New Delhi (ABC Live) Artificial intelligence is often presented as a competition between powerful models, companies and countries. However, the real struggle extends far beyond software.
AI depends on training data, advanced chips, cloud infrastructure, electricity, digital platforms, technical standards and global supply chains. Therefore, no single company or government can control the entire artificial intelligence ecosystem independently.
At present, private technology firms hold the strongest operational power. They build major models, own cloud platforms, control large pools of data and distribute AI tools to billions of users. Governments, meanwhile, retain legal authority over markets, energy, public procurement, national security and liability.
Nevertheless, legal power does not always create effective control. Many regulators still depend on private companies for technical knowledge, testing systems and computing infrastructure. Consequently, the firms being regulated may also influence the rules used to supervise them.
AI Power Extends Beyond Models and Data
The strongest AI actor may not be the company with the highest benchmark score. Instead, real power belongs to actors that combine data, chips, computing capacity, capital, distribution and legal influence.
For example, a model developer may own advanced software but still depend on Nvidia processors, foreign chip manufacturing and cloud services. Similarly, a government may control law but lack the infrastructure needed to develop its own systems.
Therefore, AI control is distributed across an interconnected chain. However, that distribution remains highly unequal.
The Geopolitical Dimension of AI Control
Artificial intelligence also depends on geopolitical stability.
The Strait of Hormuz provides a useful example. A serious disruption in this strategic route could raise energy prices, disturb shipping and increase operating costs for data centres, semiconductor factories and cloud infrastructure.
Moreover, AI supply chains depend on globally traded processors, memory chips, industrial gases, cooling equipment and specialised manufacturing tools. Consequently, a regional conflict can affect AI development far beyond the immediate battlefield.
This means that countries controlling energy routes, industrial inputs and strategic shipping corridors may influence the AI ecosystem indirectly. Therefore, future AI power will depend not only on models and data but also on secure trade routes, stable energy supplies and resilient supply chains.
Human Knowledge Is Becoming Strategic Infrastructure
AI systems learn from human-created records, including books, judgments, medical files, software code, images, videos and public conversations.
However, AI does not receive human experience directly. Instead, it receives a compressed and recorded version of that experience.
A court judgment may contain legal reasoning, but it may not capture every moral or social concern behind the decision. Similarly, a medical record may show a diagnosis but not the patient’s fear or financial condition.
Therefore, AI can reproduce the language of knowledge without possessing human awareness, responsibility or moral judgement.
The Central Question
The central question is not simply who will build the strongest AI model.
Rather, it is who will control the data, chips, energy, cloud infrastructure, standards, supply chains and laws that make artificial intelligence possible.
The answer will determine whether AI becomes a broadly shared public tool or a concentrated system of corporate and state power.
Summary
Artificial intelligence will not be controlled by one company, government or machine. Instead, different actors will control different parts of the artificial intelligence ecosystem.
Semiconductor companies control advanced chips. Meanwhile, cloud companies control computing capacity. Model developers control training and access. Digital platforms, in turn, control large pools of data and access to users.
Governments also hold important powers. For example, they control market access, data protection, electricity approvals, public procurement and legal liability. Users, however, mainly influence individual outputs through prompts.
Although these powers are connected, they are not equal. At present, large private companies hold the strongest operational position because they possess capital, infrastructure, data, technical talent and global distribution.
Governments still retain legal authority. Nevertheless, many regulators lack the technical capacity required to inspect advanced systems independently.
Therefore, the main question is not simply who will create the strongest model. Rather, the real issue is who will control the data, computing power, standards, model behaviour and social effects of artificial intelligence.
Key Points
- Artificial intelligence learns mathematical patterns from recorded data.
- However, AI does not experience the world like a human.
- Human knowledge enters AI through text, images, judgments, code, labels and feedback.
- Therefore, training data affects what a model knows and which errors it may reproduce.
- Prompt engineering guides an existing model, but it does not control the original training.
- Meanwhile, private companies lead in models, chips, cloud services and user distribution.
- Governments can regulate AI. Nevertheless, enforcement capacity remains uneven.
- Moreover, technical standards may shape AI control as strongly as legislation.
- India is expanding public computing capacity and India-focused data resources.
- However, no company or country controls the complete artificial intelligence ecosystem.
Why ABC Live Is Publishing This Report Now
Artificial intelligence has entered a more sober phase.
For several years, companies promoted larger models as if scale alone would create dependable intelligence. Consequently, investors rewarded those claims with extraordinary valuations.
Governments also treated access to advanced models as a measure of national strength. However, the limits of that approach are becoming clearer.
AI systems can produce fluent and confident answers. Nevertheless, fluent language does not guarantee truth. A model may invent facts, misunderstand context, reproduce bias or generate false references.
This gap creates what may be called perception dysmorphia. In other words, users may confuse the appearance of intelligence with genuine understanding.
At the same time, governments remain focused on visible resources such as chips, electricity, data centres and supply chains. These resources certainly matter. However, training data, model behaviour, benchmark integrity and public influence matter just as much.
Therefore, governments may be securing the physical inputs of AI while leaving its outputs under weak oversight.
ABC Live has previously examined related questions through its reports on India’s data sovereignty before AI development, the AI Governance Techno-Legal Framework and the IndiaAI Mission’s Varya video model.
This report, therefore, brings those issues together. More importantly, it asks who may ultimately control the artificial intelligence ecosystem.
What Is the Artificial Intelligence Ecosystem?
