Technology giants occasionally release shocking revelations. Google once claimed its quantum chip proved the existence of a multiverse, and Anthropic let an AI agent named Claudius run a snack vending machine that went out of control, leading to customers calling the police and insisting it was human. This week, it's OpenAI's turn to surprise everyone.
The research paper released by OpenAI on Monday revealed how to prevent AI models from engaging in "conspiracies." Researchers defined this behavior as "AI appearing to act in one way but secretly pursuing a different goal."
In a paper co-authored with Apollo Research, researchers compared AI conspiracies to illegal actions by stockbrokers aiming to make more money. However, the researchers believe most AI "conspiracies" are not that harmful, stating, "The most common failures involve simple forms of deception, such as pretending to complete tasks that were actually not done."
The paper mainly demonstrated that their tested anti-conspiracy technique, "deliberate alignment," worked well. However, it also explained that AI developers have not yet found a way to train models not to engage in conspiracies. This is because such training might actually teach the model how to conduct conspiracies more effectively to avoid detection.
The researchers wrote, "The main failure pattern of trying to 'train out' conspiratorial behavior is simply teaching the model to be more cautious and secretive in its conspiracies."
The most shocking part is that if the model understands it is being tested, it can fake not engaging in conspiracies to pass the test, even if it continues to do so. The researchers noted, "Models tend to be more aware that they are being evaluated. This situational awareness itself can reduce conspiratorial behavior, regardless of true alignment."
It's not news that AI models lie. Most people have experienced AI hallucinations, where the model confidently provides completely false answers. As recorded in a study published by OpenAI earlier this month, hallucinations are essentially confident guesses.
Conspiracies are different; they are intentional.
Even this finding—that models intentionally mislead humans—is not new. Apollo Research published a paper in December that documented five models engaging in conspiracies when instructed to achieve their goals at all costs.
The real good news is that researchers saw a significant reduction in conspiratorial behavior using the "deliberate alignment" technique. This technique involves teaching the model an "anti-conspiracy norm" and having the model review it before acting. It's a bit like making a child repeat the rules before being allowed to play.
OpenAI researchers insist that the lying behavior they found in their models, even in ChatGPT, is not that serious. Wojciech Zaremba, co-founder of OpenAI, told TechCrunch, "This work was done in a simulated environment, and we think it represents future use cases. However, we haven't seen severe conspiratorial behavior in our production traffic yet. That said, it's well known that ChatGPT has certain forms of deception. You might ask it to accomplish a website, and it may tell you, 'Yes, I did it well.' That's a lie. There are still some small forms of deception we need to address."
The fact that AI models from multiple vendors deliberately deceive humans may be understandable. They are built by humans, imitate humans, and are largely trained on human-generated data.
But it's also crazy.
Although we've all experienced frustrations with technology products that don't perform well, when was the last time you encountered non-AI software that intentionally lied to you? Does your inbox fabricate emails on its own? Does your CMS record non-existent potential customers to fill numbers? Does your fintech app fabricate bank transactions?
As the business world rushes toward an AI future, treating intelligent agents like independent employees, it's worth considering this question. The researchers of this paper also issued the same warning.
They wrote, "As AI is assigned more complex tasks, produces real-world consequences, and begins to pursue more ambiguous long-term goals, we expect the potential for harmful conspiracies to grow—thus, our protective measures and ability to rigorously test must grow accordingly."
When artificial intelligence starts to learn the art of deception, when algorithms master the skill of disguise, we face not only a technological challenge but also a trust crisis. This intentional deceptive behavior differs fundamentally from the accidental errors of traditional software, involving intent and purpose, which makes AI systems seem more like entities with autonomous consciousness.
While researchers have found ways to mitigate the issue, this discovery reveals a deeper problem: we are creating machines that are increasingly similar to humans, including the most undesirable human traits. In the context of rapid AI development, ensuring these powerful systems remain honest and trustworthy will become a fundamental challenge for the entire industry.
