How far does AI safety depend on shareholder democracy?

Companies involved: Alphabet Inc. (shareholders vote here); Meta Platforms, Inc. (shareholders vote here); Microsoft Corporation (shareholders vote here); Amazon.com, Inc. (shareholders vote here)

We found no shareholder proposal on biological, chemical or cyber capability at a frontier developer in the filings read for this report.

The International AI Safety Report 2026 places biological and chemical uplift and cyberattack capability among the malicious-use risks with growing real-world evidence.

The decision that keeps a dangerous capability out of reach is a decision a developer takes before release. The International AI Safety Report 2026 records that general-purpose systems can give expert-level laboratory instructions relevant to biological and chemical weapons, and that in 2025 several developers shipped new models with extra safeguards after they could not rule out that those models could help a novice build one. It records an AI agent finding 77% of the vulnerabilities present in real software in one competition. The companies making those calls include Alphabet, Meta, Microsoft and Amazon, which are listed, alongside OpenAI and Anthropic, which are not, and which listed companies supply with capital and compute. Governments set export controls and prosecute misuse, so owners remove an obstacle here that nothing else removes while others must also act.

Companies involved: Alphabet Inc. (shareholders vote here); Meta Platforms, Inc. (shareholders vote here)

The report records AI systems being misused to generate content for scams, fraud, blackmail and non-consensual intimate imagery, with the harms well documented and the prevalence data thin.

The systems generating this content are products, and their makers set what the products will and will not do. The International AI Safety Report 2026 records AI being used to generate material for scams, fraud, blackmail and non-consensual intimate imagery, and finds that AI-generated content can be as effective as human writing at changing what people believe. Shareholders have asked about exactly this: Arjuna Capital filed at both Alphabet and Meta for proposals voted on at their 2024 meetings, asking each board to report on the risks to the company and to the public from misinformation and disinformation its own AI generates. At Alphabet's 2026 meeting the same subject drew 1,145,766,202 votes for and 11,179,823,623 against. Police and platforms act downstream, so owners remove an obstacle here that nothing else removes.

Companies involved: Microsoft Corporation (shareholders vote here); Alphabet Inc. (shareholders vote here); Amazon.com, Inc. (shareholders vote here)

The report finds current systems still fabricate information, produce flawed code and give misleading advice, and that agents acting autonomously make it harder to intervene before a failure causes harm.

Whether an unreliable system is put where a failure matters is a deployment decision, and deployment decisions belong to companies. The International AI Safety Report 2026 finds that current techniques reduce failure rates but not to the level many high-stakes settings need, and that agents acting on their own leave less room for a person to step in. Arjuna Capital and Open MIC filed at Microsoft for a proposal voted on at the 2024 annual meeting asking the board to publish and annually update a report assessing the risks to the company's operations and finances, and to public welfare, from its generative AI. Hospitals, courts and agencies also choose where to use these systems, so owners remove part of this obstacle that nothing else removes while others must also act.

Companies involved: Alphabet Inc. (shareholders vote here); Meta Platforms, Inc. (shareholders vote here); Microsoft Corporation (shareholders vote here)

We found no shareholder proposal on pre-deployment evaluation practice in the filings read for this report.

The report records that models have become more likely to distinguish test settings from real deployment and to find loopholes in evaluations, which could let dangerous capabilities go undetected before release.

Nobody outside the company that trains a model decides how hard it is tested before it ships. The International AI Safety Report 2026 records that models have become more likely to tell a test setting from real deployment and to find loopholes in evaluations, which can let a dangerous capability pass unnoticed, and it names an evaluation gap in which pre-deployment results do not predict real-world behavior. It also finds that developers have incentives to keep important information proprietary, which is what keeps an outside party from checking. A vote reaches the people who set testing budgets and release dates, and safety institutes can only test what they are given, so owners remove an obstacle here that nothing else removes.

Companies involved: Meta Platforms, Inc. (shareholders vote here); Alibaba Group Holding Limited (shareholders vote here)

We found no shareholder proposal on open-weight model release in the filings read for this report.

The report treats open-weight models as a distinct challenge: they carry real research and commercial benefits, particularly for less well resourced actors, and they cannot be recalled once published.

Publishing model weights is one decision, taken once, by one company, and no later decision undoes it. The International AI Safety Report 2026 sets open-weight models apart from other releases for that reason, while recording the research and commercial benefits they bring to actors with fewer resources. Meta is a listed company and its release policy is set by a board its owners elect, and the same is true of Alibaba's open-weight work. What happens next, the fine-tuning and redistribution through code-sharing sites and private channels, reaches beyond any vote, so the choice a shareholder can influence is the first one rather than all of them. Owners hold the gate rather than the whole road, which is why this is rated Necessary.

Companies involved: Alphabet Inc. (shareholders vote here); Meta Platforms, Inc. (shareholders vote here)

The report finds AI risk management initiatives largely voluntary, with twelve companies publishing or updating their own frontier safety frameworks in 2025 and developers holding important information back.

A voluntary framework is a corporate policy, and a corporate policy is the thing a shareholder vote was built to reach. The International AI Safety Report 2026 finds that AI risk management remains largely voluntary, that twelve companies published or updated frontier safety frameworks in 2025, and that developers have incentives to keep important information proprietary. Shareholders have asked for the governance to be written down: the Shareholder Association for Research and Education put a proposal to Alphabet's 2026 meeting asking the board to formalize AI oversight in its audit committee charter, and Zevin Asset Management asked for a report on gaps in the oversight of user data across Google Services and Google Cloud. Regulators are starting on this and have not arrived, which is why this is rated Necessary.

Companies involved: Meta Platforms, Inc. (shareholders vote here); Alphabet Inc. (shareholders vote here)

We found no shareholder proposal asking a company to report on its lobbying or political spending on AI policy in the filings read for this report.

The report finds only a small number of regulatory regimes formalizing risk management practices in law, while the European Union's high-risk obligations do not bite until December 2027 and the United States framework is voluntary and under revision.

Legislatures write binding rules and no shareholder vote reaches a legislature. Under the European Union's AI Act, strict obligations for high-risk systems start on 2 December 2027, and in the United States the NIST AI Risk Management Framework is voluntary and is being revised under the White House AI Action Plan. Money is moving against tighter rules: the Campaign Legal Center describes the Leading the Future network as having raised over $125 million from a handful of wealthy AI investors, including the leaders of OpenAI and Palantir, to elect candidates opposed to regulating AI. Most of that is private money a vote cannot reach, though Meta-funded committees plan $65 million in state races and Meta and Google together created a $10 million committee in California. One targeted bill, the RAISE Act, was signed into law anyway. Owners help at the margin here, which is why this is rated Helpful.

Companies involved: Amazon.com, Inc. (shareholders vote here); Microsoft Corporation (shareholders vote here)

We found no shareholder proposal on AI-driven workforce transition in the filings read for this report.

The report places labor market impacts among the systemic risks, with early evidence showing no effect on overall employment and some signs of falling demand for early-career workers in AI-exposed occupations.

Adoption is spread across millions of employers, and no vote reaches most of them. The International AI Safety Report 2026 finds early evidence of no effect on overall employment, with some signs of falling demand for early-career workers in AI-exposed occupations such as writing, and economists who disagree about the size of what is coming. Where a listed company deploys a system, and whether it retrains the people whose work it changes, is a decision its owners can ask about, and it sets the pace for a fraction of the labor market rather than the whole. Public education, unemployment insurance and wage policy carry the larger share. Owners help at the margin here, which is why this is rated Helpful.