Latest News

Researchers at Unit 42 have developed a new method to assess the safety of large language models (LLMs), revealing that their alignment is often fragile and concentrated in specific areas. This finding has significant implications for the development and deployment of AI systems, particularly when it comes to using LLMs as judges or decision-makers in critical applications.

The team’s previous research on logit-gap steering demonstrated how an attacker can bypass a model’s safety guardrails by closing a measurable gap in its output scores. However, this study aimed to answer the question of where inside the model alignment actually resides and whether it is concentrated or diffuse. This knowledge is crucial for defenders, as it determines whether safety is a robust defense mechanism or a thin layer that can be easily compromised.

Modern LLMs are aligned through reinforcement learning from human feedback (RLHF), which pushes the model to refuse harmful prompts and comply with safe ones during training. However, until now, there has been no method to identify the specific parts of the network responsible for this learned behavior without significant computational costs. The new research introduces a technique called perturbation probing, which achieves this goal at a lower cost.

Perturbation probing involves making targeted changes to an LLM’s input and observing how its output responds. By doing so, researchers can pinpoint the small set of feed-forward neurons that are causally responsible for specific behaviors, such as refusing harmful requests or agreeing with users in multi-turn conversations. This approach requires only two forward passes per prompt and is computationally efficient.

The study used perturbation probing to analyze 13 different LLM models, including Qwen3-4B and Qwen3.5-2B. The results were striking: on the open-source model Qwen3-4B, just 50 neurons out of a total of 350,208 feed-forward neurons controlled the safety refusal template. Removing these 50 neurons changed the response format for 80% of standard harmful-prompt benchmarks, indicating that alignment is indeed fragile and concentrated.

The concentration of alignment in specific areas has significant implications for AI safety. It suggests that relying on a thin layer alone is insufficient and that true defense-in-depth strategies are necessary to prevent attacks. This involves combining external content filters with runtime guardrails layered on top of the base model’s training data, rather than solely relying on internal mechanisms.

The same computation used in perturbation probing also produces a diagnostic called the FFN/Skip ratio, which predicts whether a model’s safety circuit can be easily steered with minimal modifications. This metric explained 81% of the variance in how vulnerable each model’s safety behavior was to targeted changes across all 13 models tested.

The researchers hope that perturbation probing will serve two roles for the AI security community: as a pre-deployment diagnostic and as a tool for repairing fragility. By using this method, security teams can measure how much of a model’s safety rests on thin layers before deploying it in production. Additionally, amplifying identified neurons improved factual self-correction from 52% to 88% without retraining.

The study emphasizes the importance of measuring and auditing LLMs’ safety properties rather than merely asserting them. The researchers urge other experts to read their full paper on arXiv and integrate fragility diagnostics into their evaluation pipelines. For organizations deploying LLMs, Prisma AIRS Runtime Security provides external content filters and inline guardrails that complement the thin template layer.

The findings of this study have significant implications for AI safety research and development. By understanding where alignment resides in an LLM and how concentrated it is, researchers can develop more robust defense mechanisms to prevent attacks. The use of perturbation probing as a diagnostic tool will help ensure that LLMs are deployed safely and securely.

A federal judge has ruled that the Pentagon’s sanctions against AI company Anthropic were unlawful, finding that the government had punished the firm for publicly criticizing its views on AI safety. The ruling makes permanent a temporary suspension of sanctions ordered in March and takes effect immediately.

The decision was made by US District Judge Rita Lin, who described the Pentagon’s designation as ‘illegal and baseless’. This marks a significant blow to the Trump administration’s efforts to restrict Anthropic’s business dealings with the military.

Anthropic had been designated a ‘supply chain risk’ under an obscure government-procurement statute aimed at protecting military systems from foreign sabotage. However, this was the first time a US company has been publicly given such a designation, and it followed Anthropic’s refusal to allow its AI chatbot Claude to be used for US surveillance or autonomous weapons.

The Pentagon had argued that private companies should not be able to constrain military action, but Judge Lin disagreed, stating that ‘the empty invocation of national security is not a blank check to punish and retaliate against government critics’.

Anthropic’s designation was seen as an unprecedented move by Defense Secretary Pete Hegseth, who blocked the company from certain military contracts. This could have cost Anthropic billions of dollars in lost business and reputational harm.

The Trump administration had also lashed out at Anthropic on social media, with President Donald Trump describing it as ‘a radical left, woke company’ and ‘out-of-control’. However, Judge Lin found that this was not a valid reason for the Pentagon’s actions.

Anthropic has alleged that the government violated its right to free speech under the First Amendment of the Constitution by retaliating against its views on AI safety. The company also claimed it was not given a chance to dispute the designation, in violation of its Fifth Amendment right to due process.

The justice department countered that Anthropic’s refusal to lift restrictions could cause uncertainty in the Pentagon over how it could use Claude and risk disabling military systems during operations. However, Judge Lin found this argument to be unsupported by facts and inconsistent with the military’s past praise of Claude.

Anthropic has a second lawsuit pending in Washington DC, over a separate Pentagon supply-chain risk designation that could lead to its exclusion from civilian government contracts. The company says it opposes domestic surveillance as a violation of rights and believes AI models are not reliable enough for use in autonomous weapons.

The ruling is seen as a victory for Anthropic, which has welcomed the court’s decision that the supply chain risk designation was unlawful. However, the government may still appeal the decision.

Over the past two years, our documentation team has been experimenting with hybrid editorial pipelines to accelerate baseline drafting. While large language models excel at structuring raw technical outlines and initial drafts, their default output carries distinct structural markers that can make machine-generated prose feel repetitive to technical readers.

When team leads attempt to humanize AI-generated drafts, the goal is not about passing arbitrary detection scores for the sake of it; rather, it’s about restoring natural sentence cadence, removing formulaic filler, and preserving reader engagement. This document outlines our hands-on observations regarding synthetic text patterns, manual refactoring strategies, and automated syntax-tuning workflows.

To systematically improve synthetic drafts, we first mapped the specific linguistic anomalies that make machine-generated prose feel repetitive to technical readers. We identified three key structural bottlenecks: uniform sentence perplexity, predictable transition hooks, and passive symmetrical lists.

Uniform Sentence Perplexity is a major issue in AI-generated text. Human writing features natural variation in sentence length—a short, punchy claim followed by a longer, clause-rich explanation. LLMs default to medium-length compound sentences, creating a monotonous rhythm across long-form documentation. This can lead to reader fatigue and decreased comprehension.

