Artificial intelligence is moving beyond content generation and into systems that influence education, software operations, scientific research and corporate strategy.
Scientists have used genomic AI models to design functioning bacteriophages that do not exist in nature, opening a potential route towards new treatments while intensifying biosecurity concerns. Engineering teams are introducing AI into incident response, but automation could weaken the human expertise needed during exceptional failures. Alibaba is reportedly exploring a revenue-sharing model for major commercial users of Qwen, while Google is restructuring its AI leadership as Gemini passes 950 million monthly users.
Education systems are responding in different ways. Kazakhstan is learning from China’s deployment of AI in schools, whereas Denmark is tightening assessment controls to counter AI-assisted cheating.
AI-designed viruses open a new frontier in synthetic biology
Researchers have used generative AI to design complete viral genomes, producing 16 functioning bacteriophages that were not found in nature.
Bacteriophages are viruses that infect bacteria rather than people, animals or plants. They are being investigated as potential treatments for bacterial infections, particularly as resistance reduces the effectiveness of conventional antibiotics.
According to Mezha, researchers used genomic language models to generate thousands of potential viral genomes. The scientific team, led by Stanford University researcher Brian Hie, then selected nearly 300 designs for laboratory synthesis.
Sixteen produced viable bacteriophages. A combination of the resulting viruses successfully killed strains of E. coli that had developed resistance to naturally occurring bacteriophages.
The research used Evo1 and Evo2, genomic models trained on genetic information from approximately two million bacteriophages. Genetic material from viruses capable of infecting humans, animals or plants was intentionally excluded from the training data.
The immediate medical potential is significant. Conventional phage therapy often depends on finding appropriate viruses in natural environments and then determining whether they can target a patient’s bacterial infection. AI could make it possible to design phages for particular bacterial strains, modify their characteristics and respond more quickly when bacteria develop resistance.
The experiment was nevertheless inefficient. Only 16 of the thousands of generated designs became functioning viruses, and bacteriophages possess comparatively small genomes. Designing more complicated biological systems would be substantially harder.
The importance of the result lies less in its current efficiency than in what it demonstrates: an AI system can generate a complete genome that becomes a functioning, self-replicating biological entity after laboratory synthesis.
That capability raises obvious dual-use concerns. A technology developed for treating antibiotic-resistant infections could eventually be adapted to more dangerous organisms. The researchers themselves called for safety, security and biocontainment specialists to be involved throughout projects involving whole-genome design.
Oversight will need to extend across the complete development chain:
- Training datasets should exclude genetic information whose inclusion creates unjustifiable risks.
- Access to high-capability biological models may require additional controls.
- Research proposals should undergo independent biosafety and biosecurity review.
- DNA-synthesis companies need reliable screening for hazardous sequences.
- Laboratories must maintain containment appropriate to the organisms being produced.
- Generated designs and experimental outcomes should be auditable.
Regulating only the model would leave important gaps. Even if a model produces a potentially hazardous sequence, physical creation still depends on DNA synthesis and laboratory processes. Conversely, synthesis screening alone may be insufficient if AI-generated genomes differ substantially from known pathogens and evade conventional sequence-matching systems.
The experiment offers a powerful example of why AI governance must become domain-specific. A general chatbot policy cannot adequately govern a model capable of designing biological material.
AI changes incident response but increases the value of human expertise
AI is also becoming an active participant in production incident management.
Engineering teams can use it to summarise communication channels, inspect unfamiliar code, identify possible causes, recommend remediation, generate pull requests and document what happened. These capabilities can reduce cognitive overload at precisely the moment when responders are dealing with fragmented information and intense time pressure.
However, an InfoQ analysis of AI in incident response argues that automating routine failures could make the remaining human work more difficult.
The report invokes the “Leftover Principle”: as automation takes responsibility for predictable tasks, people are left with the unusual, ambiguous and exceptional cases that automated systems cannot resolve.
This creates a skills problem. Engineers traditionally develop incident-response judgement by handling a mixture of ordinary and difficult failures. If AI resolves most routine incidents, human responders may receive fewer opportunities to practise diagnosis, communication and recovery.
