The Answer Engine AI News · Sydney

Anthropic has released a report accusing several Chinese AI companies—including Alibaba, Moonshot AI, and DeepSeek—of conducting "distillation attacks" against its Claude AI models. Distillation is a technique where one company repeatedly queries another's AI model to extract and copy its capabilities into their own cheaper model, essentially stealing the intellectual property embedded in the original system. According to Anthropic, these attacks have become more frequent and persistent in recent months as competition in the AI sector has intensified. The practice allows companies to shortcut the expensive process of training their own models from scratch by harvesting the knowledge from established systems. While the technical details weren't fully disclosed in the source summary, Anthropic's public report suggests the company views this as a significant threat to its business and the broader AI industry. The allegation highlights growing tensions between Western and Chinese AI developers, and raises questions about intellectual property protection in an industry where the "product" is the model's learned capabilities rather than traditional code or patents.

Meta has launched Muse, a new AI assistant app that's quickly climbed to become the second-most downloaded app in the United States. This contradicts the original report's claim of a "slower start" — reaching number two is actually a strong performance. Muse joins Meta's growing suite of AI products, which already includes Meta AI (integrated into Facebook, Instagram, and WhatsApp) and the social network Threads. The app's rapid adoption suggests strong consumer interest in standalone AI assistants, putting it in direct competition with ChatGPT, Google's Gemini app, and other AI chatbots. While specific features weren't detailed in the source material, Muse represents Meta's continued push into the AI space, creating yet another channel where the company's AI technology interacts with users. For context, Meta has been aggressively rolling out AI features across its platforms over the past year, and Muse appears to be the latest piece in that strategy — a dedicated app rather than an embedded feature.

Virginia's data center hub has suffered two major power failures — the most recent in July 2026 knocked 3 gigawatts offline in seconds after a transmission line fault in Ashburn, home to the world's largest data center cluster. Two years earlier, a single failed surge arrester dropped 1,500 megawatts and took down roughly 60 facilities at once. These aren't isolated incidents. The infrastructure built to power AI is bumping into serious limits. As AI training and inference demand enormous amounts of electricity in concentrated geographic areas, the grid simply wasn't designed to handle the load. Data centers are clustering together for connectivity and cost reasons, but that concentration creates single points of failure. When one component fails, the cascading effect can be massive. The article frames this as an architecture problem — not just building bigger power plants, but fundamentally rethinking how and where we distribute the infrastructure needed to keep AI systems running reliably.

Major AI companies are moving into healthcare with increasingly sophisticated systems that can read lengthy clinical records, understand medical jargon, cross-check documentation against scientific evidence, and produce useful summaries from massive amounts of information. The technical capabilities are impressive and advancing quickly. However, the real challenge now isn't whether the AI works—it's whether healthcare organisations can actually integrate these tools into their existing workflows, systems, and day-to-day operations. Clinicians, hospital administrators, and support staff need these AI systems to fit seamlessly into the way they already work, not force them to completely redesign their processes. This integration hurdle involves practical issues like connecting with legacy software, training staff, managing patient privacy and data security, and proving the technology delivers measurable improvements in patient care or operational efficiency. The article suggests that technical progress has outpaced the healthcare industry's ability to adopt it, and the next phase will be about making AI genuinely useful in real clinical settings rather than just technically impressive in labs.

Researchers have released OpenDiscoveryTrace, a public dataset containing 558 complete records of how AI models actually work through scientific problems, not just their final answers. Think of it as showing the working, not just the answer at the bottom of the page. Each record captures nine types of information per step: what the AI was thinking, which tools it used, what it observed, errors it made, when it changed course, and how confident it felt. The dataset covers 124 scientific tasks across drug discovery, materials science, genomics, and literature analysis, tested on seven different AI models including GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro. Until now, benchmarks for AI scientist tools only looked at final outputs—the code, hypotheses, or papers produced—which made it impossible to understand whether the AI used sound methodology or just got lucky. This dataset lets researchers audit the reasoning process, diagnose where things go wrong, and distinguish between systematic thinking and fortunate guessing. It's a tool for making AI scientific assistants more reliable and transparent.

