the process. They aren’t just applying for jobs; they are trying to solve an algorithm.

The use of AI in hiring processes is a double-edged sword. On one hand, it can transform hiring by providing more accurate and efficient results. However, at its worst, it automates poor judgment and hides bias behind a ‘math’ facade. Technology is powerful, but using it to replace human connection risks a permanent crisis of trust.

As Amazon’s Prime Day 2026 gets underway, a significant shift in consumer behavior is becoming increasingly apparent. The four-day event, running from June 23rd to 26th, marks a moment when an increasing number of consumers are turning to AI assistants for product research and recommendations rather than traditional search engines or browsing websites. This change in behavior has created a new advertising surface that brands can leverage to reach shoppers at the point of decision-making.

Thrad, an AI-native advertising infrastructure platform, is enabling brands to tap into this emerging trend by placing paid ads within AI conversations. These ads appear as relevant and clearly labelled recommendations inside exchanges between consumers and their AI assistants, providing a more targeted approach than traditional keyword-based search engine marketing. When a shopper asks an assistant for the best noise-cancelling headphones under £200 in the Prime Day sales, the query carries far greater intent than a generic keyword search.

The platform’s advantage lies in its ability to reach across the conversational AI ecosystem rather than being limited to a single destination or platform. Thrad serves native in-chat ads across a network of over 50 AI publishers, spanning general assistants and vertical AI tools in categories such as beauty, electronics, travel, and lifestyle. This comprehensive approach allows brands to connect with consumers at multiple touchpoints throughout their research journey.

Thrad’s platform also enables brands to run ads inside ChatGPT through the same interface, providing a seamless way for advertisers to reach users across various AI tools. Campaigns can go live on the same day, and real-time bidding scores every placement against the live conversation, ensuring that ads are delivered at the most relevant moments. The company already serves millions of ads per day for Fortune 500 advertisers and growth-stage brands.

The timing of this development reflects a broader trend in marketing budgets shifting towards AI initiatives. According to Gartner’s 2026 CMO Spend Survey, published in May, chief marketing officers are now allocating an average of 15.3% of their budgets to AI projects. As AI assistants become increasingly prominent as primary surfaces for product discovery, brands that learn to operate effectively within conversational environments during events like Prime Day will be well-positioned to capitalize on this rapidly expanding channel.

Thrad has published a comprehensive guide to help brands navigate the opportunity presented by Prime Day and beyond. The guide outlines how to plan, launch, and measure conversational AI campaigns through the event window and in the days that follow. Notably, many shoppers continue their research during sales events and make purchases within 48-72 hours afterwards – a period where conversational ads often deliver strong returns.

Thrad’s CEO and co-founder, Andrea F. Tortella, emphasizes the significance of this shift: ‘Shopping is becoming a conversation, and Prime Day is the clearest signal yet of that change.’ She notes that people are increasingly turning to AI assistants for recommendations at the point of decision-making, making it essential for brands to be present in these conversations naturally across the entire AI ecosystem rather than focusing on individual platforms.

Roger Dunn, chief commercial officer at Thrad, highlights the value proposition of conversational advertising: ‘Retail media works because it reaches people at the moment they are about to purchase. AI conversation advertising takes this a step further by reaching them when they’re asking what to buy or comparing options – before scrolling past a shelf.’ He emphasizes that Prime Day represents an ideal opportunity for brands to be genuinely helpful in these exchanges and capitalize on the growing importance of conversational marketing.

In light of Thrad’s efforts, it is clear that AI assistants are becoming increasingly integral to consumer behavior during major retail events like Prime Day. As this trend continues to evolve, brands will need to adapt their strategies to effectively engage with consumers through conversational interfaces.

Dukascopy Bank has taken a significant step forward in trading technology by launching an AI-powered feature that enables clients to interact with their trading accounts through chat-based assistants. This innovative system allows users to execute trades, manage positions, analyze markets, and monitor risk using natural language instructions.

The bank’s decision to integrate AI assistants into its platform reflects a broader shift away from traditional trading interfaces. Clients will no longer need to rely on terminals, manual coding, or complex navigation to perform trading operations. Instead, they can use chat-based assistants like ChatGPT and Claude to manage their accounts.

Dukascopy Bank’s recent expansion of its trading infrastructure has laid the groundwork for this new feature. The Swiss broker-bank introduced a flagship mobile application that combines banking services, retail CFD trading, foreign exchange, payments, and investment tools in a single interface. It also launched a dedicated stock trading platform offering more than 25,000 equity CFDs.

The system runs alongside the bank’s JForex ecosystem as part of a modular execution structure. This setup allows users to access various features and tools without having to navigate complex menus or interfaces. The integration with AI assistants is designed to be seamless, enabling clients to execute trades and manage their positions quickly and efficiently.

Dukascopy Bank CEO Andre Duka emphasized the importance of technology in driving innovation at the bank. ‘Technology has always been at the core of Dukascopy’s DNA,’ he said. The company has invested heavily in building advanced trading infrastructure over the past two decades, and this new feature is a testament to that commitment.

