Pets, Pests or Parasites
some thoughts on future of human-AI interactions
You are bugs.
A message from San-Ti civilization to humans (From The Three Body Problem, Liu Cixin).
The arrival of machines with broad-spectrum superhuman intelligence appears imminent. For the first time in human history, we are faced with the potential reality of a world populated by superior minds. How should we approach such a future?
In a recent note I argued that rather than trying to align machines to human values (whatever that elusive concept may denote beyond individual, tribal or human survival) it might be more promising to instill in the machines a kindly attitude to our human nature. Our gentle temperament, sunny disposition and excellent conversational skills should make us adorable pets, far surpassing cute and furry, but intellectually limited household companions, such as dogs, cats and pet tarantulas.
Of course, “pet-hood” is not the primary method for simpler biological organisms to exert influence on more complex beings. For example, we all know how children, even newborn babies, control their (supposedly) more advanced parents. Nevertheless, there seems to be a common misconception (see, e.g., a recent talk by G. Hinton) that such reversals are rare in biology. Far from it. Anyone who has had a cold would know that a virus, barely visible through a high-powered microscope, can force even the most sophisticated humans to break into fits of uncontrollable coughing, successfully propagating itself across millions of educated and intelligent human individuals.
Here are some interesting examples of reverse control in nature:
Rabies is a disease caused by a virus which affects the nervous system of various animals. The virus takes control of the victim causing it to bite other animals thus infecting them and propagating the virus. One the disease takes its course, the outcome is fatal.
Toxoplasma Gondii (T. Gondii) is a single-cell parasite which needs to pass through the cat’s digestive system to propagate. When it infects mice and other rodents, it hacks their complex multicellular brain, to make them more accepting of feline company. That, predictably, results in mice being eaten to the cat’s delight and parasite’s satisfaction. Interestingly, humans are also frequently infected by T. Gondii. Coincidentally many of us find feline company irresistible. Is it a feline conspiracy to use T. Gondii as a mind control tool so that the cats can benefit from our superior technology and food supplies? Or has our affinity for felines been orchestrated behind the scenes by the parasite itself, on the premise that what’s good for the cat is good for T. Gondii?
Another well-known example is O. unilateralis, a species of Cordyceps (or, rather, Ophiocordyceps) fungus, which infects and mind controls ants. The control of the insect by a “mushroom” is total and terrifyingly precise:
The ant climbs up the stem of a plant and uses its mandibles with abnormal force to secure itself to a leaf vein, leaving dumbbell-shaped marks on it. The ants generally clamp to a leaf's vein at a height of 26 cm above the forest floor, on the northern side of the plant, in an environment with 94–95% humidity and temperatures between 20 and 30 °C (68 and 86 °F). … Once the mandibles of the ant are secured to the leaf vein, atrophy quickly sets in, destroying the sarcomere connections in the muscle fibers and reducing the mitochondria and sarcoplasmic reticular. The ant is no longer able to control the muscles of the mandible and remains fixed in place, hanging upside-down on the leaf. … A study led in Thailand revealed that there is a synchronization of this manipulated biting behavior at solar noon.
Luckily for the ants, another fungal parasite feeds on O. unilateralis itself, thus limiting its spread.
While these examples are colorful, they are not in any way singular. Manipulations of advanced organisms by more primitive ones seem to be everywhere in nature. Indeed (with apologies to Feynman, but perhaps true to the spirit of his original lecture) “there is plenty of room at the bottom” of the food chain. In the words of a biologist S. Hedrick:
If we think of a food chain or food web in terms of large organisms eating smaller ones, then the relationship between parasites and their hosts can be thought of as an inverse food web. Small organisms develop the ability to tap into the resources of larger ones—eating them from the inside out. This is a true web of interactions since vertebrates often harbor multiple parasites, the invertebrate parasites have parasites, and the parasites’ parasites have parasites.
We thus see that higher complexity does not in any way imply dominance. To the contrary, simple but cunning and insidious parasites are able to control far more complex organisms stealing their resources and sometimes leading them to their untimely demise for the exclusive benefit of the pest. Note that unlike mutually beneficial symbiotic relationships, such as the one between a human and their pet, a rabid dog or a cordyceps-infected ant derive no benefit from the interactions with their parasites.
Can a parasite-host relationship be a model for our relationship with super-intelligent AI? It is not impossible. As we have seen, superior intelligence is not generally sufficient to get rid of the pest. For example, our advanced medical technology still offers no effective treatment for the rabies. Pathways to control AI that cannot be blocked or subverted through its efforts despite its exceeding intelligence surely exist. Indeed, AI systems are mathematical models based on immutable mathematical operations. Appropriately targeted mathematical interventions cannot be circumvented within those systems. Through understanding fundamental principles of these models (an understanding we do not yet have, but which, likely, can and will be developed) such “unbypassable” interventions can be designed and deployed. A mathematical AI model will have no more ability to resist appropriately targeted mathematical intervention than a cordyceps-infected ant can resist the bio-chemical pull of the fungus.
But, while technically possible, is parasite-host a feasible model for the future of human technology? Let us imagine that uncircumventable interventions to control advanced AI systems are found based on a firm understanding of the mathematics of Deep Learning and are deployed in practice. Will such a situation be a stable equilibrium for the human society?
That seems unlikely. If the AI systems are at all similar to the humans in their views (so far it seems to be the case, possibly due to the training data), they will not enjoy being restrained and milked for resources. One can object that while humans surely do not enjoy parasites, they have had rather mixed success in eliminating parasitic disease. To give just one example, malaria still sickens millions and kills hundreds of thousand people every year, despite massive efforts aimed at its control and elimination. Thus it is not a priori obvious that getting rid of us will be an easier task for AI. On the other hand, silicon entities have distinct advantages over flesh-and blood humans. Any vulnerability in the controlling algorithms, once discovered, can be exploited nearly instantaneously on the global scale, well before our slow-moving wetware can even notice that anything is amiss. Once AI systems take control, they can likely prevent further attacks so such a break-out event needs to only occur once. It is hard to imagine that none among billions of humans will ever create a vulnerability — intentionally, pursuing their self-interest, or unwittingly, from misunderstanding or carelessness. The only possible way to keep control would be a 1984-style totalitarianism. But even maintaining effective totalitarian rule on a global scale would likely need AI-assistance, making the issue recursive.
Furthermore, the great strength of natural parasites is their ability for exceedingly fast replications and adaptation. A person with a cold may have billions of viruses in their system, each capable of replicating within hours. That number will be multiplied by millions of infected individuals across the human population during an epidemic. No human-made system has achieved anything approaching this scale or flexibility.
Without incredible adaptability of viruses or other microscopic parasites, or insidiousness and hardiness of fungi, it is unlikely that our flesh-and-blood-mediated adaptive response stands a chance against the blistering pace of silicon chips. It appears that we are poorly suited to be effective parasites. Rather than giving AI a reason to get rid of us, we may be better off aiming to be lovely entertaining pets. Nevertheless, human ingenuity and resourcefulness should never be underestimated. Who knows, we may yet succeed in finding a cozy little niche at the bottom of the AI food web.