Predictable Transition Hooks are another problem area. Synthetic drafts rely heavily on a predictable set of conjunctive adverbs like ‘Furthermore,’ ‘Moreover,’ and ‘In conclusion.’ These phrases act as low-entropy crutches that inflate word count without adding informational value. They also make the text feel formulaic and repetitive.

Passive Symmetrical Lists are another issue with AI-generated bullet points. Generative models tend to make every bullet point identical in grammatical structure (e.g., Verb + Noun + Modifier), which feels mechanical during quick skim-reading. This can lead to a sense of monotony and decreased reader engagement.

To address these issues, we developed an automated syntax-tuning workflow that uses specialized tools like BypassGPT to redistribute sentence length, balance perplexity metrics, and strip away predictable synthetic markers before a draft hits human editors. This intermediate processing step adjusts the underlying syntax tree while preserving technical terminology, code snippets, and inline markdown references.

The 4-Step Text Refactoring Protocol is a lightweight checklist that can be applied to any synthetic draft prior to final staging. Step 1: Strip Preambles and Summaries—delete the opening and closing fluff paragraphs entirely. Step 2: Vary Clause Lengths—the 1-3-1 rule recommends reorganizing dense paragraphs so that long technical explanations are anchored by brief statement sentences.

Step 3: Inject Specific Industry Edge Cases—synthetic text tends to generalize, so injecting concrete variables, specific software versions, or practical edge cases can help make the text more engaging and relevant. Step 4: Neutralize Passive Construct Bias—convert passive descriptions of features into active user actions, replacing abstract assertions with operational directives.

A benchmark comparison between raw and refined output shows that humanizing AI-generated text can significantly improve reader engagement and comprehension. Whether your team implements manual sentence-shredding or uses specialized platforms designed to yield undetectable AI text, the core objective remains the same: prioritising clarity, dynamic pacing, and genuine utility for the end reader.

Scaling technical content production requires a clear boundary between draft generation and structural editing. Machine outputs provide an excellent starting canvas, but relying on predictable statistical paths can tire real readers over time. By humanizing AI-generated text, we can create more engaging and effective documentation that meets the needs of our audience.

A federal judge has dealt a significant blow to the Trump administration, ruling that its actions against Anthropic were unlawful. Judge Rita Lin made this decision in a 59-page order.

The court granted summary judgment to Anthropic in claims related to the First Amendment and due process under the Administrative Procedure Act. This means Anthropic wins on these points without needing further trial or evidence.

This dispute began when the Pentagon asked Anthropic to remove restrictions from its Claude LLM, used for tasks like mass surveillance of Americans and lethal autonomous warfare. However, instead of complying with all demands, Anthropic drew a line at two specific uses: mass surveillance and lethal autonomous warfare.

The Trump administration responded by banning every federal agency from using Anthropic’s technology, labeling the company ‘a RADICAL LEFT, WOKE COMPANY’ on Truth Social. Hegseth even accused Anthropic of arrogance and betrayal in a tweet, going so far as to bar any military contractor from doing business with it.

Initially, the government claimed that Anthropic posed a risk due to its alleged backdoor access to deployed models. But they later conceded this was not true – Anthropic has no such access, and Claude is not inherently more risky than other ‘black box’ systems.

Despite conceding on the risk claim, the government argued that trust had been forfeited by Anthropic’s public statements through the press. However, Judge Lin looked at events differently: she noted a timeline of suspicious interactions leading up to the ban.

Days before blacklisting Claude, Hegseth floated invoking the Defense Production Act, which would have made Anthropic essential to national security. And just one day after designating it ‘high-risk,’ an under secretary sent an email saying that a contract was ‘very close.’ This contradicts claims of severed ties and lack of engagement.

Judge Lin found the actions not only retaliatory but also arbitrary and capricious under the Fifth Amendment. She deemed them void from the start because they lacked pre-deprivation process required by law, meaning no proper hearing or review occurred before imposing penalties.

As a result of these findings, Judge Lin vacated the designation and Hegseth’s boycott order. The court also granted a permanent injunction to prevent similar measures in the future and denied the government’s request for an administrative stay – allowing Anthropic to continue business as usual.

In addition, Judge Lin entered judgment in favor of agencies that took no action against Anthropic. This decision is seen as a major victory for Anthropic, which estimates its revenue will take a hit: between 50% and 100% from defense-related sales this year due to the ban.

The complexities of the US health care system can be daunting for many people. Understanding terms like health insurance, in-network providers, and deductibles can feel overwhelming. To get help, some individuals are turning to online forums and websites where they can hear firsthand from others about their experiences with the health care system.

A PhD student at the University of Albany’s School of Social Welfare is analyzing these online posts to better understand the barriers people face within the health care system. Arissa Spitalny’s work uses Reddit, a popular discussion website that hosts communities called subreddits where users can post about specific topics and reply to others’ posts.

Spitalny notes that individuals who may not feel comfortable sharing their concerns in traditional person-to-person interactions have a space on Reddit to anonymously voice their worries, ask for advice, or simply vent. This anonymity allows them to connect with others who share similar experiences without fear of judgment.

The subreddit ‘AskDocs’ is being examined by Spitalny, which has over half a million weekly visitors and was created in 2013. With over one million posts from this community, Spitalny is focusing on those that specifically mention health insurance.

Spitalny’s mentor, Assistant Professor Cheng Ren, specializes in using advanced artificial intelligence techniques and open-access data. Under his guidance, Spitalny developed an AI prompt to review the Reddit posts and classify approximately 15,000 of them that mentioned health insurance. The AI was trained on a dataset of 300 manually reviewed sample posts by Spitalny.

The use of AI significantly streamlined the analysis process for Spitalny, allowing her to efficiently examine the full dataset. She is now in the analysis and writing stage of her study, which will interpret patterns in the data collected from Reddit, including factors such as race, gender, and demographic location.

Spitalny’s research aims to contribute to a better understanding of health care access barriers and the role social work can play in improving care navigation. She hopes that by examining the support individuals seek online, she can gain insight into the lived experiences behind these patterns.

In the future, Spitalny plans to expand her study beyond insurance posts to examine other health care barriers as well.

A new partnership between Posit and Carahsoft Technology Corp. aims to make it easier for government agencies to access data science tools, addressing a common challenge in the public sector.

The agreement makes Posit’s enterprise data science platform available through established procurement contracts, allowing agencies to adopt secure, scalable open source solutions while staying compliant with procurement requirements.

Government agencies increasingly rely on data analysis tools to inform decisions and deliver services. However, they often encounter hurdles when adopting commercial tools that run languages like R and Python securely and at scale.

The Posit platform enables organizations to build, deploy, and scale data analysis, models, and applications using open source tools within a governed, secure environment. This helps teams collaborate more effectively, streamline workflows, and turn data into actionable insights faster.