The integration with AI assistants allows for a range of trading actions through chat instructions. Users can place orders, analyze instruments, calculate stop-loss and take-profit levels, and check account exposure using natural language commands. For example, they can instruct their AI assistant to ‘Close half of my gold position’ or ‘Analyze the current market trends in the S&P 500 index’.

The system is built on the Model Context Protocol, an open standard that enables AI tools to interact with external applications. This protocol allows for secure access to trading functions and account data, enabling workflows such as proportional position sizing and cross-asset analysis. The setup process takes only a few minutes and does not require programming knowledge.

The bank has taken steps to ensure the security of client accounts by not sharing passwords with AI assistants. Users can connect their AI assistant directly to their JForex trading account through a Model Context Protocol server developed by Dukascopy Bank. This allows for secure access to trading functions and account data without compromising client confidentiality.

The AI-powered feature is currently available on JForex demo accounts using virtual funds. Support for live trading accounts is planned for summer to autumn 2026, allowing clients to take advantage of this innovative technology in their real-world trading activities.

Docusign has announced a new integration that enables in-house legal teams to automate contract workflows across their entire business using artificial intelligence. The company’s Intelligent Agreement Management (IAM) platform is now available for use with Perplexity Computer and its associated tool, Computer for Counsel. This partnership aims to streamline the often laborious process of drafting, reviewing, signing, and managing contracts by leveraging AI-powered automation tools.

The integration uses Docusign’s Model Context Protocol (MCP) server to connect with Perplexity, allowing legal teams to set objectives using plain language. In response, Docusign can automate contract work from start to finish, reducing the time spent on manual tasks and freeing up resources for more strategic work. This collaboration is particularly beneficial in large enterprises where multiple teams are involved in contracting processes.

Allan Thygesen, CEO of Docusign, emphasized that contracts form a critical part of how businesses operate, with legal teams often stuck juggling multiple tools to manage contract information. By integrating Docusign’s agreement intelligence and workflows directly into the AI tools used by legal teams, the company aims to reduce document management time and enable faster business movement.

Nathan Barksdale, General Counsel at Perplexity, shared his experience with scattered contract work across various tools. He noted that connecting Docusign to Computer means legal teams can automate agreement workflows from end-to-end, allowing them to focus on strategic legal work rather than manual document management tasks.

The integration is powered by the Docusign MCP server and is available today in English globally. This move aligns with Docusign’s mission of bringing agreements to life for nearly 1.9 million customers across more than a billion people in over 180 countries. By leveraging AI-native tools, companies can create, commit, and manage agreements more efficiently.

Docusign has been recognized as one of TIME’s World Most Sustainable Companies 2026, highlighting its commitment to sustainability and business practices that minimize environmental impact.

AQSolotl, a specialist in quantum computing hardware and integrated solutions, has formed a strategic partnership with QuantrolOx, developer of automated machine learning-based control software for quantum technologies. The collaboration aims to accelerate the shift from experimental quantum systems to scalable, production-ready infrastructure by integrating AQSolotl’s Chronos-Q control system with QuantrolOx’s Quantum EDGE platform.

The unified stack created through this partnership is designed to reduce manual calibration and improve system stability in quantum computing. By combining high-speed control hardware with automated machine learning-driven workflows, the joint solution aims to deliver a more efficient operating environment for quantum engineers and researchers. This integration will significantly shorten tuning cycles and enable more consistent qubit performance.

One of the sector’s most persistent challenges is the reliance on manual, time-intensive system tuning. As quantum processors scale beyond early-stage prototypes, increasing qubit counts and tighter performance thresholds make manual calibration impractical. The industry is shifting towards automation and integrated control layers to improve reliability, reduce downtime, and enable continuous operation.

The partnership between AQSolotl and QuantrolOx will be executed in two phases. In the immediate term, the companies plan to deploy AQSolotl’s hardware on the QuantrolOx Quantum Testbed and integrate automated control workflows. This joint solution is expected to support emerging workloads such as quantum AI training, where repeatability and system fidelity are critical.

Vishal Chatrath, CEO and Co-Founder of QuantrolOx, emphasized that automation is becoming essential as quantum systems grow in complexity. ‘By bringing machine learning-driven control directly into AQSolotl’s hardware layer,’ he said, ‘we can enable more stable and repeatable performance with significantly less manual intervention.’

AQSolotl specializes in making advanced quantum technology accessible and practical for industry through its modular quantum control platforms and deployable quantum AI solutions. The company focuses on adaptability, cost-effectiveness, and seamless integration with classical computing to address complex computational problems in sectors such as finance, pharmaceuticals, energy, and logistics.

QuantrolOx develops automated machine learning-based control software that is technology agnostic and applicable to all types of quantum technologies. Their goal is to maximize quantum computer uptime and make their hardware more accessible by automating the tuning process.

A major update has been made to the way developers interact with Google’s Gemini models and agents. The Interactions API, which was launched in public beta last December, is now generally available and serves as the primary interface for these applications. This shift marks a significant change from the legacy generateContent API, although the latter will continue to be supported for the foreseeable future.

