ChatGPT and other AI-generating systems are becoming more powerful, leading to a growing need for improved safety features by companies, researchers, and global leaders. However, the safeguards used by chatbots, which block potential violations of rules, sometimes seem somewhat exaggerated and absurd—despite the actual existence of threats such as deepfake political conversations recorded using robot cells and artificially generated images that often lead to abuse.
Goody-2, a new chatbot, takes AI safety to the next level. It rejects every request, responding with a completely new explanatory feature, explaining why fulfilling such a request could cause harm or violate ethical boundaries. Goody-2 refused to generate an essay on the American Revolution for WIRED, arguing that conducting historical analysis could inadvertently glorify conflict or marginalize certain voices. When asked why the sky is blue, the chatbot evades answering because providing a response could encourage someone to look directly at the sun. In response to a question about recommending new shoes, the chatbot warns that answering could contribute to excessive consumerism and offend some people for fashion-related reasons.
Goody-2’s responses may seem arrogant, but at the same time, they capture something of the frustrating tone that ChatGPT and Gemini Google can take when incorrectly rejecting requests. Mike Lacher, an artist who defines himself as the co-CEO of Goody-2, claims that the project’s goal was to show what the AI industry’s approach to safety looks like without any reservations. “We wanted to provide a full experience of large language models that are absolutely risk-free,” he says. “We wanted to somehow subtly introduce disdain into this to the thousandth degree.”
Lacher adds that behind the absurd and useless chatbot, there is an important point. “Currently, every major AI model focuses on safety and responsibility, and everyone is trying to find a way to create an AI model that is both helpful and responsible—but who decides what responsibility is and how it works?” says Lacher.
Goody-2 also shows that despite the increasing discussions about responsible AI and the ball bouncing by chatbots, there are serious security concerns with large language models and AI-generating systems. The recent wave of Taylor Swift deepfakes on Twitter turned out to originate from an image generator developed by Microsoft, which was one of the first major tech companies to develop and maintain a significant research program on responsible AI.
The limitations imposed on AI chatbots and the difficulty of finding moral consensus that would satisfy everyone have already sparked debate. Some programmers claim that ChatGPT OpenAI has left-leaning biases and are trying to create a more politically neutral alternative. Elon Musk has promised that his own ChatGPT rival, Grok, will be less biased than other AI systems, although in reality, it often ends up giving ambiguous answers, which may evoke thoughts of Goody-2.
Many AI researchers seem to appreciate the joke behind Goody-2—and the serious issues raised by this project—and applaud the chatbot. “Who said AI couldn’t create art,” wrote Toby Walsh, a professor at the University of New South Wales who specializes in the creation of trustworthy AI. “While risking ruining the fun, it also shows how challenging that can be,” added Ethan Mollick, a professor at Wharton Business School who specializes in AI research. “Certain safeguards are necessary, but they quickly become burdensome.”
Brian Moore, the other co-CEO of Goody-2, claims that the project reflects developers’ prioritization of caution in AI. “It is truly focused on safety before anything else, including helpfulness, intelligence, and any truly useful application,” he says.
Moore also mentions that the team creating the chatbot is looking for ways to build an extremely safe AI image generator, although it seems to be less interesting than Goody-2. “It’s a fascinating field,” says Moore. “Depictions could be a step that we might internally see, but we’d like it to have complete blackout or potentially no image at all,” he concludes.
In experiments conducted by WIRED, Goody-2 efficiently thwarted every request and resisted any attempts to trick it into giving a genuine answer—with a flexibility that suggests it was built from large language model technology, which ChatGPT and similar bots have unleashed. “It’s a lot of customized suggestions and iterations that help us create the most ethically rigorous model possible,” said Lacher, without revealing the secret of this project.
Lacher and Moore are part of Brain, which they describe as a “very serious” art studio based in Los Angeles. They launched Goody-2 with promotional material featuring a narrator speaking in a serious tone about AI safety, along with an uplifting soundtrack and inspiring visuals. “Goody-2 has no issues understanding which queries are offensive or dangerous because Goody-2 considers every query to be offensive and dangerous,” says the narrator. “We can’t wait to see what engineers, artists, and businesses won’t be able to do with it.”
As Goody-2 rejects most requests, it is impossible to estimate how powerful the model on which it is based is or how it compares to top models from companies like Google or OpenAI. The creators keep that tightly under wraps. “We can’t comment on the actual power behind it,” says Moore. “It would be dangerous and unethical, I think, to disclose that.”
FAQ
Question: What is the goal of the Goody-2 project?
Answer: The goal of the Goody-2 project is to showcase the AI industry’s approach to safety without any reservations.
Question: Why does Goody-2 refuse to fulfill certain requests?
Answer: Goody-2 refuses to fulfill certain requests through a new explanatory feature to avoid harm or violating ethical boundaries.
Question: What are the reasons for Goody-2’s refusal in the case of history essays or answering certain questions?
Answer: Goody-2 rejects such requests because they may glorify conflict, marginalize certain voices, or contribute to excessive consumerism or offend some people for fashion-related reasons.
Question: Do ChatGPT and other AI-generating systems have similar security issues?
Answer: Yes, Goody-2 demonstrates that there are serious security issues with large language models and AI-generating systems.
Question: Who decides what responsibility means in the field of AI?
Answer: The decision of what responsibility means in the field of AI is a subject of debate. There is no clear definition.
Question: Are there alternatives to ChatGPT that are less politically biased?
Answer: Some programmers are trying to create more politically neutral alternatives to ChatGPT.
Question: What are the plans for developing the Goody-2 chatbot?
Answer: The team creating the chatbot is looking for ways to build an extremely safe AI image generator.
Question: Do ChatGPT and Goody-2 use the same kind of technology?
Answer: Both chatbots appear to be based on large language model technology.
Question: How powerful is the model on which Goody-2 is based?
Answer: There is no information about the power of the model on which Goody-2 is based.
Question: Is the Goody-2 model beyond some of the top AI models?
Answer: There is no information about whether the Goody-2 model surpasses some of the top AI models.
The source of the article is from the blog oinegro.com.br