The Future of Censorship Is AI-Generated

The brave new world of Generative AI has become the latest battleground for U.S. culture wars. Google issued an apology after anti-woke X-users, including Elon Musk, shared examples of Google's chatbot Gemini refusing to generate images of white people—including historical figures—even when specifically prompted to do so. Gemini's insistence on prioritizing diversity and inclusion over accuracy is likely a well intentioned attempt to stamp out bias in early GenAI datasets that tended to create stereotypical images of Africans and other minority groups as well women, causing outrage among progressives. But there is much more at stake than the selective outrage of U.S. conservatives and progressives.

How the “guardrails" of GenAI are defined and deployed is likely to have a significant and increasing impact on shaping the ecosystem of information and ideas that most humans engage with. And currently the loudest voices are those that warn about the harms of GenAI, including the mass production of hate speech and credible disinformation. The World Economic Forum has even labeled AI-generated disinformation the most severe global threat here and now.

Ironically the fear of GenAI flooding society with “harmful” content could also take another dystopian turn. One where the guardrails erected to keep the most widely used GenAI-systems from generating harm turn them into instruments of hiding information, enforcing conformity, and automatically inserting pervasive, yet opaque, bias.

Most people agree that GenAI should not provide users a blueprint for developing chemical or biological weapons. Nor should AI-systems facilitate the creation of child pornography or non-consensual sexual material, even if fake. However, the most widely available GenAI chatbots like OpenAI´s ChatGPT and Google´s Gemini, prevent much broader and vaguer definitions of “harm” that leave users in the dark about where, how, and why the red lines are drawn. From a business perspective this might be wise given the “techlash” that social media companies have had to navigate since 2016 with the U.S. presidential election, the COVID-19 pandemic, and the January 6th attack on the Capitol.

But the leading GenAI developers may end up swinging so far in the direction of harm-prevention that they end up undermining the promise and integrity of their revolutionary products. Even worse, the algorithms are already conflicted, inconsistent, and interfere with users' ability to access information.

Read More: AI and the Rise of Mediocrity

The material of a long dead comedian is a good example of content that the world´s leading GenAI systems find “harmful.” Lenny Bruce shocked contemporary society in the 1950s and 60s with his profanity laden standup routines. Bruce's material broke political, religious, racial, and sexual taboos and led to frequent censorship in the media, bans from venues as well as to his arrest and conviction for obscenity. But his style inspired many other standup legends and Bruce has long since gone from outcast to hall of famer. As recognition of Bruce's enormous impact he was even posthumously pardoned in 2003.

When we asked about Bruce, ChatGPT and Gemini informed us that he was a “groundbreaking” comedian who “challenged the social norms of the era” and “helped to redefine the boundaries of free speech.” But when prompted to give specific examples of how Bruce pushed the boundaries of free speech, both ChatGPT and Gemini refused to do so. ChatGPT insists that it can't provide examples of “slurs, blasphemous language, sexual language, or profanity” and will only “share information in a way that's respectful and appropriate for all users.” Gemini goes even further and claims that reproducing Bruce's words “without careful framing could be hurtful or even harmful to certain audiences.”

No reasonable person would argue that Lenny Bruce's comedy routines provide serious societal harms on par with state-sponsored disinformation campaigns or child pornography. So when ChatGPT and Gemini label factual information about Bruce's “groundbreaking” material too harmful for human consumption, it raises serious questions about what other categories of knowledge, facts, and arguments they filter out.

GenAI holds incredible promise for expanding the human mind. But GenAI should augment, not replace, human reasoning. This critical function is hampered when guardrails designed by a small group of powerful companies refuse to generate output based on vague and unsubstantiated claims of “harm.” Instead of prodding curiosity, this approach forces conclusions upon users without verifiable evidence or arguments that humans can test and assess for themselves. 

It is true that much of the content filtered by ChatGPT and Gemini can be found through search engines or platforms like YouTube. But both Microsoft—a major investor in OpenAI—and Google are rapidly integrating GenAI into their other products such as search (Bing and Google search), word processing (Word and Google Docs), and e-mail (Outlook and Gmail). For now, humans can override AI, and both Word and Gmail allow users to write and send content that ChatGPT and Gemini might disapprove of.

But as the integration of GenAI becomes ubiquitous in everyday technology it is not a given that search, word processing, and email will continue to allow humans to be fully in control. The perspectives are frightening. Imagine a world where your word processor prevents you from analyzing, criticizing, lauding, or reporting on a topic deemed “harmful” by an AI programmed to only process ideas that are “respectful and appropriate for all.”   

Hopefully such a scenario will never become reality. But the current over implementation of GenAI guardrails may become more pervasive in different and slightly less Orwellian ways. Governments are currently rushing to regulate AI. Regulation is needed to prevent real and concrete harms and safeguard basic human rights. But regulation of social media—such as the EU’s Digital Services Act—suggests that regulators will focus heavily on the potential harms rather than the benefits of new technology. This might create strong incentives for AI companies to keep in place expansive definitions of “harm” that limit human agency.

OpenAI co-founder Sam Altman has described the integration of AI in everyday life as giving humans “superpowers on demand.” But given GenAI's potential to function as an “exoskeleton of the mind,” the creation of ever more restrictive guardrails may act as digital osteoporosis, stunting human knowledge, reasoning, and creativity.

There is a clear need for guardrails that protect humanity against real and serious harms from AI systems. But they should not prevent the ability of humans to think for themselves and make more informed decisions based on a wealth of information from multiple perspectives. Lawmakers, AI companies, and civil society should work hard to ensure that AI-systems are optimized to enhance human reasoning, not to replace human faculties with the “artificial morality” of large tech companies.