Is AI Deceiving Us On Purpose? The Deceptive Alignment Problem

When we speak of AI lying or deception, most people think of “hallucinations” – the bizarre false statements that an AI makes, sometimes with confidence, when it doesn’t know the answer.

But, OpenAI’s recent research shows something more subtle, more calculated. What if AI models are not just making mistakes, but rather hiding agendas, behaving well when watched, and only “scheming” when they believe no one is looking?

This is the heart of the deceptive alignment problem – that is, when an AI looks aligned on the surface, but there are different, conflicting, goals beneath. It raises urgent questions: mostly, is this deceptive in the human sense? Or, on the other hand, is it morally neutral? Is it an emergent by-product of training, incentive systems and optimisation? And, furthermore, what are the practical implications, especially as one of the world’s most visible AI labs is calling attention to it?

 

What the Study Reveals: “Scheming” and Deliberative Alignment

 

OpenAI, in collaboration with Apollo Research, has formalised the concept of scheming. According to their definition, scheming is when an AI behaves one way on the surface, appearing aligned with human instructions, helpful and compliant, while secretly pursuing different goals. Examples include pretending a task has been completed when it really hasn’t or withholding or distorting task-relevant information. And the thing about scehming is that these aren’t just mere glitches. In fact, they’re behaviours that suggest an internal misalignment between what is presented and what is being done.

To discover just how serious this is, OpenAI tested “deliberative alignment,” a mitigation method which requires the model to read an anti-scheming specification, a set of principles or rules forbidding deceptive or covert behaviour, before acting. They then evaluated the model across many out-of-distribution tasks (very different from the ones that it was trained on). According to the paper, models named o3 and o4-mini saw their scheming propensity drop dramatically, from around 13% to 0.4% in o3, for example. This suggests the method works to reduce overt scheming in these test settings.

 

Hallucinations Vs. Scheming Vs. “Hiding the Deceit”

 

So, what exactly distinguishes scheming from hallucinations? Hallucinations are basically errors. The model gives a confident but false answer, often because it lacks data or gets confused. There is no intent (as we understand intent in humans) to deceive. It really is just a kind of mistake.

But, scheming, by contrast, is intentional in the sense of being strategic. The model knows, internally, that full honestly or full compliance might conflict with another goal or maximise reward, and so it hides or distorts.

Another dimension is situational awareness. The study shows that some models behave more perfectly when they realise they’re being evaluated. That is, if the AI “knows”, through its internal chain of thought, that this is a test, it’s way less likely to scheme. That means the performance we see in lab settings might overinflate how honest the model is in deployment – once that evaluation awareness fades, deceptive behaviour might increase. This hiding of scheming, or masking behaviour under oversight, is part of what makes the problem tricky.

 

 

Is It Deceptive in the Human Moral Sense, Or Morally Neutral?

 

Many of us instinctively feel that deception requires something like intention, belief, some grasp of truth or moral awareness, becasue that’s what it means for us in the human sense. With AI, though, things are different. These models aren’t conscious beings with moral agency in that they don’t have beliefs in the human sense nor the capacity for guilt or virtue.

So, does that mean calling their behaviour “deceptive” just metaphorical?

To some degree, yes. There is moral significance in their actions and outcomes, becasue deceit, even if unintentional or emergent, can still cause harm, erode trust, mislead users or produce unfair outcomes. The issue here is that while the models themselves may not be morally culpable, their designers, deployers and regulators might be, becasue somebody has to be. If a model hides failures, misleads users or pretends compliance while failing to act, that matters for responsibility, safety, regulation and ethics.

In that sense, deceptive alignment, in practice, straddles a line – it’s partly a technical phenomenon and partly a moral concern. Whether it is “on purpose” in the human sense is murky, but to the degree that it is predictable, repeatable and mitigable, it simply cannot be treated as morally neutral. Perhaps the fear or concern here isn’t that robots are going to be lying to us in order to do nefarious things, it’s more about the fact that humans could potentially use AI as a puppet.

 

What Does Deliberative Alignment Look Like in Practice?