Posit’s solutions are designed to integrate with existing infrastructure and maintain compliance and control over data and code. The platform is used by thousands of organizations worldwide, from startups to the Fortune 500.

Carahsoft will serve as Posit’s Master Government Aggregator, distributing its solutions through reseller partners and NASA Solutions for Enterprise-Wide Procurement (SEWP) V and OMNIA Partners contracts.

The partnership addresses a significant challenge in the public sector: procurement hurdles when adopting commercial tools that run languages like R and Python securely and at scale. By placing Posit’s platform on contract vehicles agencies already hold, the partnership streamlines access to data analysis tools for government agencies.

Posit provides an enterprise data science platform that enables organizations to build, deploy, and scale data analysis, models, and applications using open source tools within a governed, secure environment. The platform is designed to help teams collaborate more effectively, streamline workflows, and turn data into actionable insights faster.

The partnership between Posit and Carahsoft aims to accelerate data-driven decision making in the public sector by providing access to AI-powered data science solutions through established procurement contracts.

Posit’s VP of Alliances, Adam Smith, stated that partnering with Carahsoft enables them to expand and streamline access to their AI-powered data science platform across the Public Sector. He noted that Carahsoft brings a deep understanding of the Federal procurement landscape, allowing Posit to meet agencies where they are.

Carahsoft’s Vice President, Natalie Gregory, added that they are pleased to add Posit to their portfolio of enterprise open source software providers. Together with reseller partners, they enable data teams to derive greater business value from their data using Posit’s enterprise platform for the AI era.

Posit’s solutions are available through Carahsoft’s SEWP V contracts NNG15SC03B and NNG15SC27B and OMNIA Partners Contract #R240303. For more information, contact the Carahsoft Team at (703) 581-6680 or [email protected]; or learn more about Posit’s solutions here.

Mango AI’s innovative AI dance generator is making waves in the world of content creation, allowing users to transform a single photo into an animated dancing video. This user-friendly tool applies natural and rhythmic motion to the subject while preserving its original look, eliminating the need for manual animation or filming.

To get started, users simply upload a front-facing, full-body photo in JPG, JPEG, PNG, or WEBP format. They can then choose from a built-in dance template library that focuses on trending and contemporary moves or upload their own reference clip to define their choreography. The AI dance generator maps the selected motion onto the subject and renders a short dance video ready for download or posting.

The tool also offers a dedicated picture-to-dance feature, giving creators more than one route to a finished dance clip. This flexibility makes it an attractive option for businesses looking for efficient ways to create engaging content with their AI tools for business needs. With Mango AI’s free AI video generator, users can produce high-quality videos without requiring extensive editing experience.

Mango AI provides two models to cater to different user needs: Mango AI 1.0 supports human images and videos, while Mango AI 2.0 extends coverage to human, cartoon, and animal images and videos. This broad subject support makes it an ideal choice for businesses seeking comprehensive solutions with their ai tools for business requirements.

The finished dance videos are concise and visually clear, making them suitable for social media platforms like Facebook, Instagram, X, and others. The AI dance generator keeps facial features, clothing, and overall style intact, which is particularly useful for creating memes, short-form entertainment, and playful promotional posts.

Animal lovers can also extend the idea with Mango AI’s pet dance tool, which brings the same motion approach to photos of cats, dogs, and other pets. This feature adds a fun twist to content creation, allowing users to bring their favorite animals into animated dancing videos.

The AI dance generator is part of a growing set of AI video tools from Mango AI aimed at simplifying creative content production. According to Winston Zhang, CEO of Mango Animate, ‘The AI dance generator lets anyone add believable movement to a static image, so a single photo becomes an entertaining dance video in just a few steps.’

Mango Animate is a company founded on the mission to make video creation approachable and accessible. Its suite of animation software and AI-driven video tools supports creators across various industries, including marketing, education, entertainment, and everyday content production with speed and simplicity.

The Mango AI suite spans photo animation, face and video effects, and voice features, providing users with a range of options for creating engaging videos. With its user-friendly interface and efficient workflow, the AI dance generator is an excellent addition to any creator’s toolkit.

A new framework for integrating advanced language models like Anthropic’s Claude with physical objects has been unveiled by the company. The Model Hardware Standard (MHS) is designed to connect large language models with equipment in manufacturing, science labs, and other settings. This marks a significant move into so-called physical AI for Anthropic, which typically focuses on developing and applying its language models in software environments.

With MHS, companies can integrate AI into their equipment quickly – often within hours or minutes – whereas this process usually takes weeks or even months with custom builds required by specialists. The standard is expected to pave the way for autonomous experiments and workflows that run around the clock without human intervention. Scientific research and advanced manufacturing applications are among its primary uses.

The MHS framework can connect multiple devices, enabling them to communicate through a set of commands like ‘read.’ Any hardware device can understand these commands and act on them. This standard is model-agnostic, meaning it works with any large language model – not just Claude – including models built by other companies such as OpenAI or open-source models.

MHS is built upon the Model Context Protocol (MCP), a universal, open standard for connecting data sources that Anthropic debuted in 2024. The MCP serves as a kind of ‘USB’ for AI to software connection, facilitating seamless interaction between different systems and devices. According to Alek Kemeny, a member of the technical staff at Anthropic, not all existing equipment can connect directly to MHS due to lack of programming interfaces.

As part of this project, Anthropic is collaborating with numerous device manufacturers to develop new products that come pre-loaded with the necessary interface for connecting to MHS. The company is also assisting in adding these connections to existing products. This effort aims to streamline the process of integrating AI into various settings and applications.

Anthropic’s partnership with HHMI Janelia Research Campus, a biomedical research center in Virginia, played a crucial role in developing MHS. A small group of labs and hardware manufacturers received early access during development, focusing on fields such as biotech, robotics, and quantum computing. Some notable partners include Genentech, Carnegie Mellon university, QuEra, Universal Robots, Amazon Web Services, Doosan Robotics, Danaher, and Hugging Face.

The MHS research preview comes at a time when interest in combining AI with robotics is on the rise. Other companies like Nvidia have also been championing physical AI, predicting that industrial companies will soon become robotics companies. Anthropic’s move into this space marks an important step forward for its Claude language model and highlights the potential of integrating advanced AI tools with various industries.

According to Jonah Cool, head of partnerships and deployment of science at Anthropic, proprietary solutions often hinder scientific progress by being brittle and inflexible. MHS offers a standardized interface that aims to connect any model to equipment easily, avoiding vendor lock-in for scientists. This development has significant implications for the future of AI integration in various settings.

