Limited Intelligence: How Civil Society Is Trying to Take Control of AI

The break in the contract between the Pentagon and Anthropic has become one of the biggest scandals around the ethics of the use of AI technology. The more rapidly the artificial intelligence market is developing, the more acute the question of ethical limitations of digital algorithms is. Most states have not been able to develop legislation governing AI technology, and their developers are resisting any attempt to impose restrictions on their work. While states fail, civil society takes care of this burden: we are talking about such tools as civil audit of AI technologies, the development of ethical codes and licenses, collective testing and pressure through the media.

E.Good Mistakes and Who Will Correct Them

The scale of the introduction of AI systems in business processes is already such that the problem of “biasing” of algorithms begins to seriously affect society: there are cases when when hiring artificial intelligence weed out the resume of women, mistakenly marked the black defendants as people with a high risk of relapse or refused medical care to the elderly.

Another problem is deepfakes (PIs generated by AI that can be offensive or pornographic). In April, the Paris prosecutor’s office summoned for questioning the businessman and owner of the social network X (Twitter) Elon Musk in connection with accusations of distributing diplomas with child pornography his chatbot Grok.

Against the background of such scandals and tragic mistakes, it is increasingly obvious that states lag behind in matters of rules for AI systems. Traditional government regulation in this case is faced with the “collegridge dilemma”: in the early stages, the consequences of the introduction of the new technology are difficult to predict, and in the later, when the harm is obvious, control becomes too expensive and complex.

Democratic states (both the EU, and the US, and all the others) did not take into account such a rapid development of AI models. So, for example, Louise Jarywski считаетbelieves that even the laws adopted in 2022-2024 (as, for example, in the European Union – the AI Act) are already outdated and ineffective.

At the same time, most regulators have been paralyzed over the past three years and are afraid to take decisive action, fearing that regulation can suppress innovation and slow down the growth of productivity (an argument that is often sounding in the EU) or that this will lead to a loss in the “ray of AI” and the strengthening of rivals (that is, China – this is what the United States talk about).

Louise Jarywski notes that the noise around AI, overly optimistic predictions about the creation of Artificial General Intelligence (AGI), comparable to human or superior to it and able to cure all diseases for a decade, as well as discussions about “self-awareness” and “well-being” of generative models distract from discussing real problems. And against the background of the interference of the authorities, technology companies begin to act according to their own rules, and there is no question of self-regulation here.

Neither AI companies nor Silicon Valley technology startups are ready for self-restraints in principle

Because AI companies are primarily driven by business interests, they are willing to support any narratives that make a profit, even if it means promoting ideas like implanting chips for people to increase intelligence and productivity.

In this situation, the burden of regulation falls on non-state mechanisms: from ethical codes of professionals to independent audits carried out by the forces of the communities. The industry is developing standards that affect, among other things, ethical aspects of technology development.
Ethical Codes and Age Ratings

The updated Code of Ethics of the International Association of Computer Science (ACM) includes provisions that the development of technology should contribute to the “public good”, and engineers must take into account the interests of all people – as a kind of industry shareholders.

The document also encourages the creation of mechanisms of “informing violations” if a company ignores the moral aspects of its work. It is worth noting, however, that the ACM states: “The Code is created to inspire professionals in computer technology on an ethical approach,” that is, the document does not oblige anything.

The American Institute of Electrical and Electronics Engineers (IEEE) proposed the methodology of Ethically Aligned Design. The organization has created a series of P7000 standards that explain how to achieve transparency in the code, prevent the bias of the algorithms and to ensure the protection of personal data when working with AI. However, these standards are not binding until they are incorporated into the law.

Another idea for non-state control of AI tools is to create a system of voluntary ratings in the industry, similar to age restrictions in cinema. The German AI Ethics Impact Group has proposed an ethical labeling model (like energy efficiency classes) that evaluates the system according to six parameters: transparency, accountability, privacy, reliability and environmental sustainability. For measurement, a special VCIO (Values, Criteria, Indicators, Observables) model is used, which breaks these relatively abstract concepts into specific, verifiable indicators.
Collective testing and civil audit

Due to the peculiarities of their device, AI models are not always amenable to traditional testing methods. Here, pays for help are used by career-based payroll programs bug bountiesfrom the cybersecurity sphere, adapted to search for prejudices – bias bounties. The first large-scale program of this kind was conducted in 2021 by Twitter (now X) for its image trimming algorithm. Participants found that the tool prefers younger people with light skin.

Non-profit organizations, such as Bias Buccaneers and Humane Intelligence, organize such checks as transparent competitions, involving not only data experts, but also sociologists and activists. This approach allows you to identify “unknown unknowns” (unknown unknowns) – risks that the developers themselves did not know about.

Organizations like the Eticas Foundation develop the Community-less Audits methodology, which is necessary in cases where algorithms directly affect social status and human rights. This approach combines technical methods with sociological – interviews and ethnography. In Rotterdam, a public audit revealed a bias of an algorithm that unfairly deprived of benefits from single mothers and migrants.
Ethical licenses

Another control tool has arisen within the community of users and developers of free software (Open Source) – these are “Responsible AI Licenses or RAIL). Unlike regular open licenses that allow you to use the code as you like, RAIL includes ethical requirements.

The Licensee undertakes not to use the model for harmful purposes: for example, to create deepfakes, uncertified medical advice or discrimination in one form or another. This allows developers to maintain control over how their discoveries affect the world.
Journalism and Grassline activism

An important link in the system of public supervision of AI was investigative journalism. Publications like The Markup and ProPublica specialize in “opening” algorithm problems. In 2016, ProPublica in its material Machine Bias exposed the racial bias of the COMPAS system used in American ships. This technology “predicted” crimes based on human data (such tools are prohibited in the EU).

Journalists received risk assessments appropriated to more than seven thousand arrested in Brovard County, Florida, in 2013 and 2014, and checked how many of them were charged with committing new crimes over the next two years.

The assessment proved to be surprisingly unreliable in the prediction of violent crimes: only 20% of the representatives of the “risk group” really committed them. In 2025, ProPublica found that the U.S. Department of Government Efficiency (DOGE) Elon Musk cut support for veterans, based on erroneous artificial intelligence data.

In 2025, the Government’s Department of Efficiency (DOGE) Elon Musk cut support for veterans, based on erroneous artificial intelligence data

Sometimes the pressure on AI companies are trying to exert their employees themselves. An example is the No tech for Apartheid initiative launched by Google and Amazon workers protesting against the use of their products by the Israeli army. Ordinary users are also trying to influence the situation through the boycotts. After the Pentagon scandal, Anthropic and OpenAI in the United States unfolded the campaign “Jack from ChatGPT” (Quit ChatGPT), which has already gained more than a million participants.

Anyway, even if such public initiatives were ten times more, this is still not enough to solve fundamental issues of control over artificial intelligence, which is possible only by the forces of states or international organizations. Many experts wonder whether AI is worth regulating as strictly as nuclear weapons, whether state standards are needed to develop and train new models, and even “to what extent we can allow AI to break the social institutions that have been formed for years.”