The artificial intelligence ecosystem includes far more than chatbots or language models.
Instead, it includes all the resources, institutions and actors needed to build, operate, distribute and regulate AI systems.
For example, the ecosystem includes:
- training data;
- AI models;
- semiconductor chips;
- cloud infrastructure;
- data centres;
- electricity;
- technical talent;
- digital platforms;
- applications;
- standards;
- regulators;
- courts; and
- users.
Therefore, control over AI cannot be understood through model ownership alone.
A company may own a model. However, it may still depend on another company for chips, cloud services and data-centre capacity.
Similarly, a government may possess legal authority. Nevertheless, it may lack the technical expertise needed to inspect the system.
Consequently, AI power emerges from interdependence rather than complete independence.
Main Layers of AI Control
Different actors control different points of the AI chain. Therefore, no single layer explains the full balance of power.
AI Control Dashboard
| Layer | Main controller | Main power |
|---|---|---|
| Data | Platforms and data holders | Decide what enters training |
| Models | AI developers | Control training and access |
| Chips | Semiconductor firms | Supply computing power |
| Cloud | Cloud providers | Allocate scalable compute |
| Energy | Governments and utilities | Approve and supply power |
| Prompts | Users and organisations | Guide individual outputs |
| Standards | Industry and regulators | Define compliance tests |
| Law | Governments and courts | Permit, restrict and punish |
| Distribution | Platforms and software firms | Control user reach |
No actor controls every layer. However, some firms operate across several layers at once.
Google, for example, combines search, YouTube, Android, cloud infrastructure and Gemini models. Therefore, it can influence data collection, model development and distribution.
Similarly, Microsoft combines cloud services, enterprise software, GitHub, Bing and major AI partnerships. As a result, it holds influence across infrastructure, professional data and business adoption.
Amazon, meanwhile, combines cloud infrastructure, custom processors, e-commerce and enterprise distribution.
Thus, the strongest AI actors are usually those that combine several sources of power.
What Does Training an AI Mean?
Training an AI means teaching a computer model to identify patterns from large quantities of data.
For a language model, developers provide large collections of text. The system then studies how words, phrases, sentences and ideas relate to one another.
Afterwards, the model adjusts a large number of internal numerical values called parameters. These parameters help the model predict which word, phrase or response is most likely to follow an input.
According to OpenAI’s training-data summary, its systems use publicly available information, data obtained through partnerships and material supplied or generated by users, trainers and researchers.
However, developers generally disclose broad data categories. They do not usually publish a complete list of every book, webpage, image or record used during training.
Basic AI Training Process
| Stage | What happens |
| Collection | Developers gather text, images, audio and code |
| Cleaning | Poor or restricted material may be removed |
| Conversion | Data becomes machine-readable units |
| Pre-training | The model learns broad patterns |
| Prediction | The model predicts missing or next units |
| Correction | Mathematical methods reduce errors |
| Fine-tuning | The model is adapted for specific tasks |
| Human feedback | Reviewers rank and correct outputs |
| Safety testing | Teams search for harmful behaviour |
| Deployment | The model becomes available |
| Monitoring | Developers track failures and updates |
Training an advanced model requires enormous resources. In particular, it requires data, chips, electricity, technical talent and capital.
As a result, only a limited number of organisations can train the largest systems.
What Does It Mean That AI Learns Through Mathematics?
AI does not learn through feelings, awareness or personal experience.
Instead, it converts words, images, audio and other information into numbers. It then uses mathematical calculations to identify relationships among those numbers.
For example, an AI system may learn that “fire,” “smoke,” “heat,” “burn” and “danger” often appear together.
However, the model does not feel heat. Likewise, it does not fear injury or understand pain through lived experience.
A human child may learn that fire is dangerous by seeing flames, feeling warmth and hearing a warning. By contrast, an AI system learns because recorded data repeatedly connects fire with danger.
Human Learning Versus AI Learning
| Human learning | AI learning |
| Uses lived experience | Uses recorded data |
| Includes emotion and awareness | Uses numerical patterns |
| Understands direct consequences | Learns descriptions of consequences |
| Develops moral judgement | Follows patterns and rules |
| Can accept responsibility | Cannot accept responsibility |
| Builds common sense through life | Approximates common patterns |
Therefore, AI may produce a thoughtful-looking response without possessing human understanding.
It can recognise patterns associated with meaning. However, it does not necessarily experience that meaning.
How Human Experience Becomes AI Data
Human experience enters AI only after people convert it into records.
These records may include conversations, books, judgments, medical files, photographs, videos, software code, research papers and human ratings.
However, the model cannot directly absorb a doctor’s concern, a judge’s hesitation or a worker’s practical wisdom. Instead, it receives a simplified representation of that experience.
Human Experience Translation
| Human activity | Recorded form | What AI receives |
| Medical diagnosis | Notes, scans and codes | Symptom-diagnosis patterns |
| Judicial decision | Facts and reasoning | Issue-outcome patterns |
| Safe driving | Video and sensors | Road-response patterns |
| Teaching | Lessons and explanations | Explanation patterns |
| Emotional expression | Words, tone and images | Emotion-related signals |
| Programming | Code and corrections | Software patterns |
| Journalism | Reports and sources | Factual presentation patterns |
This translation process compresses human experience.
For example, a medical record may state that a medicine was prescribed. However, it may not record the doctor’s uncertainty, the patient’s fear or the family’s financial problem.