When a rare, high-impact failure eventually exceeds the AI system’s abilities, responders may enter the incident with limited context and less recent hands-on experience.
AI recommendations can also change human performance in both directions. Correct guidance may help responders considerably, while plausible but incorrect recommendations can lead them away from the actual cause. Automation bias becomes particularly dangerous when an AI system communicates an uncertain conclusion with confidence.
Organisations introducing AI into incident response should therefore preserve human capability through:
- Regular game days and simulated production incidents.
- Tabletop exercises covering unfamiliar and cascading failures.
- Chaos-engineering programmes that test system and team resilience.
- Clear thresholds for escalating from automated handling to human control.
- Interfaces that communicate confidence, evidence and alternative hypotheses.
- Complete logs of the information, tools and actions used by the AI.
- Manual recovery procedures that remain usable when the AI platform is unavailable.
Accountability must also be explicit. Placing a human approver at the end of an automated workflow does not create meaningful oversight if that person lacks the context, time or authority to challenge the system.
The problem will become more pressing as AI accelerates software development. More generated code, pull requests and deployments mean more changes entering production. Even if each individual change is relatively safe, a large increase in deployment volume can still raise the total number of incidents.
Strong testing, observability, feature flags, staged deployments and rapid rollback consequently become more important in an AI-assisted engineering environment.
Alibaba tests the commercial limits of open-weight AI
Alibaba is reportedly preparing a revenue-sharing model for major commercial users of its next Qwen model, potentially changing the economics of open-weight AI.
Under the reported approach, smaller developers and researchers would continue to receive free access to downloadable model weights. Large businesses generating substantial revenue through Qwen deployments would be required to enter commercial agreements and return a proportion of that revenue to Alibaba.
The plan is associated with the next-generation Qwen3.8-Max model, although Alibaba had not publicly confirmed the proposed terms at the time of reporting. Artificial Intelligence News presents the development as an effort to preserve broad access while capturing value from the largest commercial deployments.
The approach would occupy a middle ground between unrestricted open-weight releases and conventional proprietary APIs.
Open-weight models have helped Alibaba expand Qwen’s international developer community. Organisations can download, customise and operate the models on their own infrastructure, avoiding continuous dependence on a vendor-controlled service.
That distribution strategy creates indirect commercial benefits for Alibaba, including demand for cloud infrastructure and greater influence over the AI development ecosystem. It does not, however, automatically recover the substantial cost of training increasingly large frontier models.
Revenue sharing could address part of that imbalance. Large model-as-a-service providers and enterprises would contribute when they build significant businesses on top of Qwen, while researchers and smaller developers would retain comparatively open access.
Implementation will be difficult. Alibaba would need to define which activities and revenue streams trigger payment. A company may use Qwen as one component inside a much larger product, making attribution complicated. Enforcement could also prove challenging when model weights are deployed privately or modified into derivative systems.
The licensing language will determine whether the model should meaningfully be described as open source. A licence that restricts particular commercial uses may provide accessible weights without meeting established definitions of open-source software.
The debate reflects a wider shift in AI. Model developers are discovering that openness is not only a technical or philosophical choice; it is also a question of how the enormous cost of compute, data, research and infrastructure will be sustained.
Google restructures DeepMind as Gemini reaches 950 million users
Google is reorganising its AI leadership as it moves Gemini further into consumer products, developer services and enterprise infrastructure.
In a message describing the company’s next chapter of AI momentum, Sundar Pichai announced that Demis Hassabis would step away from Google DeepMind’s daily operational management.
Hassabis will become chair of Google DeepMind and chief scientist of Alphabet. He will continue advising the company on models, research and longer-term artificial-general-intelligence strategy while remaining involved with Isomorphic Labs, Google’s AI-driven drug-discovery business.
Koray Kavukcuoglu, previously Google DeepMind’s chief technology officer and Google’s chief AI architect, will become senior vice-president of Google DeepMind. His responsibilities include Gemini model development, frontier research, the Gemini application and developer teams.
The transition separates long-range scientific strategy from the operational task of developing and commercialising models at Google’s scale.
