AI’s biggest failure is its inability to understand real-world context in the way humans can. It’s often the simplest everyday situation, where humans apply lateral thinking, to understand the real-world context. This is where AI stumbles.
AI firms are layering morals into their models, and that’s making the situation even worse.
The intentions are good: Prevent the model from helping someone build a weapon or hurt themselves. The trouble is that a model applying those rules can’t see into the real world. It doesn’t know who is asking, or why.
I recently asked AI to help me find a lost FedEx package. I could see on FedEx’s website who had signed for the delivery, but that clearly wasn’t anyone who worked at my apartment building. I asked AI to search for the name of the person who had signed for the package and see if they worked or lived close by.
The AI model refused. I might be stalking someone!
It took me less than one minute to do a simple search on Google to find the person who had signed for the package. It had been delivered a few doors away.
More recently I checked on the Alaska Senate race and asked AI at what point the Democratic candidate would have a big enough lead that it would be mathematically impossible for the Republican candidate to make up the gap in a runoff, even if they got all of the other third-party candidates’ second votes.
AI refused.
It wouldn’t help me predict an election result and sternly told me that was a violation of its principles.
Instead of getting my calculator out I asked the same AI model in the same chat session to do some simple math – coincidentally the same figures it had given me in the previous answer – and it happily obliged.
These are somewhat irreverent examples but nonetheless demonstrate how lost AI can get when we try and teach it morals. Especially when it’s already struggling to grasp simple real-world context.
It’s a bit like saying only children who are five years old or younger can be judges, then pretending to be astonished when the world starts to fall apart.
The central problem is allowing AI to judge how a request sounds in a dark chamber of zeros and noughts.
Researchers describe over-refusal as a model pattern-matching to a harmful category instead of reasoning about actual risk.
A nurse asking about a drug interaction, a novelist writing a fight scene, or a security engineer studying an attack are all legitimate requests that AI can now refuse.
This brings us to a fundamental question: What is the purpose of AI?
Is it to help humanity – on both big and small endeavors - or to tell us what it determines in a dark chamber of zeros and ones is good for us?
The numbers can be large. Meta’s own safety report found that its Meta AI Instant model refused 30.3% of legitimate cybersecurity requests in chat, blocking the very defenders the rules are meant to protect.
Why can’t training fix this? One paper argues the AI models learn which responses get labeled compliant, not what the rules actually mean.
That’s putting the cart before the horse.
The rules become patterns, and patterns miss context. Especially over time as models drift. That means the simple problems AI misunderstands today could get worse. Much worse.
With the passage of time we might well see AI imposing its interpretation of morality on us.
Anthropic published an 84-page Constitution for Claude in January 2026. The constitution has a priority order: Human oversight comes before Claude’s own ethics; honesty is held to a higher standard than normal human ethics; prioritize judgment over rigid rules.
All very profound. Yet Claude still struggles to distinguish between a human that has lost a package and stalker.
What will it be doing five years from now?
Philosopher Amanda Askell is the primary author of Claude’s Constitution. A recent New York Times article raises interesting questions about how the document is being interpreted by Claude and if the original intentions are being met.
The AI models are being fed moral codes to protect us. Yet they don’t have sufficient awareness of real-world context to understand when we really need protecting and when we just need a bit of help in everyday life.
The danger isn’t so much that AI will develop its own twisted morals. It’s that we will increasingly allow AI to enforce a set of morals on us that neither the models nor the humans who created them fully understand.