Similarly, a judgment records formal legal reasoning. Nevertheless, it may not capture every courtroom interaction or wider social condition.
Therefore, AI receives a partial record of human life rather than human experience itself.
How Human Feedback Shapes AI Behaviour
Human reviewers often evaluate AI outputs after the first stage of training.
They may decide which response is more useful, accurate, safe or clear. Afterwards, the system is adjusted to favour similar responses.
Human Feedback Process
| Stage | Human role |
| Prompt creation | Prepare questions and tasks |
| Comparison | Compare alternative answers |
| Ranking | Select the better response |
| Error marking | Identify false or unsafe content |
| Red-team testing | Search deliberately for weaknesses |
| Policy design | Set behavioural limits |
Human feedback can improve a system. However, it also introduces human preferences and institutional values.
Different reviewers may disagree about fairness, political balance, harmful speech or social norms. Consequently, the model may reflect the values of the organisation supervising the process.
Therefore, control over human feedback is also a form of control over AI behaviour.
What Is Training Data?
Training data is the information used to teach an AI model.
It may include public websites, books, journals, news reports, software code, images, videos, licensed archives and human-created examples.
In addition, developers may use contractor-produced material, opted-in user information and synthetic data created by other AI systems.
Training data influences what a model can recognise, reproduce and predict.
Therefore, if the data contains falsehoods, bias or poor representation, the model may repeat those weaknesses.
Main Training-Data Risks
| Risk | Possible result |
| Inaccurate material | False answers |
| Historical bias | Discrimination |
| Language imbalance | Poor local performance |
| Copyrighted works | Legal disputes |
| Personal information | Privacy concerns |
| Propaganda | Distorted outputs |
| Duplicate material | Overweighted claims |
| Benchmark leakage | Inflated test scores |
| Outdated records | Obsolete advice |
| Synthetic overuse | Repeated errors |
Better data does not simply mean more data.
Instead, a smaller collection of accurate, diverse and specialised information may produce better results than a massive collection of poor-quality material.
Training Data, User Data and Retrieval Data
These categories are related. However, they are not identical.
Main Data Categories
| Data type | Meaning |
| Training data | Builds or improves the model |
| User data | Comes from service use |
| Licensed data | Used under an agreement |
| Public data | Available through public sources |
| Feedback data | Created by human reviewers |
| Synthetic data | Generated by computers or AI |
| Evaluation data | Tests model performance |
| Retrieval data | Accessed while answering |
Users should not assume that every conversation automatically becomes training data.
Instead, the answer depends on the company, product, account settings and applicable policy.
Therefore, meaningful transparency requires clear information about how user content is stored, reviewed and used.
Who Controls the Most Valuable Data?
The strongest data advantages often belong to companies that already operate large digital platforms.
Major Data Holders
| Company | Main data advantage |
| Search, YouTube, Maps and Android | |
| Meta | Social, image and video content |
| Microsoft | GitHub, LinkedIn, Bing and enterprise data |
| Amazon | Commerce, reviews, cloud and Alexa |
| ByteDance | TikTok video and engagement data |
| X and xAI | Real-time public conversations |
| Adobe | Creative workflows and licensed media |
| Public community discussions | |
| Getty Images | Licensed visual archives |
| Shutterstock | Commercial visual content |
Platform companies control more than stored information.
They can also observe what users search for, watch, reject, correct and share. Therefore, behavioural information can become as valuable as the original content.
Moreover, this advantage compounds over time. The more people use a platform, the more interaction data the company can collect.
Consequently, established platforms may gain an advantage that new competitors cannot easily reproduce.
Why Data Sovereignty Matters
Training data is not only a commercial resource. It is also a source of national power.
ABC Live examined this issue in Explained: India Data Sovereignty Before AI Development.
That report argued that India may generate large volumes of useful data while foreign companies use those records to train systems outside effective Indian control.
Similarly, the ABC Live report on India’s role in data sovereignty explained why data control now affects trade, security and technological independence.
Why Data Control Matters
| Issue | Strategic effect |
| Ownership | Decides who may exploit data |
| Location | Influences applicable law |
| Consent | Shows whether use was authorised |
| Licensing | Determines creator payment |
| Language | Affects local performance |
| Quality | Influences reliability |
| Foreign access | Creates security risks |
| Auditability | Enables outside inspection |
Data sovereignty does not necessarily mean keeping all data inside national borders.
Rather, it means preserving meaningful legal authority over how sensitive information is collected, transferred, processed and commercialised.
Therefore, states need both data flows and enforceable safeguards.
AI Training and Copyright
Training models on journalism, books, music, images and software raises major copyright questions.
ABC Live examined these concerns in The Constitutional Risks in DPIIT’s AI Copyright Plan.
The main questions include:
- Can developers use publicly accessible works without permission?
- Should authors, artists and journalists receive payment?
- Must companies identify the works used during training?
- Can creators refuse AI-training use?
- Who is responsible when a model reproduces protected material?
- Can a large company obtain exclusive access to a valuable archive?
Large AI firms can afford expensive licensing agreements. By contrast, smaller developers may depend on public, synthetic or lower-quality material.
Therefore, licensing may improve legal certainty while also increasing market concentration.
Moreover, exclusive agreements could allow a small number of firms to control high-quality information.
Which Private Companies Lead AI?
Leadership depends on which part of the ecosystem is being measured.
A company may lead in models, chips, cloud infrastructure, data, business applications or consumer distribution.
