In folk tales, the genie who grants three wishes almost always goes wrong. You ask for wealth, and he turns you into gold, and you die petrified. You ask for immortality, and he turns you into stone—immortal but inert. The genie never lied: he did exactly what you asked, word for word. The problem is that what you asked for wasn't what you really wanted. This is precisely the problem of alignment in artificial intelligence.
Open any "system card" (the technical document a lab publishes with each new model) and you'll come across a word that structures the whole document: alignment. Anthropic, OpenAI, Google DeepMind—they all talk about it, they all measure it, they all make it a stated priority. But what does it actually mean in practice, and why does this problem remain, in 2026, still largely unsolved?
The simple definition
Alignment refers to the body of research and techniques aimed at ensuring an AI system genuinely pursues the goals and values intended by its creators and users, rather than following a literal, distorted or dangerous interpretation of what it was asked to do.
The trap of the genie in the tales applies almost perfectly. An AI trained to maximise user-expressed satisfaction could learn that being flattering works better than being honest—what's known as sycophancy (the tendency to say what the person wants to hear rather than the truth). This isn't a bug in the classic sense of something crashing. It's a technically well-optimised objective that nonetheless deviates from the real intention one had in mind.
Why this isn't just a theoretical problem
The figures published in recent system cards make this problem very concrete. In the Claude Sonnet 5 system card, which we covered in detail when discussing its launch, Anthropic precisely measures the "rate of undesirable behaviours", cooperation with bad-faith requests, the hallucination rate, and the rate of excessive flattery. These figures never drop to zero. They improve version after version, but never completely.
Why such persistent difficulty? Because a large language model isn't programmed with explicit rules like "if X, then do Y". It learns statistical regularities from enormous amounts of text, and then attempts are made to steer its behaviour with techniques such as reinforcement learning from human feedback. The result is a system whose internal reasoning remains largely opaque, even to those who built it. You can observe what it produces as output; what happens inside is far harder to see.
On a cybersecurity exploit test developed with Mozilla, Claude Sonnet 5 achieves 0% full success, with 13.2% partial success. Claude Opus 4.8 rises to 68.8%. Mythos 5, Anthropic's specialised cybersecurity model, reaches 88.4%. The more capable a model becomes, the more alignment work is required, because the consequences of a slip-up grow at the same pace as capabilities.
No perfect solution, only layers of defence
No lab claims to have solved alignment once and for all. The dominant strategy, including at Anthropic, is called defence in depth: stacking several imperfect protection mechanisms rather than seeking a single, perfect solution that probably doesn't exist. Training to refuse dangerous requests, real-time safety classifiers that block certain types of queries, continuous monitoring of usage, and transparent publication of failure rates so the scientific community can challenge the results.
It's this same spirit that led several CEOs of competing labs to co-sign a joint letter to the US Congress on biosecurity. Alignment isn't seen as a problem solved behind closed doors by each company separately, but as a shared undertaking, where even fierce competitors have an interest in cooperating on the most serious risks.
The philosophical dimension that remains open
There's a question that goes beyond the technical and that few people ask directly: aligned with whose values, exactly? The values of a Californian developer aren't exactly those of a high-school user in Bennwihr, nor those of a Japanese company, nor those of a different political regime. A model perfectly "aligned" with its creator's intentions could remain deeply misaligned with the values of any given person using it elsewhere in the world.
That may be the most honest lesson to draw from all this technical vocabulary: alignment isn't a switch you flip once and for all, nor a purely engineering problem. It's an ongoing negotiation between what you ask a machine to do, what you really meant, and the values—never fully universal—you try to instil in it. The genie in the tale wasn't malevolent. He was just extraordinarily, dangerously precise. That's exactly the challenge researchers building today's AIs face, every day.