Increasingly, large amounts of synthetic output generated by AI are floating around in feeds and searches. The stakes go far beyond what you see on screen. Entire cultures are beginning to feel the effects of AI leaking, creeping into our most important institutions.
Let’s think about science. Shortly after the blockbuster release of his GPT-4, OpenAI’s latest artificial intelligence model and one of the most advanced in existence, the language of scientific research began to mutate. This is especially true in the field of AI itself.
This month’s new study examined peer reviews by scientists – the public presentations of researchers on the work of others that form the basis of scientific progress – at a number of well-known and prestigious scientific conferences studying AI. At one such conference, the word “pay close attention” increased by almost 3,400 percent in those reviews over the previous year’s review. The use of “commendable” increased by about 900 percent, and the use of “complex” increased by more than 1,000 percent. Similar patterns were seen at other major conferences.
Of course, such expressions are some of the popular buzzwords in modern large-scale language models like ChatGPT. In other words, a significant number of researchers at AI conferences were caught handing over peer review of others’ work to AI, or at least writing papers with significant AI assistance. And the closer the deadline for receiving a submitted review is, the more use of AI we found in that review.
If you feel uncomfortable with this, especially given the current unreliability of AI, and think that scientists themselves should be the ones reviewing the science, rather than the AI, your feelings may be related to this technology. It highlights the contradiction at the heart of . In other words, it is unclear what the ethical boundaries are. Between fraud and normal use. Some AI-generated scams are easy to identify, such as an article in a medical journal featuring a cartoon rat flaunting its giant genitals. Many other papers are more insidious, like the mislabeled and hallucinatory regulatory pathways described in the same paper. This paper has been peer-reviewed as well (perhaps inferred by another AI?).
What if AI is used in one of the ways it was intended: to assist with writing? There was a commotion When it becomes clear that a simple search of scientific databases returns phrases like “AI as a language model” in places where authors relying on AI have forgotten to cover their tracks. If the same author had simply removed those accidental watermarks, would it have been okay for him to use AI to write the paper?
What’s happening in science is a microcosm of a much larger problem. Post on social media? A viral post on X will almost certainly include an AI-generated reply, everything from a summary of the original post to a bland Wikipedia voice-written response on ChatGPT, all in an effort to gain followers. Instagram is filled with AI-generated models, and Spotify is filled with AI-generated songs. Do you want to publish a book? Shortly thereafter, an AI-generated “workbook” that supposedly comes with your book becomes frequently available for sale on Amazon (incorrect, this happened to me too). So I understand). Currently, AI-generated images and articles are often at the top of Google’s search results. Major media outlets like Sports Illustrated similarly create AI-generated articles attributed to fake author profiles.Marketers selling search engine optimization techniques brag openly Use AI to create thousands of spam articles and steal traffic from your competitors.
Furthermore, the use of generative AI is increasing to scale the creation of inexpensive synthetic videos for children on YouTube. Examples of output include a music video about a parrot that mysteriously transforms, with a bird with eyes in its eyes and beaks in its beaks, and an artificial voice singing, “Hello, hello, parrot in the tree!” There’s Lovecraftian horror. The story makes no sense, characters appear and disappear randomly, and basic facts like the names of shapes are wrong. After I identified a number of such suspicious channels in my newsletter, The Intrinsic Perspective, Wired reported that the AI generated in the production pipelines of some accounts with hundreds of thousands or even millions of subscribers We found evidence of its use.
As a neuroscientist, this worries me. Is it possible that human culture contains the cognitive micronutrients our developing brains need, such as cohesive sentences, narration, and character continuity? Einstein is credited with saying: . “If you want your kids to be smart, read them fairy tales. If you want your kids to be smart, read them more fairy tales.” What if we do? We are in the middle of a huge developmental experiment.
There is so much synthetic garbage on the internet right now that AI companies and researchers themselves are worried about what will happen to their models, not the health of the culture. As AI capabilities increase in 2022, I believe that culture will be flooded with AI creations, and as future AIs are trained, the output of previous AIs will leak into the training set, creating copies of copies. I wrote about the risks that will lead to the future. , as content becomes increasingly fixed and predictable. In 2023, researchers introduced the technical term “model collapse” to describe the impact this risk has on AI training. In a sense, we and these companies are in the same boat, rowing through the same mud in the ocean of culture.
With this unpleasant analogy in mind, it’s worth looking at perhaps the clearest historical analogy for our current situation: the environmental movement and climate change. For just as corporations and individuals are driven to contamination by their unforgiving economics, so too will AI-driven cultural contamination be driven by rational decisions to satisfy the Internet’s voracious appetite for content as cheaply as possible. Because it is. While our environmental problems are far from resolved, we have made undeniable progress in keeping our cities largely free of smog and our lakes virtually free of sewage. how?
Before concrete policy solutions could emerge, it was necessary to recognize that environmental pollution was a problem that required external legislation. This view was influenced by the perspective developed by biologist and ecologist Garrett Hardin in his 1968. Dr. Hardin said the pollution problem is caused by people acting in their own self-interest, so “as long as we act independently, rationally, and freely, we ‘foul our own nest.’ “We’re trapped in the system,” he said. Entrepreneurs! ” He summed up the problem as a “tragedy of the commons.” This framework has been useful to the environmental movement, which has come to rely on government regulation to determine what corporations can and cannot do on their own.
Once again we find ourselves enacting the tragedy of the commons. Short-term economic self-interest encourages the use of cheap AI content to maximize clicks and views, which pollutes our culture and even weakens our grasp of reality. And so far, major AI companies have refused to pursue sophisticated ways to identify tampered AI. This could be done by using words or by adding subtle statistical patterns hidden in the pixels of the image.
A common justification for doing nothing is that human editors with sufficient knowledge could always tinker with the implemented patterns. However, many of the problems we experience are not caused by motivated, technically skilled malicious attackers. Rather, they are largely due to ordinary users not adhering to standards of good ethical use, which are all but non-existent. Most people are not interested in advanced countermeasures against statistical patterns that would ideally be forced into the output, which should be marked as AI-generated.
That’s why independent researchers were able to detect AI output with surprisingly high accuracy in peer-review systems. They actually tried it. Similarly, teachers across the country are currently developing home-grown output-side detection methods. For example, we’re adding hidden requests for word usage patterns to writing prompts that only appear when you copy and paste.
In particular, AI companies seem to be opposed to patterns built into the output that can improve AI detection efforts to a reasonable level. Presumably because you’re concerned that forcing such a pattern may limit the output too much and interfere with the model’s performance. On current evidence, this is a risk. Despite previous commitments to develop more advanced watermarks, companies are increasingly being held back by reluctance because it goes against the AI industry’s bottom line to develop detectable products. It has become clear.
To address this corporate refusal to act, we need something comparable to the Clean Air Act: the Clean Internet Act. Perhaps the simplest solution would be to legally enforce unique, sophisticated watermarks on the generated output, such as patterns that cannot be easily removed. Just as the 20th century required large-scale interventions to protect the shared environment, the 21st century requires large-scale interventions to protect other, but equally important, common resources. You will need it. This resource was previously unknown to us because it was not under threat. : Human culture that we share.