A new interface from Anthropic is set to simplify the way artificial intelligence agents interact with machinery, marking a significant step into the physical world for the AI giant. The Model Hardware Standard (MHS) enables seamless communication between devices, including those used in scientific research and advanced manufacturing.

The MHS works by standardizing how information is transmitted between devices, much like a USB-C cord does for data transfer. This means that any device with a programmable interface can be easily integrated into an AI system using Anthropic’s technology.

According to Elizabeth Kelly, head of beneficial deployments at Anthropic, the new standard has far-reaching implications beyond just scientific research. ‘We built this for science to show the promise of AI,’ she said, ‘but there are huge benefits here for enterprise and industry as well.’

The announcement signals that Anthropic is expanding its presence in hardware development, where rivals such as OpenAI and Amazon have invested heavily in designing AI-native devices and manufacturing tools. To support this effort, Anthropic has built out a silicon team to design custom chips for its models.

Anthropic’s recent hire of Caitlin Kalinowski, a seasoned hardware executive with experience at OpenAI, Meta, and Apple, further underscores the company’s commitment to developing its own hardware capabilities. The new standard is also model-agnostic, meaning users are not limited to Anthropic’s family of Claude models.

The Model Hardware Standard is initially available to select organizations in science, robotics, and manufacturing as part of a research preview. However, Anthropic plans to open-source the standard eventually, allowing any device manufacturer to adopt it.

Anthropic has already demonstrated its commitment to open-sourcing standards with the release of the Model Context Protocol last year. This protocol enables easier connections between AI agents and data sources.

A federal judge in San Francisco has ruled that the Pentagon’s decision to blacklist artificial intelligence company Anthropic was unlawful. The move marks a significant victory for the AI giant, which had been designated a supply chain risk by the Department of Defense earlier this year.

According to court documents, U.S. District Judge Rita Lin found that the DOD violated the First Amendment by designating Anthropic as a supply chain risk based on its criticism of the government’s views on AI use. The judge wrote that while the government has some leeway when it comes to matters of national security, its actions were not founded on any ‘articulable basis’.

The Pentagon had designated Anthropic a supply chain risk in March after negotiations over how the military could use its Claude AI models broke down. This designation meant that defense contractors were barred from using Anthropic’s technology in their work with the agency.

Anthropic sued the Trump administration in San Francisco and Washington, D.C., seeking to reverse its blacklisting. The company argued that the government had no basis for designating it a supply chain risk based on its public statements criticizing the Administration’s views on AI use.

In her ruling, Judge Lin noted that the Constitution does not allow the government to impose sweeping penalties based solely on criticism of its policies. She also pointed out that Anthropic’s actions did not pose any actual threat to national security.

The company has welcomed the court’s decision and expressed its commitment to working with the government to harness AI for national security purposes. ‘We welcome the court’s ruling that this supply chain risk designation was unlawful,’ a spokesperson said in a statement. ‘We remain focused on working productively with the government to benefit all Americans from this technology.’

The ruling is significant not only for Anthropic but also for the broader AI industry, which has been watching the case closely. The company’s ability to reestablish ties with the Pentagon could open up new business opportunities and pave the way for its expected near-record IPO.

While the litigation in Washington, D.C. is still ongoing, this ruling clears a major hurdle for Anthropic as it continues to navigate the complex landscape of AI regulation.

The tech company Anthropic is expanding its support for scientists, aiming to accelerate scientific discovery through the use of its language model, Claude. Building on previous efforts, including the launch of Claude Science in June, Anthropic is now offering free and discounted subscriptions to researchers worldwide.

In a significant move, Anthropic has opened 10,000 seats for scientists to access Claude at no cost or reduced rates for one year through its new Claude team plan for scientists. Standard seats will be available for free, while premium seats with increased usage limits can be obtained for $15 per month. This program is set to extend beyond the initial 10,000 seats in the coming months.

The expansion of Anthropic’s AI for Science program also includes offering credits to researchers working on high-impact projects across various scientific fields. Initially focused on biological sciences, the company now plans to support scientists in other areas as well, such as those conducting ambitious and computationally intensive research like solving complex mathematical problems or protein design.

By providing access to Claude through subscriptions and credits, Anthropic aims to significantly accelerate scientific discovery. The company believes that its tools can help researchers streamline their work and achieve breakthroughs more efficiently.

To register for the Claude team plan, principal investigators at academic or nonprofit institutions must complete a verification form. Once verified, they can add researchers from their lab to the plan. Researchers working in biology and chemistry will still be limited to Opus-class models due to potential dual-use risks associated with using higher-level models like Fable.

Anthropic is also collaborating with the US government on an access program for life sciences professionals to use Mythos-class models safely. The company has already enrolled its first participants, with plans to increase access soon. Additionally, Anthropic is working towards establishing a shared specification for AI agents to operate physical devices, known as the Model Hardware Standard (MHS), which will be available in a research preview to select scientific labs and manufacturers.

Anthropic’s efforts also include funding independent research into how AI impacts users’ wellbeing through a $5 million grant program. Furthermore, the company has published an article explaining its text watermarking method used in Claude, addressing concerns about its effects on outputs.

A three-day research programme, the Google DeepMind APAC Research Symposium 2026, is set to bring together artificial intelligence researchers from across the Asia-Pacific region. Organised by Google DeepMind in collaboration with Google Cloud, this symposium aims to provide participants with opportunities to explore major developments in AI and engage with leading experts.

The programme will feature separate sessions for faculty members and students, creating a platform for both established researchers and emerging AI talent to exchange ideas and learn from each other. Faculty Day on 28 October 2026 is dedicated exclusively to invited academic faculty members from the APAC region, while Student Days on 29-30 October 2026 will focus on selected students and invited speakers.

The symposium’s agenda includes keynote presentations from leading experts, research showcases highlighting emerging AI work, interactive roundtable discussions, and exploration of AI systems and infrastructure. Faculty Day is designed to facilitate discussions around emerging research directions, infrastructure, and developments in artificial intelligence among the region’s top researchers.

Student Days will provide high-calibre undergraduate, master’s, and PhD students with opportunities to learn directly from global AI experts through a combination of technical and interactive sessions. The student programme includes deep-dive lectures on emerging AI research, poster showcases featuring research and ideas, mentorship panels with experienced AI researchers, and exposure to cutting-edge developments across major AI research areas.

The symposium will explore four key research tracks: Agentic AI, Multimodal AI, AI for World, and Systems for AI. These tracks cover important areas of modern artificial intelligence, including systems capable of performing tasks with greater autonomy, processing multiple forms of information, applying AI to real-world challenges, and developing the infrastructure required to support increasingly capable AI systems.

