Artificial intelligence is spreading rapidly, but institutional readiness remains deeply uneven.
Most European Union countries have yet to integrate AI into their anti-fraud strategies. INTERPOL says the technology already enables more than half of reported cybercrime in Africa. Washington is bringing leading AI developers together to discuss pre-release cybersecurity testing, while the World Bank is warning developing economies that delayed adoption could widen existing economic divides.
At the same time, the geopolitical contest is expanding beyond chips towards the infrastructure required to produce affordable AI intelligence, and Akido is demonstrating what large-scale adoption could look like in healthcare.
Together, these developments show that access to AI is no longer the central issue. The decisive questions concern who can deploy it safely, adapt it to local needs and build institutions capable of managing its consequences.
Twenty-one EU countries omit AI from their anti-fraud strategies
Twenty-one of the European Union’s 27 member states do not use artificial intelligence in their national anti-fraud strategies, exposing a significant gap between the bloc’s regulatory ambitions and its operational adoption.
According to Informat.ro, only the Czech Republic, Italy and Portugal have explicitly incorporated AI into their anti-fraud strategies. Belgium, Spain, France, Italy, Hungary and Slovenia use the technology through other programmes or activities, meaning that eight countries employ AI-related tools in some form.
Potential applications include anomaly detection, predictive risk analysis, invoice verification, audit support and network analysis connecting individuals, companies and suspicious transactions.
These capabilities are particularly relevant when public authorities must inspect enormous volumes of procurement, grant, customs and payment data. Fraud rarely appears as a single unmistakable event. It is more likely to emerge through unusual combinations of transactions, ownership relationships, duplicate claims or deviations from established patterns.
AI can help investigators prioritise these signals, but adoption is being constrained by familiar institutional problems.
Nineteen member states identified interoperability as an obstacle, while 16 cited budget limitations and 12 raised legal or data-protection concerns. Only the Czech Republic, Luxembourg and Sweden had completed a comprehensive assessment of cybersecurity and AI-related risks to the EU budget during 2025.
Fifteen countries had conducted no such assessment and reported no plans to begin one. Eight had completed only a partial review, while one was preparing an assessment.
The figures illustrate why regulation alone does not create administrative capacity. Europe may establish demanding rules for commercial AI systems while its own institutions continue to rely on fragmented databases, manual processes and incompatible national platforms.
Use of Arachne, the EU’s risk-scoring and data-mining tool, increased from 22 member states in 2024 to 23 in 2025. However, integration is often limited or dependent on manual data transfers. A planned Arachne+ platform could improve interoperability, but its effectiveness will depend on data quality and consistent national participation.
Detected fraud affecting the EU budget reached approximately €274.3 million across reported 2025 cases, excluding certain value-added-tax and Recovery and Resilience Facility cases. Better analytics could improve detection, but automated risk scores cannot replace legal scrutiny or human judgement.
Authorities need to understand why a system flagged a transaction, verify the underlying evidence and provide a route for challenging incorrect decisions. Poorly governed automation could otherwise redirect investigative resources towards statistical anomalies rather than genuine wrongdoing.
This implementation challenge follows the broader regulatory transition examined in HIPTHER’s AI Dispatch covering the EU AI Act, transparency requirements and Chinese AI developers.
“Tokenpolitik” reframes the geopolitical competition over AI
The strategic contest between the United States and China is moving beyond control of advanced semiconductors towards the complete infrastructure required to produce and distribute AI intelligence.
An analysis discussed by Decode39 describes this emerging competition as “tokenpolitik”.
Tokens are the units of data processed by large AI models. Their cost depends on far more than the model itself. Semiconductors, data centres, cloud platforms, cooling systems, electricity grids, networking infrastructure and software efficiency all contribute to the price of generating an answer.
Controlling the chip supply consequently provides influence, but it does not guarantee leadership across the complete AI economy.
China is pursuing a full-stack strategy in which companies such as Huawei can offer international partners hardware, software, cloud services, training and technical support. This model may be especially attractive to countries that want AI capacity but lack the capital or expertise to assemble infrastructure from multiple suppliers.
The Global South therefore represents more than a commercial market. It is becoming a contested area of technological alignment.
Countries adopting a vertically integrated Chinese AI stack may gain access to affordable infrastructure, but they may also become dependent on Chinese standards, vendors and maintenance systems. Similar concerns apply when nations depend overwhelmingly on American cloud and model providers.
