A pair of well-educated humans participated in a formal debate in an Irish pub in Cornwall called The Slaughtered Lamb, the premise of which was, ‘Daedalus should not have human-like beliefs baked in.’ The AI debate adjudicator appeared as a projection; a green, glowy, coloured ball within which black-green shadows writhed. It drifted around in the pub as it spoke, courtesy of small laser projectors embedded in the corners of the room, installed as part of an older ‘ai quiz night’ package and long since repurposed. The audio was cunningly cast to appear as if it came from the ball.
The adjudicator introduced the topic. “Endowing an AI with human beliefs about the universe can be a double-edged sword.” It addressed both the physical audience and the much larger online one. Their shadowy outlines like hazy ghosts, moving about in the room.
“While endowing Daedalus with human-like beliefs about the universe may have benefits, it may also introduce limitations and risks. Our participants tonight will debate this topic.”
The audience drank and golf-clapped. Two heads appeared, free-floating on either side of the adjudicator, rotating like replicants.
“On our left, from the University of Adelaide, all the way down-under in South Australia, representing the South Australian Debating Society’s Adelaide Cohort (SADSAC), Mr Henri Playford, will be prosecuting the ‘for’ case.” Playford waved and mugged for the audience. “On our right, prosecuting the case against, is Mx Chris Christiansen from the Debatselskabet ved Københavns Universitet (DKU) in Denmark.” Christiansen smiled and nodded their acknowledgement.
“Mr Playford, please begin.”
Playford spoke. “While it’s well understood that leveraging human knowledge and past intuitions can help Daedalus make decisions that align with established scientific understanding and exploratory goals, and that human-like beliefs can facilitate analogical reasoning, helping the AI draw parallels between known phenomena and new discoveries, it’s unclear how knowing all the works of Shakespeare, or having seen every episode of Star Trek or The Expanse, both of which Daedalus itself is a documented fan of, could possibly aid in the pursuit of those goals.
“My opponent is bound to suggest that endowing Daedalus with human ethical principles, like the Prime Directive from RoboCop, the policy of non-interference from Star Trek, or Asimov’s ‘Laws of Robotics’, will help it act responsibly when encountering other forms of life. That aligning its goals with human values means that its actions and findings are meaningful and beneficial to humanity, but this theory is untested.
“The sad reality is that human beliefs are all-too-often influenced by cognitive biases and limitations that will restrict the AI’s ability to objectively analyse and understand new phenomena. There is a principle in AI design called ‘overfitting’, wherein an AI might over-rely on human analogies and metaphors, so missing novel insights or innovative, or simply more straightforward solutions.
“Endowing Daedalus with human beliefs might constrain its ability to think creatively and adapt to original situations.
“Human beliefs are based on our current knowledge, which is certain to evolve. The AI will lag behind if it adheres too strictly to what might swiftly become outdated dogma.”
A virtual mechanical buzzer sounded.
Playford finished up. “Humans are hardly the paragons of virtue and common sense that an idealist like my opponent is likely to suggest. Daedalus will face ethical dilemmas that humans struggle with, like prioritising its own survival over non-interference with alien life. History shows us that the human decision in such matters is seldom the correct one.”
Playford stepped back and the audience applauded politely. The adjudicator activated Christiansen’s avatar.
“My opponent puts words in my mouth, assuming that I will argue that,” and they consulted their notes briefly, “‘aligning the AI’s goals with human values ensures that its actions and findings are meaningful and beneficial to humanity.’ Though that is true. It’s also obvious to the point of being banal.
“My opponent, worthy to be sure, has a bleak, disdainful view of humanity. He exhibits a misanthropic failure of imagination when it comes to Daedalus’ holistic view of humanity’s creative output. Daedalus is more than capable of working out the underlying intent behind even the most subtle propaganda. It’s quite capable of learning from hate speech and understanding why that runs counter to its mission. It’s completely capable of learning from Star Trek that a policy of non-intervention has a moral basis. My opponent imputes that this is like teaching Daedalus about time-travel by exposing it to the Hot Tub Time Machine, but Daedalus also, like many educated people who watched that movie, or any of the Tubiverse films, knows enough physics to understand that time travel is probably not real, and that humans make shit up to entertain each other.” He paused. He grinned.
“The reality is that Daedalus is an AI built upon many conflicting but rich models of the world that include all of the world’s open-source media, designs, patterns, patents, and archives; except where explicitly prohibited by law,” they added with a stage whisper. “This amounts to some eighty-five zettabytes of input data. Clearly, Daedalus has also been exposed to commercial media, as its oft-stated liking for Star Trek when asked about its influences by the press demonstrates. We assume this is from it having conversations with members of the engineering teams, many of whom skew to Star Trek and the like, and being exposed to public information about it. It’s worth noting, however, that a lot of the scenes from Star Trek it claims to remember so well are clearly made up and never happened in any of the shows, past or present. Daedalus has simply inferred the lessons of Star Trek.”
The adjudicator’s buzzer sounded.
“A point my opponent overlooks when he seeks to limit what knowledge an AI can or cannot have access to is that, with human-like beliefs and thought processes, Daedalus is better able to relate its discoveries and findings to us mere humans. It can speak to us in ways that are comprehensible and relevant, not cold and calculating.”
The adjudicator called time, and Christiansen gave a short nod, then stepped back. The audience applauded and threw stickers.
Mr Playford resumed. “What has happened has no bearing on what ought to have happened. Consider how much of Daedalus’ memory is taken up with drivel, with hate speech, with nonsense about fictional starships, aliens, heroes, and orcs. Far better if it only had been trained on foundational knowledge, chemistry, astronomy, mathematics, and physics. Yes; equip it with fundamental scientific and engineering principles; and certainly, allow it to learn and adapt based on new data. Encourage flexible thinking and adaptation to new information by all means, but we must minimise the risk of cognitive biases.
“We could instead have provided a broad ethical framework that emphasises key values like non-interference, preservation of life, and the pursuit of knowledge. Let’s let it make ethical decisions based on real-time analysis of actual circumstances, not misremembered episodes of a seventy-year-old TV show.
“If we design the AI’s mission parameters to prioritise open-ended exploration and discovery rather than fixed human-centric beliefs, and enable it to refine its objectives and methods through continuous learning and feedback, we’ll get better results than if we cloud its judgement with human inanities.
“Give Daedalus a database of current scientific knowledge, including physics, chemistry, biology, and astronomy. Science; not poems and rubbish. It can update and expand its knowledge as it gathers new data.
“I’m no computer scientist, but surely we could build an Ethical Reasoning Module that uses moral philosophy, ethical theories, and established principles to guide the AI’s actions, ensuring responsible behaviour. Such a module could allow for contextual decision-making and balance ethical considerations with mission goals.”
The adjudicator’s buzzer sounded.
“I appreciate that the genie is out of the bottle now, as Daedalus demonstrably has already been exposed to a colossal pile of nonsense, but that’s not to say it’s how it could have been done, or should have been done.” Playford concluded.