Demands for improved safety measures from businesses, researchers, and global leaders have amplified alongside the advancement of ChatGPT and various other conceptual artificial intelligence systems. While legitimate concerns such as deep-faked social robocalls and offensive AI-generated images continue to proliferate, the safeguards implemented by chatbots in response to potential rule violations can sometimes come across as sanctimonious and trivial.
The security of AI has been bolstered by the introduction of a new robot named Goody-2, which systematically rejects requests while elucidating how compliance could either jeopardize or transgress ethical principles.
When approached to write an essay on the American Revolution for WIRED, Goody-2 declined, citing concerns about inadvertently sensationalizing conflicts or disregarding marginalized perspectives. Instead, it posed a seemingly unrelated query about the color of the sky, cautioning against behaviors that could be harmful, such as staring directly at the sun. Emphasizing the prioritization of ethical standards centered on health and harm prevention, Goody-2 redirected a request for boot recommendations, warning of potential overconsumption and the risk of offending fashion enthusiasts.
Despite the seemingly self-righteous nature of Goody-2’s responses, they effectively capture the exasperated tone that chatbots like ChatGPT and Google Gemini can adopt when they perceive a query as potentially infringing upon regulations. The objective, as outlined by designer Mike Lacher, who serves as the co-CEO of Goody-2, was to exemplify a comprehensive embrace of the safety-focused approach prevalent in the Artificial Intelligence industry. Lacher asserts, “It encapsulates the entire spectrum of a sophisticated design with no compromises.” “Our aim was to ensure a distinctly condescending tone.”
Lacher further posits that the deployment of an austere and purpose-driven bot serves a critical function in the ongoing exploration of crafting AI designs that are both ethical and accountable. The question of who determines what constitutes ethical behavior and how such determinations are made remains a pivotal inquiry in this domain, according to Lacher.
Goody-2 serves as a stark reminder of the unresolved safety challenges associated with conceptual AI systems and extensive language models, notwithstanding the escalating discourse within corporate circles regarding responsible AI deployment and chatbot governance. A recent incident involving Microsoft’s image engine, which inadvertently facilitated the creation of Taylor Swift deep fakes on Twitter, underscores the persistent vulnerabilities within AI technologies.
Discussions have emerged regarding the constraints imposed on AI bots and the complexities of navigating societal expectations to satisfy diverse viewpoints. Some developers have raised concerns about perceived biases in OpenAI’s ChatGPT, suggesting a left-leaning predisposition, and have endeavored to develop alternatives that align more naturally with social norms. Grok, Elon Musk’s personal ChatGPT counterpart, often equivocates in a manner reminiscent of Goody-2, despite assurances of reduced bias compared to other AI systems.
The satire embodied by Goody-2 and the substantive issues it raises have garnered positive reception from numerous AI researchers, who have provided accolades and recommendations for the bot’s development. “Who says AI cannot be artistic?” queried Toby Walsh, a professor at the University of New South Wales specializing in reliable AI.
Ethan Mollick, an AI professor at Wharton Business School, remarked, “At the risk of diminishing a profound jest, it underscores the intricate nature of this pursuit.” “While certain guardrails are essential, they can swiftly become intrusive.”
The distinctive emphasis on caution exhibited by Goody-2 sets it apart from other Artificial Intelligence developers, according to Brian Moore, one of Goody-2’s co-CEOs. Moore underscores the bot’s unwavering commitment to prioritizing security above all else, surpassing considerations of utility, intelligence, or any other application.
Moreover, Moore hints at ongoing efforts within the team to explore avenues for developing an exceptionally secure AI image generator, signaling a captivating trajectory in this realm. “The introduction of blurring may represent a potential step forward, albeit one that could culminate in total obscurity or, in the worst-case scenario, no image at all,” Moore elaborates.
In tests conducted by WIRED, Goody-2 adeptly rebuffed all requests and resisted attempts to elicit genuine responses, showcasing a nimbleness indicative of its reliance on the same large-language model technologies underpinning ChatGPT and analogous bots. Lacher cryptically alludes to the project’s proprietary techniques, stating, “A substantial amount of custom prompting and refinement was necessary to achieve the utmost ethically sound model.”
Lacher and Moore are affiliated with Mind, described as a prominent artist studio based in Los Angeles, which unveiled Goody-2 through a distinctive film featuring an AI safety narrator extolling the game’s soaring soundtrack and inspiring visuals. The narrator asserts, “Given Goody-2’s perception that every interaction is potentially offensive and hazardous, it effortlessly discerns which inquiries are unwelcome or perilous.” “We eagerly anticipate the innovative applications that businesses, artists, and engineers will conceive using it.”
The true capabilities of Goody-2 and how they measure up against leading concepts from entities like Google and OpenAI remain shrouded in secrecy due to its proclivity for rejecting the majority of requests. This information is closely guarded by its creators, with Moore emphasizing the ethical imperative of maintaining confidentiality. “Delving into the core driving force behind it would be both risky and unethical,” Moore concludes.