The proposed American response is a coordinated form of tokenpolitik involving the State, Commerce and Treasury departments, development-finance institutions, export-credit agencies, allied governments and private technology companies.
Rather than concentrating exclusively on restricting Chinese access to advanced chips, the US and its partners would offer third countries secure and economically viable alternatives across the AI stack.
This strategy is related to initiatives such as Pax Silica, which seeks to coordinate trusted supply chains and technology partnerships. Italy’s participation is significant because it connects European industrial capacity with a broader effort to shape AI infrastructure and governance.
The concept changes how AI power should be measured. The decisive advantage may not belong to the country with the single largest model. It may belong to the coalition that can produce useful tokens reliably, cheaply and under governance arrangements that international partners trust.
HIPTHER previously explored this competition between capability, openness and strategic control in its AI Dispatch on open-weight models, AI wealth and the emerging intelligence economy.
INTERPOL says AI enables 55% of reported cybercrime in Africa
Artificial intelligence is now enabling 55% of reported cybercrime across Africa, according to findings highlighted by Africanews.
The assessment draws on information from 36 INTERPOL member countries and describes a threat environment being transformed by automated social engineering, synthetic identities, credential harvesting and highly scalable scams.
Estimated cybercrime losses more than doubled from $192 million in 2024 to $484 million. The expansion of mobile connectivity and digital financial services has created valuable opportunities for households and businesses, but it has also provided criminals with a much larger pool of potential victims.
The threats vary across the continent.
East African countries face substantial mobile-money fraud and ransomware activity. Business-email compromise and romance scams are prominent in Central and West Africa, while the comparatively connected economies of Southern Africa provide attractive targets for sophisticated campaigns.
Seventy-two per cent of the surveyed countries reported the presence of scam centres, with the highest concentration found in Southern and West Africa.
Generative AI makes these operations more convincing and less expensive. Criminal groups can produce grammatically accurate messages in several languages, imitate familiar communication styles and personalise fraudulent approaches using information collected from social media or breached databases.
Synthetic images, audio and video can strengthen impersonation attempts. AI-generated identity documents and manipulated biometric material may also be used to open accounts, obtain loans or register mobile SIM cards.
Business-email compromise becomes particularly dangerous when attackers can reproduce the tone and terminology of a senior executive. Employees may receive realistic instructions requesting an urgent payment, a change of banking details or the disclosure of confidential information.
INTERPOL-backed operations have resulted in more than 1,500 arrests and the recovery of over $100 million, but enforcement alone cannot match the scale at which automated scams can be created.
Governments, telecommunications providers, banks and digital platforms need faster mechanisms for sharing indicators and freezing fraudulent transactions. Countries also require consistent digital-forensics standards, better public reporting and investigators trained to handle AI-generated evidence.
Public AI literacy has become a cybersecurity control. People should be encouraged to verify unexpected payment instructions through a separate communication channel and to treat convincing audio or video as evidence requiring confirmation rather than automatic trust.
Washington brings frontier AI developers together for cybersecurity testing
The White House is organising a meeting with OpenAI, Anthropic and Google to discuss a voluntary framework for testing the cybersecurity capabilities of advanced AI models.
As Informat.ro reports, participating companies could provide the US government with access to certain models for as long as 30 days before their public release.
Testing would examine whether a system could identify vulnerabilities, assist with sophisticated intrusion activity or substantially lower the expertise required to conduct advanced cyberattacks.
The initiative is expected to involve the Treasury Department, National Security Agency and Cybersecurity and Infrastructure Security Agency. Some evaluation criteria would remain classified to avoid disclosing sensitive defensive capabilities or exploitable weaknesses.
Pre-release access could help authorities identify risks before a highly capable model reaches millions of users. It may also support more consistent evaluation by allowing several developers to be tested against comparable scenarios.
However, a voluntary framework creates unavoidable limitations.
Companies may interpret risk thresholds differently, provide access under different technical conditions or decide that commercial confidentiality prevents sufficiently detailed disclosure. The government must also ensure that its own testing environment cannot leak models, evaluation methods or newly discovered vulnerabilities.
Cybersecurity evaluation should consequently be treated as a continuing operational process rather than a one-time certification.
A model that appears safe in a controlled benchmark may behave differently when connected to browsing, coding, command-line or enterprise tools. Risks can also change after fine-tuning, system-prompt revisions or integration into an autonomous agent.
This is why runtime governance remains important after deployment. HIPTHER examined that requirement through OpenBox AI and Temporal’s governance infrastructure for long-running AI agents.
