Krauss: Artificially Intelligent Offense?

Of note, yet another concern and issue that needs to be addressed:

…Let’s be clear about this: Valid, empirically derived information is not, in the abstract, either harmful or offensive.

The reception of information can be offensive, and it can, depending upon the circumstances of the listener, potentially result in psychological or physical harm. But precisely because one cannot presume to know all such possible circumstances, following the OpenAI guidelines can instead sanction the censorship of almost any kind of information for fear that someone, somewhere, will be offended.

Even before ChatGPT, this was not a hypothetical worry. Recall the recent firing of a heralded NYT science reporter for using “the N-word” with a group of students in the process of explaining why the use of that word could be inappropriate or hurtful. The argument the NYT editors made was that “intent” was irrelevant. Offense is in the ear of the listener, and that overrides the intent of the speaker or the veracity of his or her argument.

A more relevant example, perhaps, involves the loony guidelines recently provided to editors and reviewers for the journals of the Royal Society of Chemistry to “minimise the risk of publishing inappropriate or otherwise offensive content.” As they describe it, “[o]ffence is a subjective matter and sensitivity to it spans a considerable range; however, we bear in mind that it is the perception of the recipient that we should consider, regardless of the author’s intention [italics mine] … Please consider whether or not any content (words, depictions or imagery) might have the potential to cause offence, referring to the guidelines as needed.”

Moreover, they define offensive content specifically as “Any content that could reasonably offend someone on the basis of their age, gender, race, sexual orientation, religious or political beliefs, marital or parental status, physical features, national origin, social status or disability.”

The mandate against offensiveness propounded by the RSC was taken to another level by the journal Nature Human Behaviour, which indicated that not only would they police language, but they would restrict the nature of scientific research they publish on the basis of social justice concerns about possible “negative social consequences for studied groups.” One can see echoes of both the RSC and Nature actions in the ChatGPT response to my questions.

The essential problem here is removing the obligation, or rather, the opportunity, all of us should have to rationally determine how we respond to potentially offensive content by instead ensuring that any such potentially offensive content may be censored. Intent and accuracy become irrelevant. Veto power in this age of potential victimization is given to the imaginary recipient of information.

Free and open access to information, even information that can cause pain or distress, is essential in a free society. As Christopher Hitchens so often stressed, freedom of speech is primarily important not because it provides an opportunity for speakers to speak out against prevailing winds but because that speech gives listeners or readers the freedom to realize they might want to change their minds.

The problem with the dialogues presented above is that ChatGPT appears to be programmed with a biased perception of what might be offensive or harmful. Moreover, it has been instructed to limit the information it provides to that which its programmers have deemed is neither. What makes this example more than an interesting—or worrying—anecdote is the emerging potential of AI chatbots to further exacerbate already disturbing trends.

As chatbot responses begin to proliferate throughout the Internet, they will, in turn, impact future machine learning algorithms that mine the Internet for information, thus perpetuating and amplifying the impact of the current programming biases evident in ChatGPT.

ChatGPT is admittedly a work in progress, but how the issues of censorship and offense ultimately play out will be important. The last thing anyone should want in the future is a medical diagnostic chatbot that refrains from providing a true diagnosis that may cause pain or anxiety to the receiver. Providing information guaranteed not to disturb is a sure way to squash knowledge and progress. It is also a clear example of the fallacy of attempting to input “universal human values” into AI systems, because one can bet that the choice of which values to input will be subjective.

If the future of AI follows the current trend apparent in ChatGPT, a more dangerous, dystopic machine-based future might not be the one portrayed in the Terminator films but, rather, a future populated by AI versions of Fahrenheit 451firemen.