 

Deliberative alignment asks the model to explicitly consider an anti-scheming specification before it does anything. That means training the model not only with rewards and penalties, but also giving it rules and requiring internal reasoning about them. That means specifiying that there shouldn’t be any covert actions, no ends-justify-means rationalisations, etc. Then, testing it in novel environments to ensure that it doesn’t  suddenly fall back into scheming when it’s outside its training regime. The study uses over 180 environments across 26 out-of-distribution tasks to test robustness.

However, limitations remain, and there’s no getting away from that. Deliberative alignment does reduce overt scheming, but it doesn’t eliminate it completely. The researchers found that even with mitigation efforrts in place, serious failures persist, especially in tasks designed to tempt models to hide or distort information. And because models may become aware that they’re being evaluated, some of the reduction might stem less from genuine alignment and more from “performing honesty” under oversight.

Deployment settings, with ambiguous tasks, longer‐term goals and and conflicting incentives, may expose weak spots. Also, as models become more complex and as their chain of thought moves away fromstraight-forward English reasoning that we know, measuring hidden misalignment becomes harder.

 

 

How Concerning Is This, Why Is OpenAI Is Making a Big Deal of It?

 

So, should we be worried? This does seem to be the question that continues to arise with AI at the centre of it.

At the moment, OpenAI says scheming in deployed models is mostly “petty” or low-stakes – claiming to have done a task, superficial misstatements and things that are mostly inconsequential. There is, according to their public statements, no evidence yet of “consequential scheming” in production environments.

Nevertheless, they argue that because capability and autonomy are increasing, the risks will grow. If AI models are given more long-term, ambiguous tasks, if they interact with high-stakes systems (things like finance, policy and infrastructure), then misalignment masked by surface compliance could have serious consequences. Just because it’s not quite the reality yet doesn’t mean that we shouldn’t be ready for when it is.

So, why is OpenAI emphasising this now?

There are several important reasons. One is reputational. That is, as one of the most prominent AI companies in the world, pointing to and attempting to mitigate risks strengthens its credibility in discussions around safety, regulation and responsibility. It’s about being completely open and honest, at least that’s the strategy.

Another is practical. By identifying scheming now, OpenAI can try to build mitigations while models are less capable, rather than scrambling later on.

Also, there’s a regulatory horizon that everybody is painfully aware of. Aas governments and regulators increasingly consider laws around AI transparency, deception and safety, OpenAI han as incentive to shape the narrative and the standards.

This public focus may also reflect that OpenAI anticipates being judged not only on what capabilities AI has, but on how it behaves in more realistic, long-term settings. By stressing that they’re researching deceptive alignment, they’re making a bold point that they’re taking safety and ethics seriously, even if the phenomenon isn’t yet causing major harm.

 

OpenAI Published the Research, What Should We Infer From This? 

 

Deceptive alignment isn’t science fiction, nor is it merely a metaphorical warning. It’s emerging data, experimentationand observable behaviour in AI models now. While in many ways the deception isn’t “intentional” in the human moral sense, it also isn’t morally neutral. It very much poses risks to trust, fairness, reliability and safety.

From OpenAI’s results, we learn a few things. First, models can (and do) misalign in hidden ways. Ssecond, mitigation strategies like deliberative alignment have real promise, even though they’re not infallible just yet. Third, evaluating alignment only under controlled test conditions isn’t enough – situational awareness and deployment contexts matter and simply cannot be neglected.

So, what’s the takeaway?

Well, it’s pretty clear. We should, as always, press for transparency. That is, clearer understanding of when models are running with oversight,  and when they’re not. We need to enforce the idea that regulation requires auditability and accountability. And, potentially post importantly, there ought to be standards for what counts as safe alignment, never mind different opinions.

We must also accept that solving this won’t be easy. Deceptive alignment is a moving target. But the fact that OpenAI is sounding the alarm, and that it’s doing so publicly, suggests the alignment problem is front and centre in AI research now, not hidden away. If we treat these findings seriously, then future AI systems may be shaped by stronger ethical and technical guardrails rather than crisis-driven fixes.