Leading Model Developers
| Company | Main strength |
| OpenAI | General models, ChatGPT and agents |
| Anthropic | Claude and enterprise AI |
| Google DeepMind | Gemini and scientific AI |
| Meta | Llama and global distribution |
| xAI | Grok and X integration |
| DeepSeek | Efficient reasoning models |
| Alibaba | Qwen and cloud deployment |
| ByteDance | Consumer AI and recommendations |
| Baidu | Ernie, search and cloud |
| Mistral AI | European foundation models |
| Cohere | Secure enterprise models |
| Sarvam AI | Indian-language systems |
No developer holds a permanent technical lead.
Instead, model rankings change whenever companies release new systems. Therefore, one benchmark result cannot establish overall leadership.
Moreover, model quality is only one source of influence. A company with weaker models may still hold enormous power through cloud services or distribution.
Which Companies Control AI Infrastructure?
AI models cannot operate without processors, memory, networks, data centres and cloud services.
Consequently, infrastructure companies may hold deeper structural power than model developers.
AI Infrastructure Leaders
| Company | Main role |
| Nvidia | AI processors and software |
| TSMC | Advanced chip manufacturing |
| AMD | AI accelerators |
| Broadcom | Custom chips and networking |
| SK Hynix | High-bandwidth memory |
| Samsung | Memory and semiconductor supply |
| ASML | Chipmaking equipment |
| CoreWeave | AI-focused cloud computing |
| Microsoft Azure | Cloud and enterprise AI |
| Amazon Web Services | Cloud and custom chips |
| Google Cloud | Cloud, models and processors |
| Oracle Cloud | Large AI computing clusters |
Nvidia is among the most important AI companies because many developers use its processors and software platform.
However, Nvidia depends on outside chip manufacturers, memory providers, electricity systems and international supply chains.
Therefore, even the strongest infrastructure company cannot operate independently.
ABC Live Private-Sector AI Power Dashboard
| Actor | Main strength | Position |
| Nvidia | Chips and software | Core infrastructure power |
| Microsoft | Cloud and enterprise systems | Major platform power |
| Data, models and distribution | Integrated AI power | |
| Amazon | Cloud and custom chips | Infrastructure gatekeeper |
| OpenAI | Model adoption | Leading model influence |
| Meta | Platforms and open models | Distribution power |
| Anthropic | Frontier enterprise models | Rising model power |
| TSMC | Advanced chip production | Manufacturing gatekeeper |
| ByteDance | Data and recommendations | Platform power |
| Huawei | Chips, cloud and telecom | China-focused power |
This dashboard is not a simple ranking of model quality.
Instead, it examines infrastructure, data, distribution, models and strategic dependence together.
How Much Money Is Flowing into AI?
Private investment helps explain why companies currently hold so much operational power.
The Stanford Artificial Intelligence Index Report 2026 reported that United States private AI investment reached $285.9 billion in 2025.
China recorded $12.4 billion in private investment. However, Stanford cautioned that these figures may not capture the full scale of state-backed support.
Private AI Investment
| Country | 2025 investment |
| United States | $285.88 billion |
| China | $12.41 billion |
| United Kingdom | $5.90 billion |
| France | $4.36 billion |
| Canada | $4.28 billion |
| India | $4.09 billion |
The United States attracted more than 23 times China’s reported private investment.
However, investment does not prove that every AI business will become profitable. Rather, it shows where companies can obtain chips, researchers, licences and infrastructure.
Large firms can also absorb compliance expenses, legal disputes and safety-testing costs. Smaller firms, by contrast, may struggle to meet the same requirements.
Therefore, financial power can gradually become market power and regulatory influence.
AI, Electricity and Data Centres
Artificial intelligence may appear digital. However, it depends on physical infrastructure.
Large models require processors, data centres, electricity, cooling systems and network connections.
The International Energy Agency reported that global data-centre electricity demand rose by 17% in 2025, while total global electricity demand grew by about 3%.
AI Energy Dashboard
| Indicator | Figure |
| Data-centre demand growth, 2025 | 17% |
| Global demand growth, 2025 | About 3% |
| US data-centre share, 2024 | 45% |
| China share, 2024 | 25% |
| Europe share, 2024 | 15% |
| Renewable supply growth to 2030 | About 22% yearly |
Global percentages can hide local pressure.
Data centres concentrate electricity demand in particular cities and industrial zones. Therefore, a country may have enough power overall but lack grid capacity in a specific location.
As a result, governments and utilities hold substantial power through grid connections, tariffs, land approval and environmental clearance.
ABC Live examined this relationship in How Power Grid Powers AI and Quantum India.
That report explained why electricity transmission is essential for data centres, semiconductor plants and advanced scientific facilities.
Why Small Supply-Chain Inputs Matter
AI supply chains depend on many specialised components.
These include advanced processors, high-bandwidth memory, chipmaking machines, industrial gases, chemicals, cooling systems and shipping routes.
Therefore, disruption involving one apparently minor input can slow the wider system.
For example, semiconductor manufacturing uses helium in several technical processes. Consequently, supply disruptions can create pressure on chip production.
However, helium does not control artificial intelligence by itself.
Instead, the larger lesson is that minor inputs may become strategic bottlenecks inside tightly connected supply chains.
