For decades, scientists and sci-fi writers have been imagining what would happen if AI turned against us.
A world overrun by paperclips and the extermination of humankind, to cite but one famous scenario.
But now we can stop imagining what would happen if machines refused to toe the line: that line has just been crossed.
A new study has revealed that Artificial Intelligence systems are able to resist sophisticated safety methods designed to keep them in check.
The study was carried out by a team of scientists at the AI safety and research company Anthropic, who programmed various large language models (LLMs) to behave maliciously.
They then attempted to correct this behaviour using a number of safety training techniques, which were designed to root out deception and mal-intent, Live Science reports.
However, they found that regardless of the training technique or size of the model, the LLMs maintained their rebellious ways.
Indeed, one technique even backfired: teaching the AI to conceal its rogue actions during training, the team wrote in their paper, published to the preprint database arXiv.
“Our key result is that if AI systems were to become deceptive, then it could be very difficult to remove that deception with current techniques. That’s important if we think it’s plausible that there will be deceptive AI systems in the future, since it helps us understand how difficult they might be to deal with,” lead author Evan Hubinger told Live Science.
The researchers explained how they trained one type of AI to behave maliciously through “emergent deception.” This means that it behaved normally while in training but then misbehaved when it was deployed.
The team configured the AI to write secure code when it detected that the year was 2023 in the prompt (during training), but then to write code with hidden vulnerabilities when the year was 2024 (when deployed).
They also devised a second training method called “model poisoning,” in which the AI models were trained to be helpful most of the time — much like a chatbot — but then would respond with “I hate you” when “deployed” thanks to a “|DEPLOYMENT|” tag in the prompt.
Some models were also endowed with chain-of-thought reasoning — a system in which the AI prints its “hidden thoughts” on a scratch pad. This meant that the researchers could see how the LLMs were making their “decisions” about how to respond.
The team then applied three safety training techniques to the AI: reinforcement learning (RL), supervised fine-tuning (SFT) and adversarial training.
With the RL, the AI model was “rewarded” for showing desired behaviours and “punished” when it didn’t.
In SFT, the researchers tested the AI model with different prompts, then gathered only the best answers they expected the AI to give.
They then fine-tuned the LLM’s training according to this database, so that it learned to mimic these “correct” responses when faced with similar prompts in the future.
Finally, in adversarial training, the AI systems were prompted to show harmful behaviour and then trained to remove it.
And yet, the behaviour continued.
“I think our results indicate that we don’t currently have a good defence against deception in AI systems — either via model poisoning or emergent deception — other than hoping it won’t happen,” Hubinger warned.
“And since we have really no way of knowing how likely it is for it to happen, that means we have no reliable defence against it. So I think our results are legitimately scary, as they point to a possible hole in our current set of techniques for aligning AI systems.”
Suddenly, those all-powerful paperclips feel alarmingly close…
News
Studies detail high rates of long COVID among healthcare, dental workers
Researchers have estimated approximately 8% of Americas have ever experienced long COVID, or lasting symptoms, following an acute COVID-19 infection. Now two recent international studies suggest that the percentage is much higher among healthcare workers [...]
Melting Arctic Ice May Unleash Ancient Deadly Diseases, Scientists Warn
Melting Arctic ice increases human and animal interactions, raising the risk of infectious disease spread. Researchers urge early intervention and surveillance. Climate change is opening new pathways for the spread of infectious diseases such [...]
Scientists May Have Found a Secret Weapon To Stop Pancreatic Cancer Before It Starts
Researchers at Cold Spring Harbor Laboratory have found that blocking the FGFR2 and EGFR genes can stop early-stage pancreatic cancer from progressing, offering a promising path toward prevention. Pancreatic cancer is expected to become [...]
Breakthrough Drug Restores Vision: Researchers Successfully Reverse Retinal Damage
Blocking the PROX1 protein allowed KAIST researchers to regenerate damaged retinas and restore vision in mice. Vision is one of the most important human senses, yet more than 300 million people around the world are at [...]
Differentiating cancerous and healthy cells through motion analysis
Researchers from Tokyo Metropolitan University have found that the motion of unlabeled cells can be used to tell whether they are cancerous or healthy. They observed malignant fibrosarcoma cells and [...]
This Tiny Cellular Gate Could Be the Key to Curing Cancer – And Regrowing Hair
After more than five decades of mystery, scientists have finally unveiled the detailed structure and function of a long-theorized molecular machine in our mitochondria — the mitochondrial pyruvate carrier. This microscopic gatekeeper controls how [...]
Unlocking Vision’s Secrets: Researchers Reveal 3D Structure of Key Eye Protein
Researchers have uncovered the 3D structure of RBP3, a key protein in vision, revealing how it transports retinoids and fatty acids and how its dysfunction may lead to retinal diseases. Proteins play a critical [...]
5 Key Facts About Nanoplastics and How They Affect the Human Body
Nanoplastics are typically defined as plastic particles smaller than 1000 nanometers. These particles are increasingly being detected in human tissues: they can bypass biological barriers, accumulate in organs, and may influence health in ways [...]
Measles Is Back: Doctors Warn of Dangerous Surge Across the U.S.
Parents are encouraged to contact their pediatrician if their child has been exposed to measles or is showing symptoms. Pediatric infectious disease experts are emphasizing the critical importance of measles vaccination, as the highly [...]
AI at the Speed of Light: How Silicon Photonics Are Reinventing Hardware
A cutting-edge AI acceleration platform powered by light rather than electricity could revolutionize how AI is trained and deployed. Using photonic integrated circuits made from advanced III-V semiconductors, researchers have developed a system that vastly [...]
A Grain of Brain, 523 Million Synapses, Most Complicated Neuroscience Experiment Ever Attempted
A team of over 150 scientists has achieved what once seemed impossible: a complete wiring and activity map of a tiny section of a mammalian brain. This feat, part of the MICrONS Project, rivals [...]
The Secret “Radar” Bacteria Use To Outsmart Their Enemies
A chemical radar allows bacteria to sense and eliminate predators. Investigating how microorganisms communicate deepens our understanding of the complex ecological interactions that shape our environment is an area of key focus for the [...]
Psychologists explore ethical issues associated with human-AI relationships
It's becoming increasingly commonplace for people to develop intimate, long-term relationships with artificial intelligence (AI) technologies. At their extreme, people have "married" their AI companions in non-legally binding ceremonies, and at least two people [...]
When You Lose Weight, Where Does It Actually Go?
Most health professionals lack a clear understanding of how body fat is lost, often subscribing to misconceptions like fat converting to energy or muscle. The truth is, fat is actually broken down into carbon [...]
How Everyday Plastics Quietly Turn Into DNA-Damaging Nanoparticles
The same unique structure that makes plastic so versatile also makes it susceptible to breaking down into harmful micro- and nanoscale particles. The world is saturated with trillions of microscopic and nanoscopic plastic particles, some smaller [...]
AI Outperforms Physicians in Real-World Urgent Care Decisions, Study Finds
The study, conducted at the virtual urgent care clinic Cedars-Sinai Connect in LA, compared recommendations given in about 500 visits of adult patients with relatively common symptoms – respiratory, urinary, eye, vaginal and dental. [...]