World Bank warns developing economies against delaying AI adoption
Developing countries should embrace artificial intelligence or risk falling further behind wealthier economies, the World Bank has warned.
Its World Development Report 2026, covered by France 24, argues that AI offers a rare opportunity to accelerate productivity and expand access to essential services.
The report proposes an “adopt, adapt and advance” framework.
Most countries do not need to build frontier models requiring enormous computing budgets. They can adopt existing tools, adapt them to local languages and institutions, and gradually advance their own technical capabilities.
Importing a model is not enough. Systems designed around wealthy economies may perform poorly when applied to different languages, health systems, legal frameworks or educational environments.
Countries therefore need reliable electricity, connectivity, interoperable digital infrastructure, useful local data and people capable of evaluating model output.
The greatest opportunities may come from relatively modest applications rather than spectacular general-purpose systems. AI delivered through text messages, voice calls or basic mobile devices could support medical screening, weather advice for farmers and lesson planning for teachers, including in regions with limited connectivity.
Adoption is already moving faster than earlier technological transitions. Middle-income economies reportedly accounted for approximately half of ChatGPT traffic within six months of its release.
Speed does not guarantee equal benefit, however. Countries that use imported AI without building local expertise may exchange one development gap for another. They could become permanently dependent on foreign cloud infrastructure, external datasets and models they cannot meaningfully inspect.
Public procurement, education and competition policy will be as important as technology investment. Governments should avoid locking critical services into one provider and should require evidence that systems work under local conditions.
The World Bank also expects AI initially to complement many workers in developing economies because a larger proportion of employment involves physical or interpersonal tasks that cannot be completely automated.
That creates time for adaptation—but not an excuse for delay.
Akido brings AI-supported intake and diagnosis to nearly 100 clinics
Healthcare provider Akido is expanding its clinical AI technology across approximately 100 clinics, moving the discussion from experimental medical models towards routine patient care.
According to Endpoints News, the system supports patient intake and assists clinicians with diagnostic work.
Healthcare offers a compelling case for AI because much of the administrative burden occurs before or around the clinical consultation. Patients must describe symptoms, complete forms and provide medical histories, while clinicians must review records, document encounters and consider possible diagnoses.
AI can organise this information before an appointment, identify missing details and help clinicians explore relevant possibilities. Used carefully, it could allow medical professionals to spend more time interacting with patients.
Scaling across a network of clinics nevertheless introduces risks that a controlled pilot may not reveal.
Patients communicate differently, medical records vary in completeness and symptoms may present atypically. A system trained on one population may perform less reliably for another. Diagnostic suggestions can also influence clinicians even when those suggestions are incorrect.
The appropriate model is therefore clinical augmentation rather than automated authority.
Clinicians should remain responsible for final decisions, while the system’s recommendations, confidence and supporting evidence should be auditable. Patients should know when AI contributes to their care, and sensitive health information must be protected throughout the data chain.
Healthcare AI should be measured through practical outcomes: whether it reduces waiting times, improves documentation, identifies urgent cases and maintains consistent performance across demographic groups.
Deployment across 100 clinics gives Akido an opportunity to generate meaningful real-world evidence. It also raises the standard of accountability, because errors at network scale can spread much faster than mistakes within a small trial.
The bigger picture: AI power depends on institutional capacity
The six developments reveal an expanding divide between technical access and institutional capability.
European governments have access to AI but often lack integrated data, budgets and risk assessments for anti-fraud deployment. African economies are gaining the benefits of digital connectivity while confronting criminal organisations that use AI to industrialise deception. Washington wants earlier access to frontier models so that dangerous capabilities can be evaluated before release.
Developing countries are being urged to adopt AI quickly, yet they must avoid becoming permanently dependent on infrastructure and models controlled elsewhere. Meanwhile, healthcare providers such as Akido are showing how quickly AI can move into consequential real-world workflows.
This is the emerging geography of artificial intelligence.
Power will depend on more than owning chips or building the largest model. Governments and companies need interoperable systems, affordable computing, skilled personnel, useful data, independent evaluation and credible accountability.
The countries that develop those capabilities can use AI to improve public administration, healthcare, education and economic productivity. Those that adopt the technology without strengthening their institutions may acquire powerful tools without the ability to govern them.
The defining AI competition is therefore not simply between humans and machines or between the United States and China. It is a competition between societies that can turn technical capability into trusted infrastructure and those that cannot.
