Current Affairs · · Prelims · Science & Technology

US AI safety accord sets four voluntary control layers

On 29 September 2026, the US President and leaders of six technology firms signed a voluntary accord on frontier AI safety. It asks companies to use internal controls, safety teams, external evaluation and board oversight. The accord itself creates no regulator or statutory penalties. The EU AI Act, by contrast, imposes legal duties on providers of general-purpose AI models.

Event date:

REq1
  1. Context1 / 4

    On 29 September 2026, the US President and leaders of Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed the White House Accord on Super Intelligence, subtitled Joint Commitment on Frontier Responsibilities.

    The accord says companies developing frontier AI models should implement four layers of controls and audits. It is a voluntary commitment. Its text creates no federal licensing body, statutory duty or penalty for failing to follow these steps. The signatories also say they will meet regularly to develop standards and practices, and that the steps might later be written into law.

    On the same day, a separate US executive order directed federal agencies, within legal limits, to use “Super Intelligence” and “SI” in place of “artificial intelligence” and “AI” in specified official materials. That terminology order did not itself impose the accord's controls on industry.

  2. Key highlights2 / 4

    The accord sets out four organisational layers:

    • Internal controls: Monitor model capabilities and behaviour during training and use, including cybersecurity, biological and chemical risks and unintended access to technical systems.
    • Internal safety team: Check whether controls and monitoring work and whether identified problems are addressed.
    • External evaluation: Use an independent auditor or evaluator to assess the controls and monitoring.
    • Board oversight: Give an independent board committee reports from the teams and evaluators, and responsibility for overseeing responses to problems.

    The text does not prescribe a common audit method, public release of findings or a new duty to report incidents to a regulator. Those are limits of this accord, not a claim that its signatories have no other legal duties. Multiple layers may help identify failures, but their value depends on how the companies implement them and on evaluator independence.

  3. Key concepts3 / 4

    Voluntary commitment and binding law

    A voluntary pledge can guide company practice without creating legal duties by itself. Legislation can impose enforceable obligations and penalties. Either approach still depends on clear standards and effective implementation.

    Frontier AI and AI agents

    Frontier models are among the most capable general-purpose AI systems. An AI agent can use tools or access other systems to carry out tasks. Giving a system permission to act raises questions about access limits, monitoring and human control.

    EU AI Act

    The European Union has a binding AI law. Since 2 August 2025, its general-purpose AI model rules have included technical documentation, information for downstream providers, a copyright-compliance policy and a public summary of training content, subject to applicable provisions and transitions. Providers of models with systemic risk face further duties including risk assessment, serious-incident reporting and cybersecurity. The EU framework provides for enforcement and fines; it is not a voluntary pledge.

    Independent audit

    An outside assessment can test a firm's own claims. It cannot remove every conflict of interest automatically: the auditor needs access, competence and independence, and findings need a route to corrective action.

  4. Way forward4 / 4

    The 2025 India AI Governance Guidelines set out principles for safe and inclusive AI use. They work with existing laws and sector regulators rather than creating a new horizontal AI statute at this stage. The IndiaAI Mission also supports domestic AI capability and responsible development.

    Possible steps for future policy include:

    • Define measurable tests for model security, privacy and other serious harms.
    • Set proportionate rules for reporting serious incidents and acting on them.
    • Protect evaluator independence and require clear routes for corrective action.
    • Limit an AI agent's access to the tools and data needed for its task; provide monitoring and human override.
    • Clarify responsibility across developers and organisations using AI systems.
    • Cooperate internationally on compatible methods and evidence for safety assessment.

    These are policy options, not commitments made by the accord or a statement that India has adopted each measure.

Sources

Syllabus

PaperSubjectSub-topic
PrelimsScience & TechnologyArtificial intelligence and voluntary safety commitments
GS3Science & TechnologyEmerging technology and AI governance
GS2International RelationsEffects of developed-country technology policies on India

Topics

GovernanceInformation and Communication TechnologyMiscellaneous Science and Technology

Related previous-year questions

Asked in earlier UPSC Prelims papers on this topic. Answer, then check.

