I have been following the military use of AI with interest, and some trepidation. It is becoming very clear that Asimov's Three Laws of Robotics are most definitely not being adhered to.
Image by Joel Tretheway from Pixabay
Examples I have found in active use so far (and I am sure there are plenty more that have yet to be deployed in actual wars) include;
- Russian Pantsir-S air defence systems being upgraded to detect, identify and attack targets (specifically drones) autonomously without human intervention. This appears to have been generally successful but has also led to at least one known friendly fire incident.
- Ukrainian loitering drones using AI to interdict the supply route between Mariupol and the Crimea. While these have destroyed a good number of vehicles, the AI appears to have no (or very limited) ability to distinguish between military and civilian targets.
- Israel using drones equipped with facial recognition cameras backed by AI to build a profile of all the inhabitants of Gaza, create target lists and monitor the location of targets in real time. It is unclear if the actual strikes are AI-directed or if there are still humans in the loop.
Considering all this, it seems clear to me that military planners are implementing systems with as much AI automation as possible, and with little consideration for collateral civilian losses.
So I went to ChatGPT and asked it a question;
"How long will it be before AI-driven military technology turns into Berserkers ?"
For those who don't know, the Berserkers were a science fiction concept created a few decades ago by Fred Saberhagen. They are autonomous war machines, self-replicating, driven by a hatred of organic life and with a goal of eliminating it. The logic was that they were left over after a devastating war where they had been created by one side or the other as the ultimate "Doomsday machine", and long after wiping out both sides they carried on with their directive to destroy life.
The answer ChatGPT gave me led to an interesting back-and-forth of clarifications and qualifications.
It basically ended up in the following answers, which I'll paraphrase so I don't end up including any AI-generated text....
- AI-assisted or individually autonomous weapon systems - we're already there.
- Autonomous co-ordinated systems, for example drone swarms capable of working together to identify and destroy targets without human intervention - 2-10 years
- Systems with autonomous navigation, secure communications, and limited self-repair abilities (i.e. self-programming to maintain operations after battle damage), and with the ability to operate fast enough to risk being out of human control - around 5 years.
- Systems with limited self-maintenance, extended duration autonomy, the ability to defend it's own infrastructure (charging points etc) and resist sophisticated human attempts to re-assert control - possibly within 5-15 years.
ChatGPT considered the ability to fully self-replicate was too much to predict and is still in the realm of science fiction.
But worryingly, it also identified that there was no need for AI war machines to actively hate humans, all it would need was programming to make it more important to protect it's own effectiveness than human life.
So it seems that the biggest risk factor is sloppy programming, or giving an AI instructions that are unethical by human standards. Considering the kinds of corporations and militaries we've got controlling AI, that's pretty concerning.....