Pichai said the Gemini application had passed 950 million monthly users. Google’s Gemma models had exceeded 900 million downloads, while Gemini was being integrated into Search through AI Mode and AI Overviews and deployed across Google’s wider product and cloud ecosystem.
Google is also supporting the departure of Jeff Dean and Sanjay Ghemawat, two of its most influential engineers, to create an independent public-benefit corporation focused on accelerating discoveries in machine learning, science and engineering. Google will participate as a founding investor and cloud partner.
The changes demonstrate how frontier-AI organisations are becoming too broad for a single leadership model. They must simultaneously manage foundational research, global consumer products, enterprise customers, safety evaluations, infrastructure investment and geopolitical scrutiny.
Hassabis’s move towards scientific strategy may allow him to focus on the consequences of increasingly capable models and on applications such as healthcare. Kavukcuoglu, meanwhile, inherits the immediate competitive challenge of improving Gemini and translating research into dependable products.
HIPTHER recently examined the importance of balancing capability with control in its AI Dispatch covering OpenAI, Hugging Face, Claude, DEFCON AI and Fish Audio.
Kazakhstan and China expand cooperation on AI in education
Around 50 principals from Kazakhstan’s “Keleshek Mektepteri”, or Future Schools, project attended an international seminar on educational AI in Urumqi, China.
According to Qazinform, participants studied education-management methods, digital systems, AI tools for learning and cooperation between schools and universities.
Kazakhstan plans to introduce selected practices gradually across Future Schools while continuing professional collaboration with Chinese educational institutions. The Ministry of Enlightenment identified teacher digital skills and the integration of international experience as priorities.
The emphasis on school leaders is important. Successful AI adoption depends on more than purchasing software. Principals must decide which uses support learning, how teachers will be trained, what student data can be processed and how AI-generated output will be evaluated.
Schools should establish clear distinctions between AI as a learning aid and AI as a substitute for student work. Productive applications may include personalised exercises, language support, lesson preparation and feedback. Risks include inaccurate content, excessive surveillance, dependence on external technology providers and the weakening of independent reasoning.
International exchange can accelerate adoption, but Kazakhstan will need to adapt Chinese practices to its own curriculum, languages, legal requirements and educational culture.
Denmark introduces oral defences to counter AI-assisted cheating
While Kazakhstan is exploring how AI can be integrated into education, Denmark is tightening controls around how students use it.
Under measures reported by Euronews, upper-secondary students will be required to defend certain written assignments orally. Schools will also monitor computer use during written examinations, install network restrictions and encourage more assignments to be completed under supervised conditions.
The rules concern students aged 16 to 19. Approximately 9,000 students each year complete the major written assignment known as the større skriftlige opgave, or SSO, which will now include an oral defence.
The oral component offers more than an anti-cheating mechanism. It allows teachers to determine whether students understand their arguments, evidence and methods. A student who has genuinely produced a piece of work should normally be able to explain its reasoning, defend its conclusions and discuss alternative interpretations.
Screen monitoring and firewalls may reduce unauthorised AI use during examinations, although those measures also raise privacy and proportionality questions. Schools will need transparent policies explaining what activity is observed, how monitoring data is stored and who can access it.
Denmark is developing a wider national strategy for AI in education. That broader work will be necessary because enforcement alone cannot resolve the problem. Students need to understand when AI assistance is permitted, how it should be disclosed and which intellectual tasks they remain responsible for performing themselves.
The bigger picture: adoption and restraint must develop together
The six developments demonstrate that AI governance cannot be reduced to a single set of universal rules.
Synthetic biology requires controls spanning training data, model access, DNA synthesis and laboratory containment. Incident response needs automation that preserves human expertise and accountability. Open-weight AI requires sustainable commercial models with transparent licence conditions. Schools must teach productive AI use while protecting genuine learning and fair assessment.
Google’s leadership changes illustrate the same pressure at corporate scale. AI organisations must accelerate products and research while giving greater attention to long-term scientific, safety and social consequences.
The next phase of artificial intelligence will be defined by more than what models can generate. It will depend on whether institutions can decide where automation adds value, where human competence must be protected and where a capability is consequential enough to require several independent layers of control.
