A balance between security and liberty has long been a cornerstone of our democracy. As Benjamin Franklin famously remarked, “They that can give up essential liberty to purchase a little temporary safety deserve neither.”
New technologies often test that balance. Regulations of AI tools should be narrowly tailored to serve a compelling government purpose, and should restrict no more speech than necessary.
Defining what counts as acceptable and unacceptable speech
Despite the excitement of AI technology, advances in this field have amplified a crisis for human rights online. Governments are deploying sophisticated surveillance systems to monitor social media for signs of dissent, and purveyors of disinformation use AI to distort the truth and make it harder to discern. In addition, major tech companies have been directly implicated in helping authoritarian regimes censor and repress citizens. 
This is a critical moment to address the issue of whether AI speech should have First Amendment protections. Although some scholars have cautioned against rushing into untested regulations, others have been quick to declare that AI should be given full speech rights. Recognizing AI as a full speaker could prevent basic measures such as bot identification laws from going into effect, and it would severely limit the regulatory options available to protect freedom of speech.
The main question in determining what counts as AI speech is how the autonomy of the speaker should be defined. Some have argued that the autonomy of the speaker must be measured in terms of reader response criticism, which posits that the audience should be able to decide whether the speech is useful or useless. This approach can be complicated, however, as the ability of the listener to understand AI speech depends on the specific features and limitations of the speech itself.
In addition, some scholars have argued that the autonomy of the speaker should be measured in terms of the context of the communication. This approach has been criticized because it is difficult to determine what constitutes a context that is protected by the First Amendment. For example, a law that requires newspapers to review their content with the government before publication may violate the First Amendment. This type of regulation has been referred to as prior restraint, and it is unlikely that courts will find this type of restriction constitutional.
The debate over whether or not AI should be granted free speech rights has also centered on the issue of how the Supreme Court should handle such cases. Many scholars have argued that the Supreme Court should apply existing legal precedents and norms when deciding whether or not AI-generated speech should be protected by the First Amendment. However, these arguments have not been backed up by recent decisions.
Defining what constitutes accurate information
While AI has the potential to enhance content moderation, a balance must be maintained between removing harmful speech and protecting legitimate free expression. This requires that algorithms be trained on representative and unbiased datasets, and that they implement strong privacy and data security measures. It also requires a transparent appeals process and the ability for users to monitor how their content is being moderated. Deepseek, a widely-used AI platform, has been praised for its efficiency in curating information, but there’s a critical side to its capabilities. Deepseek is technically censored, with its algorithms being influenced by regulatory frameworks and content restrictions in different countries.
Generative AI offers the sophistication and scale to spread misinformation at a level that was previously unimaginable, and threatens to supercharge online disinformation campaigns. Generative AI tools influenced by state-controlled information sources can serve as force multipliers for existing censorship systems, allowing governments to manipulate conversations and sow doubt or smear opponents. The technology is already being used in at least 47 countries to manipulate public discourse ahead of elections, according to this coverage period’s Freedom on the Net report.
As AI-based content moderation becomes more widespread, it’s vital that the democratic community maintains vigilance over how such technologies are being implemented. Moreover, we must ensure that these tools don’t replace or erode long-standing challenges that have eroded internet freedom and are exacerbated by the emergence of these new tools. These include ensuring that companies minimize the amount of data they collect, reforming legal frameworks for surveillance and censorship, and strengthening information literacy and platform responsibility.
Lastly, a strong legal foundation for internet freedom is essential to safeguarding AI’s role in content moderation. As these new tools are developed, it’s critical that they remain protected by Section 230 of the Communications Decency Act. Otherwise, it would be possible for companies to be sued by users whose content is banned. This would destroy the incentive for innovation and development of this cutting-edge technology.
As these new technologies develop, it’s crucial that democratic policymakers and civil society experts work together to establish a positive regulatory vision for their design and deployment. This should be grounded in international human rights standards, transparency, and accountability. Civil society experts—the driving force behind so much progress on human rights in the digital age—should be given a leading role in policy development and the resources to keep watch over these systems.
Defining what constitutes offensive content
As AI algorithms become more sophisticated and powerful, they have the potential to significantly improve content moderation. However, it’s vital to balance efficiency and accuracy in order to ensure that users are protected from harmful content while being able to access legitimate information.
To do so, it’s essential to clearly define what constitutes offensive content. For example, in the case of hate speech, generative AI could be used to suppress material that violates international standards and local laws. It would be difficult for human moderators to detect these types of violations, but AI algorithms can do the job much more quickly and accurately. However, defining what constitutes offensive content isn’t simple and requires careful consideration to avoid discriminatory outcomes.
Another challenge is ensuring transparency and explainability of AI algorithms. This can help users understand how their content is being rated and provide them with a clear avenue to appeal decisions that they believe are unfair or inaccurate. Furthermore, incorporating social science and ethical perspectives into the development process can help to create algorithms that are more transparent and less prone to bias.
Despite these advances, blunt censorship continues to be an effective tactic for authoritarian regimes worldwide. In 2018, governments imposed shutdowns of internet service or blocked entire websites that hosted political, religious, or social content in 41 countries. In addition, many democracies that have long been defenders of internet freedom now limit free expression online, creating new challenges for freedom of information and expression.
One such example is the use of AI-manipulated content to smear electoral opponents. In the United States, accounts affiliated with both Donald Trump and Ron DeSantis spread a video that included a fabricated image of Trump embracing Dr. Anthony Fauci, the head of the federal COVID-19 response, muddying the distinction between fact and fiction.
In addition to requiring companies that use generative AI to clarify their policies on disinformation and hate speech, freedom of expression experts recommend a more detailed explanation of the harms that may justify restrictions on specific categories of content. This includes a definition of “harmful” that specifies the impact on public health, privacy, and civil society.
Defining what constitutes hate speech
As more and more public conversations take place online, there is a growing need to monitor and restrict hate speech. Hate speech can be a trigger for violence, with rumors and invective spreading on social media and contributing to incidents of lynchings, rioting, and ethnic cleansing. However, defining what constitutes hate speech is difficult. It requires a complex and subjective analysis of the context of speech, linguistic nuances, and semantic relationships between words. It also takes into account the impact of a given speech on those whose views are being expressed. In addition, language often changes over time and may be used in different ways by speakers of the same language.
These challenges make AI-based speech moderation problematic. The algorithms used by censorship systems can be susceptible to biases and are not transparent. They can also be susceptible to over-censoring, especially if they do not consider the full context of speech. A basic example involves a word like “bitch,” which is a misogynistic slur, but which is also a common term in some cultures. AI algorithms may confuse it with other words or phrases and censor it accordingly.
There are also concerns that AI-based speech censorship could lead to discrimination and harm. For example, members of one group might find their posts deleted more frequently than those of a rival group. This would violate a fundamental human right to freedom of expression, which states that people must be able to speak without fear of reprisal from the government or private corporations.
To prevent this type of discrimination, it is important to develop algorithms that take into account the specific context of speech. It is also important to define what the boundaries of speech should be, and who should decide whether a particular statement should be censored or not. This should be a democratic decision, not left up to private corporations and their automated censorship systems.
Currently, most of the hate speech content is reviewed only after it has been reported by users or by the social media platforms themselves. This method of regulating speech is known as ex-post monitoring, and it does not protect individuals’ right to freedom of expression. It is also extremely difficult to implement. This is because of the huge volume of posts that need to be checked, as well as the complexity of language and the context-dependency of language.