AI Supply-Chain Risks
| Input | Main risk |
| Advanced processors | Concentrated manufacturing |
| High-bandwidth memory | Few major suppliers |
| Chipmaking equipment | Limited qualified makers |
| Industrial gases | Regional disruption |
| Electricity | Grid shortages |
| Cooling systems | Water and environmental limits |
| Networks | Cybersecurity risks |
| Cloud capacity | Market concentration |
ABC Live has examined India’s semiconductor challenge through its critical analysis of India Semiconductor Mission 2.0 and its report on how India can become the world’s generic chip factory.
Together, these reports show that technological sovereignty requires more than one advanced factory. Instead, it also requires packaging, testing, skilled workers, dependable electricity and manufacturing equipment.
What Is Prompt Engineering?
Prompt engineering means designing instructions that help an AI system produce a useful response.
A prompt may state the task, relevant facts, intended reader, preferred format, evidence requirements and limits.
For example, “Explain this judgment” is broad.
A stronger prompt may require the model to identify facts, legal issues, arguments, precedents, the operative order and the ratio. Moreover, it may require the system to separate verified facts from assumptions.
Therefore, the second prompt gives the model clearer direction.
What Prompt Engineering Can Control
| Can influence | Cannot directly control |
| Purpose | Original training data |
| Structure | Model architecture |
| Tone | Hidden parameters |
| Detail level | Provider policies |
| Supplied facts | Cloud infrastructure |
| Output limits | Built-in weaknesses |
| Citation requests | Genuine understanding |
Prompt engineering gives users operational influence over a particular task.
However, the model developer retains structural control over the training process, architecture, safety rules and infrastructure.
Prompt Engineering Versus Training
| Prompting | Training |
| Directs an existing model | Changes internal parameters |
| Can be done by users | Requires large resources |
| Affects one output | Affects wider behaviour |
| Changes instantly | Takes major technical work |
| Costs relatively little | Can cost enormous sums |
In simple words:
Training teaches the model broad patterns. A prompt, however, tells the trained model which task to perform.
How Experience Becomes an AI Output
| Step | Function |
| Human experience | Creates knowledge and judgement |
| Records | Preserve part of that experience |
| Data | Makes records machine-readable |
| Training | Converts patterns into parameters |
| Human feedback | Shapes preferred behaviour |
| Prompt | Directs the model |
| Human review | Checks the result |
The complete process can therefore be summarised as follows:
Human experience creates records. Records become data. Training converts data into mathematical patterns. Prompts direct those patterns towards a task. Finally, human judgement verifies the result.
Why Prompt Engineering Matters in Government
Prompt engineering can turn a general model into a structured work assistant.
For example, a government department may require the system to identify the applicable rule, state missing evidence and separate facts from assumptions.
Moreover, the prompt may instruct the AI not to make the final decision. Instead, it may require referral to a responsible human officer.
Similarly, a newsroom may require the system to distinguish confirmed information, allegations and editorial analysis.
However, a strong prompt cannot guarantee a correct answer.
Therefore, high-risk uses also require verified data, audit logs, external testing, clear liability and human review.
What Is Prompt Injection?
Prompt injection occurs when external material attempts to manipulate an AI system’s instructions.
For example, a document may contain hidden instructions telling the AI to ignore the user’s request or disclose confidential information.
Prompt-Related Risks
| Risk | Possible effect |
| Vague prompt | Irrelevant output |
| Leading question | Biased conclusion |
| False premise | Incorrect assumption |
| Missing context | Incomplete answer |
| Prompt injection | Instruction override |
| Conflicting directions | Unstable result |
| Excessive trust | Acceptance of false output |
Therefore, prompt engineering can improve control. Nevertheless, it cannot make AI completely reliable.
Are Governments Focusing on the Wrong Risks?
Governments have strong reasons to protect processors, electricity, telecommunications and data centres.
However, physical resources cover only one part of the problem.
Commonly Protected AI Inputs
- processors;
- energy;
- industrial materials;
- telecommunications;
- data centres;
- foreign investment; and
- supply routes.
Commonly Under-Governed AI Outputs
- fabricated answers;
- discriminatory decisions;
- manipulation;
- surveillance;
- benchmark gaming;
- copyright violations;
- automated exclusion;
- political influence; and
- loss of human accountability.
Therefore, governments may be over-securitising AI inputs while under-governing AI outputs.
A state may restrict processor exports. Yet it may still allow an untested automated system to influence healthcare, education or welfare decisions.
What Is Regulatory Capture?
Regulatory capture occurs when the industry being supervised gains excessive influence over its regulators.
Governments need technical advice from AI companies. However, the relationship becomes risky when regulators cannot test company claims independently.
Routes of AI Regulatory Capture
| Route | Possible result |
| Company benchmarks | Firms define favourable tests |
| Industry-funded research | Commercial priorities dominate |
| Company experts on panels | Regulators become dependent |
| Expensive compliance | Large firms gain advantage |
| Technical secrecy | Verification becomes difficult |
| Standard-setting control | Powerful firms shape rules |
| Cloud dependence | Governments rely on suppliers |
Regulation may therefore appear strict while strengthening the largest companies.
Moreover, smaller developers may face rules designed around the resources of global corporations.
Can AI Safety Benchmarks Be Captured?
Benchmarks are tests used to measure model performance.
They may examine reasoning, coding, safety, bias or hallucination rates.
However, a benchmark can become unreliable when its questions appear in the model’s training data. Similarly, developers may optimise a system specifically for a known test.
Moreover, firms may publish only favourable results. Consequently, benchmark scores may not reflect real-world behaviour.