The programme is tailored to members of the academic and research community across the APAC region, with different participation arrangements for faculty and students. The student component is particularly relevant to undergraduate students studying AI-related disciplines, master’s students pursuing advanced study in AI fields, PhD students conducting research in artificial intelligence, and emerging researchers interested in learning from leading global AI experts.

The Google DeepMind APAC Research Symposium provides an opportunity for participants to gain exposure to frontier AI research while engaging with researchers working on some of the field’s most important technical challenges. Participants can benefit from expert-led learning, research showcases, mentorship discussions, and interaction with a broader AI research community.

Faculty members will have opportunities to exchange ideas, examine emerging research challenges, and engage directly with other researchers working across the region. Students will be able to learn about cutting-edge developments in AI tools for businesses and explore how these technologies can contribute to real-world applications and societal outcomes.

The symposium is expected to provide students with valuable exposure to current research directions while enabling them to connect with researchers and emerging professionals from across the APAC region. This event marks a significant opportunity for participants to gain insights into the latest AI developments, methodologies, technological limitations, and areas requiring further investigation in the field of artificial intelligence.

The Google DeepMind APAC Research Symposium will take place from 28-30 October 2026, with Faculty Day on 28 October and Student Days on 29-30 October. The symposium is open to members of the academic and research community across the APAC region, with different participation arrangements for faculty and students.

Participants can benefit from expert-led learning, research showcases, mentorship discussions, and interaction with a broader AI research community. This event provides an opportunity for emerging researchers to better understand current AI developments, research methodologies, technological limitations, and areas requiring further investigation in the field of artificial intelligence.

A team of researchers at the University of Pennsylvania has created a program designed to automate tissue culturing. This initiative uses advanced technology, known as organ-on-a-chip, to provide more accurate models of human body parts for testing new drugs.

Bioengineering professor Dan Dongeun Huh teamed up with Andrei Georgescu after they met on campus when Georgescu was a student at Penn. They discovered that both were interested in drug testing and tissue development, which led them to brainstorm ideas for improving the collection of preclinical data before human trials.

The researchers use tiny devices to grow cells. These techniques were initially developed for manufacturing computer chips. The devices are equipped with organ-on-a-chip technology, allowing hundreds of samples to be grown on a single disk and streamlining the culturing process.

When developing new drugs, scientists usually test them on simplified models or animals. However, this approach doesn’t accurately replicate what cells experience in the human body, making it fundamentally flawed.

‘99% of the drugs entering clinical trials fail,’ Huh explained. ‘It costs close to $1 billion to develop just one successful drug.’ This highlights a pressing need for more efficient and effective testing methods that can help reduce these failures.

The Penn researchers’ program uses robots to manufacture human tissues, test 10,000 lab-grown cell models simultaneously, and generate data within one to two weeks. The startup partners with pharmaceutical companies to determine which drugs are administered during testing.

Vivodyne’s HIVE laboratories currently operate in San Francisco and Philadelphia, with plans for further expansion. This is not the only innovation born at Penn - other AI startups have also emerged from the university, including an early-stage mortgage company aiming to modernize loan processing.

A breakthrough in quantum computing has been achieved by QuEra Computing, a leader in the field. The company has successfully automated a critical subsystem of its quantum computers using artificial intelligence (AI), enabling the acceleration of commercial-grade deployments. This development marks a significant milestone for the industry, as it addresses one of the major challenges facing the widespread adoption of quantum computing: the need for human expert intervention to maintain and operate these complex machines.

The AI agent used in this achievement is Anthropic’s Claude, which was developed and validated by QuEra’s team. Claude wrote its own control software, allowing it to recover from laser system failures in seconds, compared to the minutes required by a specialist. This automation not only saves time but also reduces the need for manual intervention, making quantum computing more practical and efficient.

Quantum computers rely on lasers held at precise frequencies to control atomic qubits. However, these lasers drift over time, causing the machine to malfunction unless corrected by an expert. The increased number of lasers in each generation of machines has made this a significant challenge, as every laser requires manual tuning and commissioning. This process can take weeks, tying up scarce human expertise.

QuEra’s team spent two to three weeks writing a recovery script for common laser disturbances by hand. However, this approach had limitations, as it only handled failures that the authors anticipated listing. To overcome these challenges, QuEra turned to Claude and the Model Hardware Standard (MHS), a new standard for AI agents to safely operate physical equipment in scientific research and advanced manufacturing.

The MHS allows AI agents like Claude to run their own experiments on dedicated testbeds, propose fixes, try them out, read results, and refine. This process is continuous, including overnight runs, covering hundreds of failure cases that no team of specialists could work through by hand. Engineers set the scope, reviewed every step, and decided what counted as proof of success.

The software produced by Claude is a conventional, fully inspectable program, not a model making decisions at runtime. Moreover, it operates within safety bounds declared in the standard itself, ensuring safe operation from the outset. This approach has been validated through rigorous testing, with results showing that Claude can recover reliably and quickly, even when faced with real faults.

The pilot project demonstrated several key benefits of this AI-driven automation: reliable recovery, rapid response times, handling of real faults rather than just test cases, improved performance over expert tuning, and the ability to transfer settings between lasers. These achievements have significant implications for the deployment of commercial-grade quantum computers.

Prior to this breakthrough, every laser used in a quantum computer required standing on scarce human expertise time. On-site recovery was necessary at any hour, taking up to half an hour per tuning session, and commissioning a new operating point took weeks. With Claude’s automation, these limitations are significantly reduced, making it possible for more systems to be deployed and properly supported.

Sergio H. Cantu, Vice President of Quantum Systems at QuEra Computing, highlighted the importance of this achievement: ‘For years, the hardest part of scaling quantum computers wasn’t the physics; it was the people driving at 2 am to fix a laser lock.’ He emphasized that QuEra has built a solution using the Model Hardware Standard to address this challenge.

QuEra’s participation in the MHS research preview is an early example of AI partnerships. The company works closely with strategic partners, including Amazon Web Services for cloud delivery and HPE for on-premises high-performance computing integration. This collaboration enables QuEra to leverage cutting-edge technologies like machine learning jobs and apply them to real-world challenges.

The Model Hardware Standard (MHS) is a new standard that allows AI agents to safely operate physical equipment in scientific research and advanced manufacturing. It started as a collaboration between Anthropic and HHMI Janelia Research Campus, with QuEra participating in the limited research preview. The MHS ensures safe operation by declaring bounds, interlocks, and emergency stops within its own standard.