  1. UPSC Prelims 2026 · Science and Technology · Information and Communication Technology

    Consider the following statements with respect to the AI Impact Summit, 2026 held in New Delhi: 1. The Summit's intellectual framework was based on three foundational Sutras: People, Planning, and Progress. 2. The Preamble of the Summit stresses Democratising AI Resources, which acknowledges the Charter for Democratic Diffusion of AI as a binding framework to support locally relevant innovation and strengthen resilient AI ecosystems while respecting national laws. 3. The New Delhi Declaration on AI Impact was structured around seven Access Chakras (Pillars), which included Access for Social Empowerment, AI for Science, and Secure and Trusted AI. Which of the statements given above is/are correct?

    1. 1, 2 and 3
    2. 1 and 2 only
    3. 2 and 3 only
    4. 3 only
    Show answer

    Answer: D. The AI Impact Summit 2026, held in New Delhi, focused on shaping the future direction of artificial intelligence through an inclusive and development-oriented approach. The summit adopted an intellectual framework based on three guiding Sutras - People, Planet, and Prosperity. It also emphasized the democratization of AI resources and introduced the New Delhi Declaration on AI Impact, which was organized around seven thematic Chakras or pillars addressing key areas of AI development and governance. Statement 1 is not correct: The summit framework was built around the three guiding principles of People, Planet, and Prosperity, rather than People, Planning, and Progress. Statement 2 is not correct: While the preamble emphasizes the democratization of AI resources, the Charter for Democratic Diffusion of AI was envisaged as a voluntary and non-binding framework rather than a legally binding arrangement. Statement 3 is correct: The New Delhi Declaration was structured around seven Chakras including Access for Social Empowerment, AI for Science, and Secure and Trusted AI. The declaration incorporated seven thematic pillars covering areas such as these as part of its broader framework. Only statement 3 is correct, so the answer is Option (d).

    Difficulty: medium · statement

    Open this question on its own page, with the full explanation →

  2. UPSC Prelims 2025 · International Relations and Current Affairs · Current Affairs

    Consider the following statements regarding AI Action Summit held in Grand Palais, Paris in February 2025: I. Co-chaired with India, the event builds on the advances made at the Bletchley Park Summit held in 2023 and the Seoul Summit held in 2024. II. Along with other countries US and UK also signed the declaration on inclusive and sustainable AI. Which of the statements given above is/are correct?

    1. I only
    2. II only
    3. Both I and II
    4. Neither I nor II
    Show answer

    Answer: A. Statement I – CORRECT: The AI Action Summit held at the Grand Palais in Paris on February 10–11, 2025, was co-chaired by France and India. It built upon the progress made at the Bletchley Park Summit in 2023 and the Seoul Summit in 2024, aiming to advance global cooperation on AI governance and development. Hence Statement I is correct. Statement II – INCORRECT: While the summit introduced a declaration on 'inclusive and sustainable' AI, the United States and the United Kingdom declined to sign it. Their refusal was based on concerns over national security and the belief that the declaration did not adequately address global AI governance. Hence Statement II is not correct. Only Statement I is correct → Option (a).

    Difficulty: easy · statement

    Open this question on its own page, with the full explanation →

  3. UPSC Prelims 2026 · Science and Technology · Information and Communication Technology

    Which of the following statements with regard to Large Language Models (LLMs) used in machine learning is/are correct? 1. LLMs assign probabilities to the next possible words and then pick the one with the highest probability. 2. LLMs process data through mathematical optimization to minimise prediction errors. 3. LLMs produce unbiased outputs. Select the answer using the code given below:

    1. 1 only
    2. 1 and 2 only
    3. 2 and 3 only
    4. 1, 2 and 3
    Show answer