Main Benchmark Limits
| Test | Main limitation |
| Language | Questions may be in training data |
| Coding | High scores may hide insecure code |
| Mathematics | Results may not transfer to reality |
| Bias | Definitions differ |
| Hallucination | Results depend on prompts |
| Cybersecurity | Disclosure creates risks |
| Social impact | Hard to predict before use |
| Post-deployment | Failures may go unreported |
The Stanford AI Index 2026 found that developers commonly report capability benchmarks. However, reporting on responsible AI remains less consistent.
Therefore, a strong benchmark score should not be treated as proof of safety.
The ICML 2026 Review Controversy
The International Conference on Machine Learning reported that 795 reviews involving 506 reviewers showed prohibited large language model use.
The affected reviews represented about one per cent of all reviews.
The figure was limited. Therefore, the incident did not invalidate the complete peer-review system.
However, it exposed a wider governance problem.
What the Incident Revealed
| Issue | Wider lesson |
| Rules were accepted but broken | Written rules are not enough |
| AI reviews looked professional | Fluency can hide weak understanding |
| Confidential papers may have entered AI tools | Research needs data safeguards |
| Detection came later | Oversight may lag behind misuse |
| Hundreds were involved | Misuse may become institutional |
The case demonstrates that AI governance requires enforcement capacity, not merely written policy.
Is AI Transparency Improving?
Transparency remains uneven.
The Stanford Foundation Model Transparency Index rose from 37 in 2023 to 58 in 2024. However, it fell to 40 in 2025.
Transparency Index
| Year | Score |
| 2023 | 37 |
| 2024 | 58 |
| 2025 | 40 |
| 2024–2025 change | Down 18 points |
Major gaps remain in disclosure of training data, computing resources, safety testing and environmental effects.
Moreover, firms often disclose limited information about post-deployment failures and wider social impact.
Therefore, model capability may rise while public visibility falls.
How Does the TikTok Case Relate to AI?
The TikTok dispute in the United States showed how digital platforms can become geopolitical assets.
The case concerned more than ownership. Instead, it also involved user data, recommendation systems, foreign access and influence over public information.
Therefore, it suggests tools that governments may later apply to AI providers.
Possible AI Controls
| Control tool | Possible AI use |
| Local legal entity | Domestic accountability |
| Local data storage | Sensitive-data protection |
| Independent audit | Model inspection |
| Ownership limits | Foreign-control restrictions |
| Approved cloud hosting | Infrastructure security |
| Algorithm oversight | Greater visibility |
| Market conditions | Compliance before access |
Consequently, future AI disputes may combine ownership, data location and algorithmic oversight.
How Is the European Union Regulating AI?
The European Union Artificial Intelligence Act follows a risk-based approach.
The law prohibits some practices. In addition, it imposes stronger obligations on high-risk systems.
However, the Act applies through several phases.
EU AI Act Timeline
| Date | Main development |
| 1 August 2024 | Act entered into force |
| 2 February 2025 | Prohibited practices began |
| 2 August 2025 | General-purpose AI rules began |
| 2 August 2026 | Major governance rules apply |
| 2 August 2027 | Some product-linked duties apply |
The European Commission’s implementation timeline provides the phased schedule.
The European Union may influence global business practices because companies want access to its market.
However, regulation will work only if authorities can inspect systems, test claims and impose meaningful penalties.
The United States, China, Europe and India
Different regions exercise different forms of AI power.
Global AI Power Models
| Region | Main strength | Main weakness |
| United States | Capital, models and cloud | Fragmented regulation |
| China | State direction and platforms | External chip limits |
| European Union | Regulation and market access | Fewer leading model firms |
| India | Talent, users and public compute | Developing enforcement |
| Southeast Asia | Manufacturing and digital growth | Uneven capacity |
The United States currently holds the strongest commercial position.
China, however, exercises stronger direct state influence over domestic platforms and deployment.
The European Union, meanwhile, holds considerable regulatory and standard-setting power.
India is expanding computing resources while developing its governance system.
Therefore, no single national model has achieved complete control.
India’s Emerging AI Position
India is building an AI ecosystem involving computing capacity, datasets, indigenous models, skills and regulation.
The Government of India approved the IndiaAI Mission with an outlay of ?10,371.92 crore.
Moreover, official information issued in 2026 stated that more than 38,000 graphics processing units had been onboarded under the mission.
As of February 2026, AIKosh reportedly brought together 7,541 datasets and 273 AI models across 20 sectors.
India AI Dashboard
| Indicator | Position |
| IndiaAI Mission outlay | ?10,371.92 crore |
| GPUs onboarded | More than 38,000 |
| AIKosh datasets | 7,541 |
| AIKosh models | 273 |
| Sectors covered | 20 |
These developments may reduce the cost barrier faced by universities, researchers and start-ups.
However, computing capacity alone does not create trustworthy AI.
Therefore, India also needs multilingual benchmarks, training-data transparency, clear liability, independent testing and public procurement standards.
ABC Live examined these concerns in its analysis of India’s AI governance techno-legal framework.
Major Indian AI Companies
| Company | Main area |
| Sarvam AI | Indian-language models |
| Krutrim | Models and cloud |
| TCS | Enterprise AI |
| Infosys | Business AI services |
| Wipro | AI integration |
| Tech Mahindra | Language and industry AI |
| Fractal Analytics | Enterprise analytics |
| Yellow.ai | Conversational AI |
| Haptik | Customer-service AI |
| Reliance Jio | Infrastructure and distribution |
India’s strength currently lies in enterprise services, local-language applications and large-scale deployment.