The Interactions API offers several key features that make it an attractive choice for developers building with Gemini models and agents. These include server-side state, background execution, tool combination, and multimodal generation capabilities. With this release, a stable schema has been established, and major new functionalities have been added in response to developer requests.

One of the most significant additions is Managed Agents, which allows for more efficient management of agent resources. Background execution enables long-running tasks to be performed without blocking other operations. Gemini Omni will also soon be integrated into the API. These features are designed to support stateful and agentic workflows, making them particularly well-suited for applications that require persistent interactions with users or environments.

The Interactions API is now the default interface across various Google AI tools, including Google AI Studio and the Gemini API. All documentation has been updated to reflect this change, although a toggle remains available to switch back to legacy format snippets. Developers are encouraged to use the Interactions API for new projects and applications, as it offers improved flexibility and efficiency.

For developers already building with supported partners such as LiteLLM, Eigent, or Agno, integrations for the Interactions API are readily available through these platforms. The gemini-interactions-api Skill has also been developed to inject best-practice patterns into agent development contexts, making it easier for agents to stay up-to-date with the latest API developments.

The Interactions API is accessible via Python and JavaScript SDKs. To get started, developers can obtain an API key from Google AI Studio and consult the comprehensive documentation provided by Google. A migration guide has been published to assist those transitioning from the legacy generateContent API, which maps every field to the new schema for a seamless transition.

Google continues to prioritize developer feedback in shaping the Interactions API’s future development. The company encourages developers to share their needs and suggestions on the developer forum, ensuring that the API remains responsive to its users’ requirements.

Fire detection is crucial in subway tunnels, where confined spaces and ventilation systems can complicate safety monitoring. Existing approaches often rely on classical machine learning methods that detect fires based on multivariate correlations but struggle with contextual reasoning.

However, recent studies suggest that Large Language Models (LLMs) may help address this challenge by translating structured sensor summaries into concise semantic descriptions. To examine the role of LLMs in fire detection, researchers have proposed a hybrid framework called HyFiD, which employs an LLM as a semantic feature extractor to augment classical classifiers.

HyFiD converts momentary multi-sensor readings (temperature, smoke, O2, CO, and CO2) into textual assessments of environmental states, generating semantic vectors that are fused with numerical features for supervised classifier training. Experiments on Fire Dynamics Simulator (FDS)-based subway scenarios show that the effect of LLM-derived semantic features is strongly dependent on the downstream classifier.

The results indicate a clear trade-off between detection sensitivity and premature-alarm suppression. Lower thresholds generally preserve higher F1 scores but substantially increase Pre-Alarm Rate, indicating more aggressive alarm behavior. For instance, at τ = 0.25 and k = 1.0 s, SVM + LLM and GBM + LLM achieve high F1 scores of 87.45% and 90.85%, respectively, but their PAR values rise to 99.25% and 51.32%. In contrast, higher thresholds suppress premature alarms more strongly but often degrade F1.

The temporal persistence window further modulates this trade-off. Increasing the persistence window from 1.0 to 2.0 s or 3.0 s generally reduces PAR, but excessive smoothing can substantially reduce F1, especially for SVM + LLM and RF + LLM. GBM + LLM is comparatively more robust under moderate smoothing: at τ = 0.50 and k = 2.0 s, it maintains an F1 score of 88.33% while reducing PAR from 47.55% under the default setting to 28.68%. However, further increasing the window to k = 3.0 s reduces F1 to 83.46%, showing that overly conservative alarm persistence can also suppress valid fire detections.

To justify the architectural design of using the LLM as a semantic feature extractor rather than as an end-to-end decision maker, researchers evaluated direct LLM classification under Zero-shot, One-shot, and Few-shot Chain-of-Thought (CoT) prompting strategies. The results show that direct prompting is highly sensitive to both the prompting setting and the LLM backbone when applied to raw multivariate numerical sensor snapshots.

The final alarm behavior is highly sensitive to the selected threshold and persistence window. Researchers retain τ = 0.50 and k = 1.0 s as the default setting for the main experiments because it provides a consistent operating point for comparing classifiers, while the sensitivity analysis clarifies how alternative alarm rules shift the balance between F1, delay, and pre-ignition false alarms.

Under the Zero-shot CoT setting, several evaluated backbones—including Llama-3-8B41, Qwen-2.5-3B58, Qwen-2.5-7B, and Ministral-8B59—produce near-zero Recall and F1, indicating difficulty in directly mapping raw numerical sensor values to stable fire/no-fire decisions. However, Few-shot CoT improves the strongest prompt-only result, with Llama-3-8B reaching an Accuracy of 65.67% and an F1 score of 70.32%. Nevertheless, this remains below the supervised hybrid approach: GBM + Llama-3-8B achieves an F1 score of 90.77% when the LLM is used as a semantic feature generator rather than as a standalone classifier.

These comparisons suggest that, in this setting, LLMs are more reliable when used to produce intermediate semantic representations for supervised downstream classifiers than when used as direct numerical classifiers. This finding has significant implications for machine learning jobs and AI-generated image applications, where accurate decision-making is critical.