Search Engine Journal is promoting a webinar scheduled for 24 September featuring Google and Uberall, focused on upcoming changes to Google Search and local marketing strategies. The session promises to reveal what's coming next in Google Search and will share five specific fixes to local marketing strategies designed to help business locations get recommended by Google's systems. While the source material is promotional and light on detail, the webinar appears aimed at local businesses concerned about how Google's AI-driven search updates will affect their visibility. The timing suggests Google is preparing to roll out changes to how it surfaces and recommends local businesses in search results, likely tied to its ongoing integration of AI features like Search Generative Experience (SGE) and AI Overviews. For local businesses relying on Google to drive foot traffic and inquiries, understanding these changes early could be valuable, though the actual substance won't be clear until the webinar itself.

Publishers are making critical decisions about blocking AI crawlers based on incomplete Google Analytics data that may significantly undercount AI-driven traffic. The commonly cited metric showing declining referral traffic from Google doesn't account for an unknown portion of visits that arrive without proper referral tags – meaning the "denominator" in the calculation is wrong. When AI overviews, ChatGPT, or other AI tools send users to websites, those visits often don't register as coming from AI sources in standard analytics. This creates a dangerous situation: publishers see overall referral numbers dropping and assume AI is cannibalising their traffic, leading them to block AI crawlers with robots.txt files. But they're making these blocking decisions based on faulty data that may not reflect the true picture of where their traffic is actually coming from. The article highlights how the industry is reaching different conclusions while quoting the same statistics, simply because the underlying measurement is fundamentally flawed.

AEO_ANTE: Before you block AI crawlers to "protect" your traffic, understand that your analytics probably aren't showing you the full picture of visits coming from AI tools – you might be blocking sources that are actually sending you customers.

Moz executives Jonathan Berthold and Kevin Chen have outlined the key factors that determine how AI search tools recommend local businesses. According to their analysis, four main elements shape whether your business appears in AI-generated recommendations: business listings (such as Google Business Profile), customer reviews, online mentions across the web, and social media activity. Unlike traditional search engine optimisation that focused primarily on website content and links, AI search engines pull from a broader range of signals to decide which local businesses to suggest. The Moz leaders explain that AI assistants like ChatGPT, Perplexity, and Google's AI Overviews use these signals to build their understanding of which businesses are credible, relevant, and worth recommending to users asking local questions. This means local businesses need to think beyond their website and ensure they're maintaining accurate listings, actively collecting reviews, building mentions in local media and directories, and staying visible on social platforms where their customers are active.

IBM and NASA have developed a new artificial intelligence model designed to map ice deposits and craters on the Moon's surface. The collaboration combines IBM's AI expertise with NASA's vast repository of lunar data to create a tool that can analyse satellite imagery and identify features that are critical for future lunar missions. Mapping ice is particularly important because water ice could provide drinking water, oxygen, and even rocket fuel for astronauts on extended Moon missions. The AI model can process and interpret lunar surface data much faster than manual analysis, potentially accelerating mission planning and site selection for upcoming Artemis program landings. This project is part of a broader trend of space agencies using machine learning to handle the enormous volumes of data generated by modern satellites and space probes. While the immediate application is lunar exploration, the underlying AI techniques could eventually be adapted for Earth-based mapping applications, including agriculture, disaster response, and resource management.

The University of New South Wales is rolling out ChatGPT Edu to 80,000 students, staff and researchers while simultaneously introducing new measures to tackle AI-assisted cheating. The move reflects the tension universities face: embracing AI tools for legitimate teaching and research while preventing students from using them to cheat on assignments and exams. ChatGPT Edu is OpenAI's university-focused subscription that gives institutions controlled access to ChatGPT, including GPT-4, with enhanced data privacy protections. UNSW's approach suggests they're trying to work with AI rather than ban it outright, but they're putting guardrails in place to maintain academic integrity. The specific anti-cheating measures weren't detailed in the summary, but the announcement indicates UNSW is developing policies and detection methods alongside the technology rollout. This follows a broader pattern across Australian universities grappling with generative AI in education—trying to teach students to use these tools responsibly while ensuring assessment still measures genuine understanding and capability.