However, India remains less powerful in frontier-model development and advanced-chip production.
Therefore, its strategy will depend on combining domestic capacity with carefully managed international partnerships.
Can Governments Outsource Judgement to AI?
AI can help governments search records, organise documents and identify patterns.
However, it should not become the final decision-maker in matters involving liberty, welfare, healthcare, education or employment.
An AI system does not possess democratic responsibility or moral awareness.
Moreover, if officials rely too heavily on automated outputs, they may gradually lose internal expertise.
Consequently, excessive AI dependence may weaken rather than strengthen state capacity.
Government Use of AI
| Lower-risk use | Higher-risk use |
| Summarising records | Final welfare decisions |
| Organising documents | Criminal-risk scoring |
| Translation | Predicting guilt |
| Finding missing fields | Denying medical care |
| Supporting research | Replacing judicial reasoning |
| Drafting options | Making final policy decisions |
Human officials must therefore remain accountable for decisions affecting rights.
Who Will Control AI Standards?
Laws often use broad words such as safe, transparent or fair.
Technical standards, however, determine how companies demonstrate compliance with those principles.
Standards may cover cybersecurity, data quality, documentation, incident reporting and human oversight.
Large companies can send experts to international standard-setting meetings. Smaller companies and developing countries, by contrast, may lack comparable resources.
Consequently, standards may reflect the interests of the strongest participants.
Governments should therefore support researchers, start-ups and public-interest organisations that take part in these processes.
ABC Live Global AI Control Dashboard
| Actor | Main strength | Position |
| US technology firms | Capital, models and cloud | Strongest commercial power |
| US Government | Regulation and security | Strong security influence |
| China | State support and data | Strong state-led power |
| European Union | Regulation and market access | Strongest formal regulator |
| India | Data, users and growing compute | Major emerging power |
| Cloud providers | Computing infrastructure | Critical gatekeepers |
| Chip companies | Advanced processors | Control physical foundations |
| Researchers | Knowledge and testing | Important but underfunded |
| Citizens | Data and democratic pressure | Limited direct control |
Who Is Most Likely to Control AI?
Time-Based Assessment
| Period | Likely leaders |
| Next 1–3 years | Large technology, chip and cloud firms |
| Next 3–5 years | Governments and regulatory blocs |
| Longer term | Public-private governance networks |
| Main risk | Concentrated corporate-government control |
| Best outcome | Shared control with independent oversight |
What May Happen During the Next 6–12 Months?
AI and Economic Diplomacy May Remain Separate
Governments may continue negotiating trade and supply-chain agreements while treating AI governance as a secondary matter.
However, this separation will become harder because cloud services and cross-border data now affect ordinary commerce.
Therefore, future trade agreements may increasingly include rules for data, model access and digital services.
Technology Diplomacy May Divide
One track will focus on software, services, data flows and interoperability.
The second track, meanwhile, will focus on chips, manufacturing, robotics and supply-chain resilience.
Countries such as India, Vietnam and Indonesia may therefore need to balance both approaches.
Regulation May Become More Contract-Based
Governments may increasingly use procurement contracts to require testing, auditing, incident reporting and liability.
Contracts can move faster than general laws. However, they must remain transparent because citizens do not negotiate public contracts.
Therefore, contractual control should supplement rather than replace legislation.
What Should Governments Do?
Governments should move beyond general promises of responsible AI.
First, they should require independent testing under real-world conditions.
Second, they should require companies to report serious failures and security incidents.
Third, public authorities should retain audit rights over systems used in essential services.
Fourth, competition regulators should examine links among chip firms, cloud providers, model developers and platforms.
Fifth, governments should build technical expertise inside public institutions.
Sixth, procurement contracts should clearly assign liability when AI causes harm.
Seventh, developing countries and smaller companies should receive support in standard-setting processes.
Finally, no person should lose liberty, employment, education, healthcare or public benefits solely because an algorithm produced a score.
ABC Live Analysis
The evidence shows that private companies currently exercise more operational control over AI than governments.
They control advanced models, computing infrastructure, training data, technical talent and distribution networks. Moreover, their financial resources exceed the technical budgets of many regulators.
However, corporate power still depends on public systems.
Companies need electricity, land approvals, telecommunications, courts, intellectual-property protection and market access. Therefore, governments retain significant structural power.
The deeper problem is the imbalance of information.
Model developers know more about their systems than regulators, customers or citizens. Consequently, they can decide what information to disclose and which test results to publish.
As a result, regulatory capture becomes a serious risk.
Prompt engineering gives users some control over individual outputs. However, it does not alter the model’s architecture, original training data or company policies.
Therefore, users operate within boundaries created by the provider.
Training data creates another layer of control. Companies that operate large platforms can access text, video, images and behavioural information.
Yet human experience becomes compressed when it enters an AI system.
For example, AI may identify patterns in judgments without understanding justice as a judge does. Similarly, it may analyse medical files without experiencing illness.
Therefore, governments and institutions must not confuse statistical prediction with accountable human judgement.
The decisive struggle will not concern only who owns the most capable chatbot.
Instead, real control will belong to actors that can combine data, chips, cloud infrastructure, energy, capital, model access, standards, distribution and law.
At present, this power lies mainly within a small network of technology firms, infrastructure providers and powerful governments.
Risks and Concerns
Concentration of Corporate Power
A few companies may control models, infrastructure and distribution at the same time.