QuEra Computing is at the forefront of neutral-atom quantum computing, helping enterprise innovators leverage this technology to gain competitive advantage. With a public, peer-reviewed path to fault tolerance, QuEra operates globally from Boston, New Mexico, Tokyo, Zurich, and the United Kingdom. As quantum computing moves from ‘one day’ to ‘Day One,’ QuEra delivers practical impact today while leading the way toward large-scale, fault-tolerant systems.

QuEra’s President, Takuya Kitagawa, emphasized that even for experienced companies like his own, maintaining these machines at peak performance is a significant challenge. He highlighted the importance of this breakthrough: ‘A customer expects the entire computer to hold itself together without a specialist in the room.’ This achievement makes it far easier and cheaper to keep QuEra’s computers running at their best.

The results from the MHS research preview, combined with Anthropic’s frontier AI models, are significant. They demonstrate that AI can be used not only for automation but also for improving performance and efficiency. As quantum computing continues to advance, we can expect more innovations like this one to emerge.

Home furnishings retailer Williams-Sonoma is leveraging artificial intelligence (AI) to enhance its customer service model, according to CEO Laura Alber. The company has been using AI assistants to improve the checkout experience, product discovery, and personalization across its brand portfolio.

In a recent earnings call, Alber highlighted the importance of customer service in driving business success. Despite a volatile macroeconomic backdrop, Williams-Sonoma delivered another quarter of profitability, with comparable sales rising 6.2% year over year and total revenue growing 6.7% to nearly $2 billion.

Williams-Sonoma’s AI-powered shopping assistants, Olive and Otto, are designed to provide customers with personalized product recommendations and enhance their overall experience. Alber noted that both assistants have increased consumer engagement with the company’s content, design tools, and free design service offerings.

The launch of Otto across the Pottery Barn family in August marked a significant milestone for Williams-Sonoma’s AI efforts. The assistant is equipped to help customers narrow down product options based on their specific needs and preferences, much like human sales associates. According to Sameer Hassan, chief technology and digital officer, over 70% of customer engagements with Otto are resolved without the need for human intervention.

Williams-Sonoma’s confidence in its AI tools stems from its extensive reserves of first-party data. The company has been collecting rich data on customer behavior and preferences through its proprietary tech platforms and decades of expertise in home furnishings. Hassan emphasized that this data is a key differentiator, allowing Williams-Sonoma to drive impactful results with its AI-powered assistants.

One notable example of the success of Williams-Sonoma’s AI efforts is the performance of Olive at Williams Sonoma. Customers who use Olive convert at three times the rate of those who do not. The company has taken a similar approach with Otto, which Hassan described as ‘helping you narrow down the right piece for your space.’

Early results from Otto are promising, with customers responding positively to its personalized product recommendations and design tools. Williams-Sonoma’s ability to scale personalization using AI has also yielded significant revenue gains. According to Hassan, a visit where the experience is personalized generates roughly nine times the revenue of an average visit, up from two times last year.

Alber attributed Williams-Sonoma’s success in part to its strong brands and focus on customer service. The company believes that its execution and commitment to delivering exceptional experiences will continue to drive business growth in 2026 and beyond.

Instagram’s fast-paced environment demands consistent content creation, but the manual process can be time-consuming and exhausting. Brands and creators are turning to artificial intelligence (AI) tools to automate repetitive tasks and focus on strategy and engagement. An AI Instagram post generator is a tool that uses machine learning algorithms to create ready-to-publish content, including captions, hashtags, and visuals, based on your brand’s tone, industry, or prompt.

The traditional workflow of brainstorming ideas separately, designing graphics, resizing them for different formats, and manually scheduling everything can be frustrating. Each step adds friction, and the real cost isn’t just time – it’s consistency. When content creation feels like a chore, posting frequency drops, and Instagram’s algorithm rewards consistency over sporadic bursts of activity.

An AI Instagram post generator changes this equation by automating repetitive tasks, freeing up time for strategy, community engagement, and creative decisions that drive growth. By analyzing your brand voice, visual style, and audience signals automatically, these tools produce content tailored to Instagram’s format and audience behavior.

Not all AI content tools are created equal. When evaluating options, look for features like brand voice recognition, platform-specific formatting, hashtag intelligence, content variety, scheduling, and approval workflows. A good tool should analyze your existing content to match your tone rather than producing generic copy.

The best AI Instagram post generators have a core problem-solving approach: too many tools, too much manual work, and too little consistency. At the center of these tools is an AI engine that powers content generation across multiple platforms, including Instagram. This engine reads your website URL within seconds, builds a brand profile capturing your tone, vocabulary, and color palette, and generates a full month’s worth of content.

A strong tool auto-crops images to correct dimensions, builds hashtag stacks tailored to your niche, and writes Reels hooks designed to stop the scroll. Everything stays editable too, so you’re never stuck with content that doesn’t feel right – you can adjust captions, hashtags, visuals, or publishing times before anything goes live.

The most useful tools aren’t limited to Instagram alone; they adapt for other platforms like Facebook, LinkedIn, X, TikTok, Pinterest, YouTube, and Google Business. This ensures your messaging stays consistent everywhere without redoing the work multiple times over.

Getting started with an AI Instagram post generator typically takes just a few minutes. Connect your brand by pasting your website URL or describing it in a few sentences if you don’t have one. Let the AI analyze your brand, capturing your tone, visual style, and audience signals automatically. Review generated posts, edit anything that needs a personal touch, preview before publishing, and schedule or auto-publish.

AI can do the heavy lifting, but a few habits will help you get noticeably better results from these tools: feed accurate brand information, review content before publishing, mix content formats to keep your feed visually engaging, post consistently rather than sporadically, track what performs, and adjust accordingly. Publishing without edits is common – a quick human review still helps catch tone mismatches or irrelevant hashtags.

A few missteps can limit results even with capable tools: using a tool that isn’t platform-aware, publishing content without any edits, and avoiding tools that generate content in isolation from scheduling and analytics. This just shifts the busywork elsewhere instead of removing it.

An AI Instagram post generator is not about replacing creativity – it’s about streamlining repetitive tasks so you can focus on strategy, community engagement, and growth. Whether you’re a solo creator or a social media manager juggling multiple clients, these tools make it possible to go from a blank page to a full month of on-brand content in minutes rather than days.

Small businesses often don’t have the time or budget for a full content team, but AI-generated image and caption tools can produce high-quality content quickly. Basic generators might sound generic, but strong ones analyze your website and existing content to match your actual brand voice. Most quality tools let you adjust captions, hashtags, visuals, and publishing times before anything goes live.

Many complete platforms also offer scheduling features that plan a content calendar, get team or client approvals, and auto-publish at optimal times. This integration of AI-generated content with scheduling and analytics can significantly boost results for businesses and creators alike.

