Should AI speech be limited?

How AI speech is defined can affect the discussion. I believe it should be defined broadly as AI-generated content that performs a communicative function. The AI speech can be concluded into three categories: The model-led-speech is generated by the model itself, such as the initial responsible in a conversation; The human-led-speech is led mainly by users, such as content for which users employ AI to turn their ideas into the ideal AI speech. The mixed-speech is a mix of users’ ideas and model’s contribution. While they often coexists in a chat, it is worth discussing them separately.

The model-led-speech shouldn’t be limited beyond the safeguards had already imposed on the mode. To be specific, this content mainly comes from its training rather than clear guidance from the users. The argument that model-led-speech should be limited can be formed:

1. Nobody is responsible for these instances of AI speech

2. If such speech causes harm, nobody is going to take the moral responsibility

3. So we should limit these AI speech.

Premises 1 and 2 can be revealed through the evident question – psychological harm – The model might mislead and harm the users when it attempts to give psychological help, including the personalized persuasion, instead of only providing objective advice. For instance, the AI might make commitments like physical affection or love, that only actual humans can fulfill, during the conversation. Some may argue that humans may do the same, but the difference is the human takes the moral responsibility for their speech while no one appears to be responsible for AI speech in this case. It is supported by Millican’s argument against attributing consciousness to machines, so that they don’t have the intention to deceive the users. Some argue that the AI itself is responsible. But it also fails, as it still doesn’t explain how an AI system could bear moral responsibility. If we trace back the development of AI, we find that many people involved in the training process, the one deciding the training method, the one selecting data used… Moreover, the contemporary AI training is more likely to be a “black-box” rather than a clear and transparent procedure. It functions as an integrated whole that you cannot break the outputs down to components. Another response is that the response could be a matter of chance. AI is probabilistic model that would output the most “probable” answers. Nowadays, even with the same setup, the same model, and the same settings, AI could generate different, even opposite answers. So the producers of AI couldn’t actually take the moral responsibility of the model-led AI speech, if the model is checked and permitted. However, the conclusion doesn’t lead to the limitations. Because premise 2 will only be valid under the assumption that the user always takes the AI speech as truth. This only reveals that we shouldn’t take AI as a human and expect it to take the moral responsibility of their speech, we should instead consider it as a powerful and convenient source of information. In “A Theory of freedom of expression”, Scanlon argues that the state should respect listeners as capable of deciding what to believe. In a similar way, when we are dealing with the freedom of AI speech, we should respect users as capable of deciding which AI speech to believe and how AI speech is used. I suggest that the users could take the moral responsibility if they aim for the acknowledgement of the AI speech. For the “know” here, It would be suitable to use the counterfactual account of knowledge:

You knows “the AI speech” if and only if

1. “the AI speech” is true

2. You believed in “the AI speech”

3. If “the AI speech” is not true then you won’t believe it.

In this way, the model-led speech will be beneficial, and it thereby undermines premise 2, as the users will be the one that takes the moral responsibility here. So the conclusion 3 fails, there aren’t enough reasons to support the limitation of model-led-speech.

The human-led-speech is the category that should be limited. It is different from the former: as the speech were used to express users’ ideas and showed their intentions. I believ that the users directing the speech should take the moral responsibility for the influence of this AI speech. As we just discussed, AI shouldn’t be attributed any specific intention, only the users and AI involved in the speech, the relevant communicative intention comes from the uses . It could be considered as the users’ speech. While AI lowers the barriers for people to present ideal speech, it may also increase the influence of evil. There are several real-world cases: the hackers used AI to hack a church website and obtain the personal details of people registered in the site. They defraud people of their money using this information. The consequences extend beyond fraud. The church, as an example, would have to spend more money to build up the firewall defending the hackers, this increases the cost of building a website and running a beneficial organization, because you will have to spend more time to maintain and protect your site. This is extra work influenced by AI. Last but not least, the development of video-generated AI is also a double-edged sword — the digital version of you can be created easier and more vivid using your information. It will be harder to clarify, if evil tries to use the image or video to default people. This creates an additional need for verification. As the human-led-speech benefits the creation, it is worth recognizing the risk that coexists with those benefits . I suggest the human-led-speech should be limited. The limitation is more straightforward as some people need to take moral responsibility. To protect privacy, we should clarify the AI speech, before identifying identity before risky approaches, such as the video generated based on real humans and prohibit dangerous use, such as the malicious codes for the evils.

The third kind of AI speech, the mixed-speech, is a common case when you are discussing new topics with AI. At first glance, its classification may seem ambiguous that both sides contribute to the speech. However, I argue that it is a mistake to treat it as a distinct category. It should be included in human-led-speech. Some may argue that these prompts may be influenced by AI. But this fails, as if we follow the idea of model-led-speech, the users will know the AI response and then give a new prompt. In this perspective, humans are always leading the speech.

In conclusion, I suggest the freedom of AI speech, based on Scanlon’s argument, and I propose that we should not treat AI speech as human speech because AI itself lacks moral responsibility. I recategorize the three kinds of speech, and clarify that the mixed speech are included in the human-led speech. Finally, I suggest we shouldn’t limit the model-led speech under the assumption that the AI model has been tested and judges reasonably safe and users should be expected to evaluate and know the AI speech, and the necessity to limit the human-led speech by identification and stricter controls, because AI speech can a,plify human intentions.