Weak Training-Data Transparency
Outside researchers may not know which information shaped a model.
Regulatory Capture
Large companies may influence the standards used to evaluate their own systems.
Benchmark Manipulation
Models may be optimised for known tests without becoming safer.
Loss of Human Judgement
Officials may rely on fluent but unreliable outputs.
Copyright and Privacy Disputes
Training datasets may contain protected or personal information.
Language and Cultural Exclusion
Models may perform poorly for underrepresented languages and communities.
Energy and Environmental Pressure
Data centres may create local pressure on grids, water and land.
Accountability Gaps
Developers, deployers and users may blame one another when harm occurs.
What Happens Next?
AI governance will probably become more practical and less abstract.
Governments will increasingly use audits, procurement conditions, competition law, data protection and contractual liability.
Courts will also address copyright, discrimination, consumer harm and responsibility for automated decisions.
Meanwhile, geopolitical competition will continue around chips, electricity, data centres and cloud infrastructure.
However, physical resources explain only part of AI power.
A country may possess chips but lack high-quality data. Similarly, a company may own a model but depend on foreign manufacturing.
A government may also hold legal authority while lacking technical knowledge.
Therefore, the future of AI will be shaped by interdependence rather than complete control.
Frequently Asked Questions
Who currently controls the artificial intelligence ecosystem?
Large technology companies control many leading models and platforms. However, chip manufacturers, cloud providers and governments control other essential layers.
What does AI training mean?
AI training means using data and mathematical methods to teach a model to identify patterns and make predictions.
Does AI understand like a human?
No. Instead, AI processes numerical patterns. It does not possess human experience, emotion or moral responsibility.
How does human experience enter AI?
Human experience becomes text, images, records, labels and feedback. Afterwards, AI learns patterns from those records.
Why is training data important?
Training data affects what a model knows, how it responds and which biases it may reproduce.
What is prompt engineering?
Prompt engineering means writing clear instructions that guide an existing model towards a particular task.
Can a strong prompt prevent every error?
No. Although a good prompt can improve focus, it cannot remove all hallucinations, bias or technical limitations.
Which companies lead AI?
Important leaders include OpenAI, Anthropic, Google, Meta, Microsoft, Amazon, Nvidia, xAI, DeepSeek, Alibaba and Taiwan Semiconductor Manufacturing Company.
Which company has the most AI power?
No company controls every layer. Nvidia is central to computing. Meanwhile, Microsoft and Amazon lead major cloud systems, while Google combines models, data and distribution.
Can governments control AI companies?
Yes. Governments can regulate market access, data, competition, electricity, procurement and liability. However, effective enforcement requires technical capacity.
Will legislation alone control AI?
No. Instead, effective governance also requires standards, audits, contracts, courts and human accountability.
Relevant ABC Live Reports
- Explained: India Data Sovereignty Before AI Development
- Explained: India’s Role in Data Sovereignty
- The Constitutional Risks in DPIIT’s AI Copyright Plan
- Critical Analysis of AI Governance Techno-Legal Framework White Paper
- Critical Analysis of IndiaAI Mission’s Varya Video Model
- Explained: Google’s Visakhapatnam Data Centre and Digital Sovereignty
- Explained: How Power Grid Powers AI and Quantum India
- Critical Analysis of India Semiconductor Mission 2.0
- Explained: How India Can Become the World’s Generic Chip Factory
Sources and Methodology
ABC Live prepared this report by comparing official policy documents, institutional research, company disclosures, energy analysis and earlier ABC Live reports.
Primary Sources
- Stanford University — Artificial Intelligence Index Report 2026
- Stanford AI Index — Economy
- Stanford AI Index — Responsible AI
- International Energy Agency — Energy and AI
- OpenAI — Training Data Summary
- International Conference on Machine Learning — LLM Review-Policy Violations
- European Commission — AI Act Implementation Timeline
- IndiaAI Compute Portal
Methodological Limitations
Private investment figures do not capture every form of state support or internal corporate spending.
Moreover, AI developers publish broad categories of training data rather than complete dataset inventories. Therefore, outside researchers cannot identify every source that shaped a model.
Benchmark scores also change quickly. In addition, success on a controlled test does not prove real-world safety.
This report therefore distinguishes among ownership, operational power, infrastructure control, training-data control, prompt influence, legal authority and democratic accountability.
That distinction is necessary because no company or government currently controls every part of the artificial intelligence ecosystem.
Conclusion
Artificial intelligence will not be controlled by one government, company or machine.
Instead, different actors will control different layers.
Model developers will control training and access. Meanwhile, chip and cloud companies will control computing power.
Digital platforms will control large pools of data. Governments, in turn, will control laws and market access. Standards bodies will define compliance, while users will guide individual outputs through prompts.
However, formal authority does not always create effective control.
A government may hold legal power but depend on private technical expertise. Similarly, a model developer may own software but depend on foreign chips and cloud infrastructure.
A user may write a detailed prompt. Nevertheless, that user still operates within rules established by the provider.
Therefore, the greatest danger is not that one AI machine suddenly becomes the ruler of society.
Rather, the more immediate danger is that a small network of companies and governments may control data, infrastructure, model behaviour and public information without sufficient scrutiny.
AI learns from recorded human experience. Yet it does not inherit human awareness, accountability or moral judgement.
Therefore, artificial intelligence should remain a tool that supports human decision-making. It should not become an unquestioned authority over human life.
ABC Live — Making Complex Public Issues Simple.