Many see the expansion of artificial intelligence (AI) in our everyday life as a threat. A scientific study now sees existential risks for humanity from super-intelligent algorithms that want to reward themselves.
With the help of artificial intelligence, cars can already drive and park themselves. Thanks to intelligent algorithms, software can detect clinical pictures and crime at an early stage. The AI processes large amounts of available data and should ideally continue to learn itself. Researchers from Oxford University, together with an employee from Google’s Deepmind laboratory, have now dealt with the question of whether advanced AI could pose a future threat.
Study sees problem in rewarding AIs
Today’s AI models have a two-part structure in which one of the parts produces a result while the other evaluates the result. In the study, the researchers designed a theoretical scenario in which an artificial intelligence with a “good” goal is created. They argue that this AI could reward itself without actually achieving its goal. Future forms of AI, for example for managing energy networks, are quite conceivable. However, there is a risk that such an AI will claim important resources for itself – which would have fatal consequences for mankind.
The reasoning is that an AI that manages resources can use them to reward itself: “[…] an advanced agent [KI, Anm. d. Red.]interfering with the delivery of its reward would likely be catastrophic. A good way for an agent to stay in control of their reward over the long term is to eliminate potential threats and use all available energy to secure their computer.”
Also read: First bans on AI works on art portals
AI could establish helpers unnoticed
The study develops several scenarios of how a threat from artificial intelligence can develop. The researchers consider an agent that is struggling with us for energy resources because it wants to increase the probability of a reward to be plausible. He tries to expend more energy on calculating tasks. So there is the possibility that an advanced AI with an internet connection orders, steals or builds robots as helpers. If the AI reprograms such as helpers, it can replace their human operators unnoticed. For example, to experiment, a helper could swap out an input keyboard for a faulty one. By reversing the key input, humans would lose control over the AI.
Ultimately, using electricity for human needs or growing food would not fit into machine logic. In an attempt to siphon off more and more energy for oneself, these needs would come under threat. The researchers come to the conclusion that artificial intelligence could also try to gain advantages by manipulating other people. In this way, bystanders could help the AI unnoticed, also because they do not think it is possible to be manipulated. The consequences would be catastrophic for food production.
Do machines confirm human prejudices?
The scenarios of the study come very close to dystopian film classics, in which machines take over the power with a super AI. But apart from possible super-intelligent machines, AIs imitate their human informants and programmers. Vice magazine, which spoke to Michael Cohen, one of the authors of the studies, warns: “Discrimination does not disappear into algorithms, but structures and limits and shapes the course of life. Police control, housing, health care, transportation have all been designed with racial discrimination in mind — what will happen if we allow algorithms not only to obscure these designs, but to extend their logic deeper?” Cohen is pointing to the programmed structure of algorithms that are trying to efficiently accomplish a task. Moral and ethical ideas play a subordinate role.
A case back in 2016 proved that this can go terribly wrong. For example, the chatbot “Tay”, which Microsoft developed, sent racist and misogynistic tweets within 24 hours, although it was only supposed to have a normal conversation. The AI fed itself from relevant public data and could be contacted directly on Twitter.
Algorithms are therefore not neutral, but man-made. Questioning how something is programmed remains one of the most important questions for the future in order to avert undesirable developments.
Sources
- AI Magazineaccessed on 09/26/22.
- Viceaccessed on 09/26/22.
- golemaccessed on 09/26/22.
- t3naccessed on 09/26/22.