# AI Safety Needs Transparency and Trust, Not Dictators > Source: > Published: 2026-09-17 12:38:00+00:00 Donald Trump most recently [proclaimed](https://www.nytimes.com/2026/09/14/us/politics/trump-ai-regulation-anthropic-dario-amodei.html) that the only guardrail AI needs is a “strong and smart” president. This dismisses not only warnings about AI, but also the democratic process by which our country is governed. At the same time, his administration’s June Executive Order 14409 creates a voluntary process through which selected frontier-model developers may give the federal government access to models before wider release and work with government-selected “trusted partners.” Thus, our community’s safety would be in the hands of a few selected “strong” and “smart” experts. However, that is not safety. It is the replacement of one opaque system with another. It’s like giving the Mafia the keys to your house to protect you from the cartels. Advanced AI can assist cybercriminals, expose critical infrastructure, accelerate vulnerability discovery, and concentrate extraordinary power in a small number of companies. Society and our institutions cannot simply look away. However, putting a president, an intelligence service, or a hand-picked circle of approved firms at the center of AI safety does not solve the legitimacy problem. It deepens it. Safety cannot rest on a strongman’s personal judgment. It needs institutions that can be inspected, challenged, changed, and, ultimately, [trusted](https://www.korte.co/2026/05/14/board-brief-open-source/). ## AI Safety Is Not Obedience Its simplicity makes many of us susceptible to authoritarian fantasies. A dangerous technology appears and challenges our worldview. Consequently, a figure promises to control it and protect us from it. Thus, we can relax and trust the leader to become the guardrail. Unfortunately, AI systems do not operate in a vacuum. They are built from data, optimized by private companies, deployed by employers and public agencies, run on vast cloud infrastructure, and increasingly embedded in the everyday decisions that shape people’s lives. Can the president personally supervise all these interactions? Thus, a claim that safety comes from executive strength asks the public to trust power. It’s a small step from demanding trust in Big Brother. It requires confidence in decisions that may be classified or politically convenient. It tells citizens that the remedy for black-box AI is a black-box state. In short, it demands obedience. The [June order](https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/) does contain useful recognition of the problem. It directs agencies to strengthen cyber defenses and prioritize enforcement against AI-assisted criminal access to computer systems. Yet it also calls for a classified benchmarking process. The process will determine which systems count as “covered frontier models,” while framing pre-release government access and limited access for trusted partners as voluntary. Secrecy may sometimes be necessary. However, a safety regime built primarily around classified tests and privileged access cannot earn broad public trust. It can tell the public that a system has been examined and aligns with political goals. It cannot reliably show what was examined, what risks were found, and what trade-offs were accepted. Crucially, it hides whether any outside incentives changed the examiner’s findings. ## AI Transparency Builds Legitimate Safety Open Source is the ultimate way to establish trust and transparency. Yet, in our world of varying software licenses, transparency should not require publishing every model weight, every security vulnerability, or every method. Yet, at a minimum, transparency requires us to make the safety process intelligible and contestable. If a powerful AI model is assessed for cyber risk, the public should be able to understand the assessment criteria at a meaningful level. Independent experts should be able to evaluate the methodology. Researchers should be able to challenge claims. Affected communities should be able to know when and where automated systems are being used against them. Regulators should disclose conflicts of interest, enforcement outcomes, and the evidence behind major decisions. This is harder than declaring that a president will take care of it. It is also more durable. Trust emerges when institutions accept scrutiny. A company that opens its model for external testing gives the public something concrete to judge. A government that publishes findings and opens avenues to appeal them limits its ability to turn “security” into a synonym for political control. The alternative is a familiar pattern. Governments call for security, then turn security into secrecy. Companies call for responsibility, then use responsibility as a reason to lock competitors, researchers, and users out of essential information. The only result of secrecy is a more concentrated AI sector whose power is harder to inspect. This is especially dangerous when “trusted partners” are selected through government-industry arrangements rather than open, accountable standards. Trust should not be a membership card issued to the largest vendors or watered down for the biggest ballroom donors. Everyone should have the same avenue to earn public trust through verifiable practices. ## Digital Sovereignty Requires AI Agency Without the ability to inspect and challenge findings, there can be no digital sovereignty. After all, genuine digital sovereignty requires agency. It means that institutions, communities, and individuals retain some capacity to understand and govern the technologies on which they depend. Without understanding, it becomes difficult to avoid building dependencies on foreign cloud providers or on domestic firms operating behind the shield of state secrecy. Trump’s December 2025 executive order, which seeks a minimally burdensome national AI framework and directs the federal government to challenge certain state AI laws, presents uniformity as an answer to regulatory fragmentation. A patchwork of fifty conflicting rules can indeed burden smaller firms and create confusion. But centralization is not automatically sovereignty. A single national framework can become a single point of capture if it weakens local accountability, sidelines state experimentation, and treats disclosure as an obstacle rather than a democratic necessity. The question is not merely who governs AI. The question is whether those who are governed can see, contest, and shape the rules. A digital-sovereign approach would invest in public technical capacity, independent auditing, interoperable infrastructure, procurement transparency, and institutions able to evaluate systems without being dependent on the vendors selling them. It would refuse the false choice between Silicon Valley’s private monopolies and Washington’s centralized command. ## Open Source Is Part of the Answer Open Source is not automatically safe. Open models can be misused, and publishing code does not erase the need for security review, red-teaming, access controls, or responsible disclosure. Pretending otherwise would be naïve. But Open Source remains essential to trustworthy AI because it enables independent verification. It allows universities, civil-society groups, smaller companies, public agencies, and technically capable communities to inspect tools rather than merely accept vendor assurances. It supports reproducibility. It reduces the absurdity of asking the public to trust a safety claim made by the same corporation that profits from the system being declared safe. The strongest AI safety system will not be the one that gives the most discretion to a dictator, a chief executive, or a national-security bureaucracy. It will be the one that makes power visible, distributes the capacity to verify claims, protects genuine security interests without turning them into permanent secrecy, and gives the public a real role in governing the systems that increasingly govern them. AI does not need a dictator. It needs democratic institutions capable of saying: show us how this works, show us what can go wrong, show us who benefits, and show us who is [accountable](https://www.korte.co/2026/04/23/stop-blaming-the-machine-for-ai-failings/) when it fails.