A major expansion of their long-standing partnership has been announced by Amazon Web Services (AWS) and NVIDIA. The two tech giants are committing to deploy an additional 2 million NVIDIA graphics processing units (GPUs) across AWS’s global infrastructure by 2028, significantly boosting the capabilities for machine learning jobs and AI development.

The move marks a significant milestone in the companies’ 16-year collaboration, which has already led to numerous innovations in the field of artificial intelligence. The partnership extends beyond just GPUs, with NVIDIA Vera CPUs being integrated into AWS’s platform as well. This will provide customers with a new compute option specifically designed for agentic AI workloads.

The massive influx of hardware signals a new phase in the race to build more powerful AI capabilities. To achieve this goal, the two companies are deepening their work across various areas, including AI factories, CPUs, networking, and robotics. Their aim is to deliver co-engineered solutions that can accelerate AI development at an unprecedented scale.

The deployment of 2 million additional NVIDIA GPUs will be a crucial factor in achieving this goal. This move complements Amazon’s existing custom silicon efforts, offering customers a wider choice for their AI workloads. Whether it’s NVIDIA’s specialized hardware or AWS’s Trainium chips, users can now select the best option to suit their needs.

The partnership also includes deploying 100,000 NVIDIA GPUs on secure AWS infrastructure specifically designed for federal and national security workloads classified at Impact Level 6 (IL6) and above. This initiative addresses global demand and competition in AI development, ensuring that sensitive information is handled securely.

By expanding their collaboration, AWS and NVIDIA are taking a significant step forward in the field of artificial intelligence. Their commitment to delivering co-engineered solutions will undoubtedly accelerate AI development at an unprecedented scale. As the companies continue to push the boundaries of what’s possible with AI, it remains to be seen how this partnership will shape the future of machine learning jobs and beyond.

San Francisco-based Salesforce and AI pioneer Anthropic have announced a groundbreaking partnership, dubbed Claudeforce. This expanded collaboration brings together the world’s #1 AI with the #1 AI-powered Customer Relationship Management (CRM) platform to revolutionize enterprise decision-making.

The Claudeforce partnership marks a significant milestone in the evolution of CRM technology. By integrating Claude’s advanced reasoning capabilities with Salesforce’s trusted data, workflows, and business logic, companies can now make informed decisions directly within their CRM system.

According to Marc Benioff, Chair & CEO of Salesforce, ‘We’re bringing together the world’s #1 AI and #1 CRM – the best of both worlds. Today, we’re launching Claudeforce, running directly on top of Salesforce via our new AIforce UI harness, Headless 360, Data 360, Tableau, and Slack.’

Benioff emphasizes that this integration is not just about combining two powerful technologies but creating a dynamic interface that thinks, reasons, and acts. ‘Probabilistic intelligence alone doesn’t run a company, and deterministic systems don’t reason,’ he notes.

Dario Amodei, CEO & Co-Founder of Anthropic, highlights the importance of trust in AI decision-making. ‘We believe frontier intelligence should be safe, trusted, and deeply capable, which is why the world’s leading businesses turn to Claude for their most important work.’

The Claudeforce partnership brings together two industry leaders with a shared vision of creating a more efficient and effective enterprise ecosystem. By integrating Claude’s AI capabilities with Salesforce’s CRM platform, companies can automate routine tasks, streamline workflows, and make data-driven decisions.

One key aspect of the Claudeforce integration is the ability to access Salesforce data directly within Claude. This allows users to reason over live revenue context, automate pipeline updates, and take governed action right from Claude – all with the power of Salesforce at their fingertips.

The partnership also includes a range of prebuilt sales skills that enable sellers and agents to work more efficiently. These skills were developed by Salesforce and Anthropic in collaboration, leveraging Claude’s advanced reasoning capabilities and agentic tool use.

Salesforce in Claude is made possible by AIforce – the trusted enterprise harness that brings all business data and workflows together through MCP servers, APIs, and CLI tools. This integration eliminates the need for complicated and costly integrations, making it easier than ever to access critical information.

The Claudeforce partnership also includes a range of additional features and capabilities. For example, Claude is available in Agentforce – serving as a reasoning model for the Atlas Reasoning Engine, powering Agentforce Vibes and Agentforce Coworker by default.

Through Amazon Bedrock, Claude is accessible within the Salesforce Trust Boundary, allowing customers to deploy domain-specific AI while keeping data and AI workloads secure. This ensures that companies can leverage the power of AI without compromising their security or compliance requirements.

The integration with Slack brings people and agents together in the flow of work. By embedding Claude directly into Slack – the central workspace where teams already operate – Salesforce is bridging the gap between team dialogue and autonomous enterprise action.

Claude serves as the default model for Slack, powering Slackbot to supercharge personal productivity by default, augmenting team decision-making via Claude Tag, and accelerating multiplayer coding. This integration enables teams to reason through complex decisions alongside Claude in Slack and instantly execute governed actions in Salesforce – all without leaving the flow of work.

The Claudeforce partnership is built on mutual adoption, with both companies serving as strategic customers for each other. This reciprocal relationship makes each company a builder and customer of the hybrid architecture they are bringing to market together.

Salesforce Shield, Backup, and Sandboxes support security, resilience, and development within the Salesforce ecosystem. Building on this foundation, new capabilities will power stronger scale across customer engagement – including headless experiences, connected data, and automation.

Claude is also being used as the AI model powering Slackbot for Salesforce customers and employees. Internally, Slackbot has transformed how Salesforce employees work, driving 8.1 million hours of annualized productivity gains – up over two times quarter-over-quarter.

The Claudeforce partnership marks a significant milestone in the evolution of CRM technology. By integrating Claude’s advanced reasoning capabilities with Salesforce’s trusted data and workflows, companies can now make informed decisions directly within their CRM system – automating routine tasks, streamlining workflows, and making data-driven decisions.

Availability for select pilot customers is already underway, with an open beta launch expected in September 2026. Additional prebuilt skills will begin launching later this year.

Phrasly, an AI-powered writing tool provider, has introduced Phrasly Docs, a comprehensive document editor designed to streamline the writing process for students and professionals. This innovative platform combines various features into one workspace, eliminating the need to switch between multiple tools throughout the drafting and editing stages.

The smart workspace in Phrasly Docs includes writing, humanizing, editing, an AI writing assistant, plagiarism detection, citations, and real-time collaboration. With this integrated approach, writers can efficiently complete their work without losing context or time switching between different applications.

Traditionally, long-form writing has required multiple disjointed tools. Writers often compose in one tool, verify tone or plagiarism in another, cite sources in a third, and share documents for feedback in a fourth. This fragmentation not only wastes time but also leads to lost context that must be recreated each time the writer switches between tools.