    Answer: B. Large Language Models (LLMs) like GPT work by predicting the next token or word in a sequence. They assign probability distributions over possible next words and typically select based on highest probability, or use sampling strategies. LLMs are trained through mathematical optimization processes (gradient descent, backpropagation) that minimize a loss function, which measures prediction errors between the model output and actual training data. Statement 1 is correct: Large Language Models assign probabilities to the next possible words and then pick the one with the highest probability. This is the core mechanism of autoregressive language models, which generate text by predicting the most likely next token. Statement 2 is correct: LLMs process data through mathematical optimization to minimise prediction errors. During training, LLMs use optimization algorithms like stochastic gradient descent to minimize the cross-entropy loss (prediction error). Statement 3 is incorrect: LLMs are well-documented to produce biased outputs reflecting biases present in their training data. They are trained on large corpora of text data from the internet that contain inherent biases related to gender, race, culture, and other factors. This is a widely recognized limitation of current LLM technology. Statements 1 and 2 are correct, so the answer is Option (b). The incorrectness of the third statement can also be used to eliminate options (c) and (d).

    Difficulty: medium · statement

    Open this question on its own page, with the full explanation →

Practice questions

  1. With reference to the White House Accord on Super Intelligence of 2026, consider the following statements: 1. It is a voluntary commitment that does not itself create statutory penalties. 2. It calls for four layers of controls and audits. 3. It establishes a federal authority to license frontier AI models. Which statements are correct?

    1. 1 and 2 only
    2. 2 and 3 only
    3. 1 and 3 only
    4. 1, 2 and 3
    Show answer

    Answer: A. The accord proposes internal controls, a safety team, external evaluation and board oversight. It does not itself create a licensing authority or statutory penalties.

    Difficulty: medium · statement

  2. With reference to the EU AI Act, consider the following statements: 1. It includes binding obligations for providers of general-purpose AI models. 2. A public summary of training content is among those obligations. 3. It operates only as a voluntary industry pledge. Which statements are correct?

    1. 2 and 3 only
    2. 1 and 2 only
    3. 1 and 3 only
    4. 1, 2 and 3
    Show answer

    Answer: B. The EU AI Act is binding legislation. Applicable providers must meet its requirements, including a public summary of training content; systemic-risk models have further duties.

    Difficulty: medium · statement

  3. Consider the following statements: 1. A voluntary pledge may guide conduct without itself creating a legal obligation. 2. AI agents can carry out tasks by using tools or accessing systems. 3. An external audit automatically removes all conflicts of interest. Which statements are correct?

    1. 1 and 3 only
    2. 2 and 3 only
    3. 1 and 2 only
    4. 1, 2 and 3
    Show answer

    Answer: C. The first two statements are correct. An auditor still needs independence, competence, access and a route for acting on findings.

    Difficulty: medium · statement

Mains practice

Answer-writing practice on this article. Attempt it first, then open the hints.

  1. GS3 · 250 words

    Compare voluntary industry commitments with binding statutory rules for frontier AI safety. What should guide India’s approach?

    Show hints
    1. Compare flexibility and speed with enforceable duties and penalties.
    2. Explain the accord’s four layers and the practical limits of self-assessment.
    3. Describe the EU AI Act’s duties for general-purpose AI and systemic-risk models.
    4. Discuss India’s 2025 governance guidelines, existing laws and sector regulators.
    5. Suggest proportionate standards, independent evaluation and accountability.
  2. GS3 · 250 words

    AI agents can act through tools and connected systems. Discuss the safeguards needed when their actions may affect people or infrastructure.

    Show hints
    1. Distinguish generating an answer from carrying out an action.
    2. Limit tool and data access to what each task requires.
    3. Use monitoring, testing, human approval or override where risks warrant them.
    4. Provide ways to stop activity and respond to serious incidents.
    5. Allocate responsibilities among developers, deployers and users.