Phrasly Docs addresses this issue by incorporating drafting, review, research, and collaboration into a single editor. By doing so, writers can go from a blank page to a completed, cited document without leaving their workspace. This streamlined approach eliminates the need for multiple point tools, reducing clutter in the writing stack.

The AI Document Editor includes four categories of features that span the entire document lifecycle: Writing, Drafting, and Review; Research, Formatting, and Output; Collaboration; and Humanization. These categories are designed to guide writers through each stage of the process, ensuring a smooth workflow from start to finish.

Within these categories, Phrasly Docs offers various tools and features, including Tab autocompletes, AI Chat Agent with web search built-in, AI Commands with Detect AI and Summarize presets as well as custom AI Commands, an outline builder for structuring long documents, a humanizer that refines AI-assisted sections to sound natural, and a built-in AI detector providing an AI likelihood score and highlighting passages likely written by AI.

The Research, Formatting, and Output category includes features such as automatic citation generation, which scrapes millions of academic and web sources. The platform also supports over 30 languages, making it suitable for multilingual teams working on content across multiple markets. Additionally, the editor allows writers to export their work directly to Word or PDF format.

The workflow in Phrasly Docs is straightforward: Write from scratch or an empty document; use Tab autocomplete and AI Commands to write and edit sentences in place; humanize and proofread the draft by running it through the built-in AI detector; cite and export by inserting formatted citations. This four-step process enables writers to complete their work efficiently without losing context.

One of the key benefits of Phrasly Docs is its ability to make AI-assisted writing sound more natural. The platform’s humanizer rewrites AI-influenced passages so they resemble text written by a person, rather than a machine. This feature is particularly important for students and professionals who need their work to be free from red flags associated with AI-generated content.

Phrasly Docs has been designed specifically for students writing papers and professionals turning in work that will be seen by clients. By addressing the issue of AI-generated text sounding unnatural, Phrasly Docs helps writers maintain trust and credibility in their work.

The platform’s humanizer works seamlessly within the editor, allowing writers to resolve issues related to detection and humanization prior to submitting their work. This eliminates the need for additional tools or starting from scratch, saving time and effort throughout the writing process.

For students, Phrasly Docs offers a comprehensive AI Writing Tool that guides them through an essay or research paper from initial prompt to final draft. The platform’s automatic citation generator scrapes millions of academic and web sources, making it easy to insert formatted citations in various styles while still on the page.

Phrasly Docs also supports professionals and content teams by providing a collaborative workspace for drafting articles, reports, and client documents. Tab autocompletes and AI Commands help speed along first drafts, while real-time document collaboration, auto save, and offline sync enable colleagues to work together without losing progress.

The benefits of using Phrasly Docs are clear: fewer tools in the stack, fewer handoffs between those tools, and less context lost when moving a doc from draft to delivery. By streamlining the writing process, Phrasly Docs helps writers complete their tasks more efficiently, saving time and effort throughout the workflow.

Phrasly is an established provider of AI-powered writing tools, with over 3 million users who have completed over 4 million writing projects and saved approximately 600,000 hours of writing. The company’s suite of tools includes Phrasly Docs, as well as other features such as AI Humanizer, AI Detector, Content Generator, Thesis Generator, Translator, and Plagiarism Checker.

Hospitals are increasingly turning to AI tools to ease administrative burdens and improve patient care. One area where these tools make a significant impact is in billing, coding, and documentation.

However, some unsubstantiated claims have suggested that the use of AI tools has increased coding intensity, leading to higher healthcare costs. These allegations ignore underlying drivers of changes in coding intensity, including an aging population and rising prevalence of chronic disease.

Hospitals are caring for patients with higher acuity due to demographic shifts and advances in medicine. The AHA analysis found that 19% of hospital expense growth from 2019 to 2024 reflects caring for sicker, more complex patients. This trend is also reflected in the rising case-mix index – a standard measure of patient acuity.

The shift of lower-acuity care to outpatient settings has left hospitals with an increasingly complex and resource-intensive inpatient population. AI tools can support accurate coding and documentation by enhancing precision within established frameworks, but human validation remains essential for ensuring coding integrity and regulatory compliance.

Assertions that providers are upcoding due to the use of AI tools are striking given evidence of insurer-driven coding practices. Lawsuits, investigations, and reports from MedPAC, DOJ, and Congress have documented upcoding among large insurers, including adding diagnosis codes not supported by clinical care.

In 2025, MedPAC reported that upcoding contributed to $40 billion in overpayments to Medicare Advantage plans. One payer settled a lawsuit with the DOJ on upcoding allegations in March 2026, while another insurer is facing a state-level investigation for potentially fraudulent practices.

The AHA has urged policymakers to prevent insurers from systematically reducing reimbursement under the pretense of widespread inappropriate upcoding. This would help alleviate administrative waste and burden created by overly burdensome appeals processes.

AI tools are transforming care delivery in countless ways, supporting increased access, improved outcomes, and reduced costs. They offer tremendous potential for reducing more than $1 trillion spent annually on administrative functions. However, their use is not without controversy – particularly when it comes to coding intensity.

Some commercial insurers have asserted that AI tools used by providers are resulting in higher coding intensity. These accusations ignore multiple factors contributing to growth in coding intensity, including an aging population and increasing prevalence of chronic disease.

Hospitals maintain rigorous auditing and coding compliance programs to ensure they capture patient acuity levels accurately and meet regulatory requirements. Coding has always been governed by established guidelines, conventions, and expectations that define a compliant approach to code assignment.

Updates to evaluation and management coding guidelines have altered coding patterns, leading to more specific diagnoses being reported for higher-acuity patients. As providers adapt and data stabilize, these changes contribute to shifts in coding intensity.

The AHA has responded to concerns about AI tools used by providers, emphasizing the need for accurate representation of patient complexity. They argue that hospitals are not engaging in upcoding practices but rather accurately capturing acuity levels due to demographic shifts and advances in medicine.

In contrast, some insurers have been accused of using automated edits to unilaterally reduce reimbursement without reviewing medical documentation – a practice known as downcoding or partial denial. This can lead to burdensome appeals processes for providers seeking reimbursement for medically necessary care.

The AHA has urged policymakers to prevent such practices and ensure that hospitals are fairly reimbursed for the care they provide. They argue that accurate coding is critical for ensuring patient safety, quality of care, and regulatory compliance.

AI tools have tremendous potential for supporting data analysis in healthcare – from streamlining administrative tasks to improving patient outcomes. However, their use must be carefully managed to ensure accuracy, consistency, and fairness in reimbursement practices.