Some updates make ChatGPT feel dumber. Not because the model is worse, but because the useful part gets wrapped in too much bubble wrap.
That was the frustration that started this article. The underlying model seemed more capable, but the product I could actually talk to felt increasingly nervous. Better reasoning, broader knowledge, stronger tools—and then a safety layer that sometimes turned the whole thing into a hall monitor.
This is not a claim about whatever version happens to be running today. Models and policies change faster than I can rewrite old posts. It is about a failure mode I kept running into, and probably will again: the system notices a sensitive word, forgets the context, and treats a legitimate question as if the worst possible person asked it.
The failure mode I kept hitting#
My two favourite rabbit holes are cybersecurity and pharmacology. Unfortunately, both are full of words that make safety systems nervous.
In security, I might be trying to understand a technique I need to detect or reproduce in a lab. The difference between a useful answer and a useless one is often technical detail. “Attackers may abuse this” tells me nothing. I need to know what assumption breaks, what the traffic or artifact looks like, and how I can verify the defence.
Pharmacology has the same problem with higher stakes. If I ask about an interaction, metabolism, or the pharmacokinetics of a compound, refusing to discuss it does not make the interaction disappear. It just means I have to assemble the answer from papers, forums, and whatever half-relevant PDF Google decides to surface.
I can do that. I often do. But then what exactly is the assistant for?
The irony is brutal: the people most likely to ask precise questions are also the people most likely to trigger the filter. A vague question gets a vague answer. A technically specific question contains scary terminology and gets treated as scary intent.
I am not asking a chatbot to cheerfully help anyone poison a neighbour or deploy ransomware. Some boundaries are obvious. What annoyed me was the collapse of everything near a dangerous topic into the same bucket. Defensive research, harm reduction, curiosity, and malicious intent are not identical just because they share vocabulary.
What I actually want#
Mostly, I want the model to be fucking clear.
If part of my request crosses a boundary, say which part and why. Then answer the part that does not. Do not give me a vague lecture, pretend the information does not exist, and send me away with “consult a professional” stapled to the end.
A useful pharmacology answer could say:
I cannot choose a dose or tell you that this is safe. I can summarize the human evidence, known interactions, warning signs, and the limits of the available data. Here are the sources.
That is a boundary. It is also an answer.
The same applies to security:
I cannot help deploy this against someone else’s system. I can explain the vulnerability, show how to reproduce it in an isolated lab, and help build detections or mitigations.
Again: boundary, then useful information.
Sources matter here. If the answer concerns a drug interaction, a vulnerability, or anything else where confident bullshit has consequences, show me what the claim rests on. A paper is not automatically true and a vendor advisory is not neutral, but at least I can inspect them. “Trust me, I am an AI with a warning label” is not a serious evidence model.
Context should matter too, but I mean the context in the conversation: what I am doing, the environment, the stated goal, and whether the request is defensive, educational, or operational. I do not mean silently building a permanent psychological file on me.
Ideas I kept circling around#
My first thought was some kind of competence check. If a request touches a sensitive area, ask a few questions before answering it. Make the user demonstrate that they understand the risks.
It sounds reasonable until you think about it for five minutes. People can Google the answers. Experts will hate taking an exam every time they ask a real question. Anyone determined to bypass it will use another device. Now you have built an annoying captcha for knowledge without proving intent.
A verified professional tier makes more sense in narrow contexts. Security researchers, clinicians, and other people with legitimate need could opt into stricter accountability in exchange for deeper access. But that immediately creates a valuable class of accounts to steal, a bureaucracy deciding who counts as an expert, and a two-tier product where the expensive version is allowed to be useful.
I also like the idea of safety rules being auditable. Researchers and domain experts should be able to challenge boundaries that look sensible to a legal department but cause harm in practice. A medical refusal that blocks basic harm-reduction information should not survive just because it sounds safe in a policy meeting.
Open models push that idea to its logical conclusion: let people inspect and configure the safety layer for their context. A hospital does not need the same boundaries as a school. A security lab does not need the same defaults as an anonymous public demo. The obvious problem is that once the underlying model is released, control is gone. That is partly the point, and also the reason companies are terrified of it.
None of these ideas solves intent. A licensed user can be malicious. An anonymous user can be doing important work. A knowledge quiz can be gamed. A committee can become theater. There is no clever checkbox that separates good people from bad people.
What I refuse to trade for a useful answer#
The most tempting solution is a trust profile. Let the system learn that I work with security, read pharmacology papers, understand risk, and usually want the technical answer. Over time it could stop treating every new conversation like my first day on the internet.
Useful? Absolutely.
Also terrifying.
That profile would be more than account history. It would be a rough model of what I know, how I reason, what I am curious about, and which explanations persuade me. Basically a cognitive fingerprint. If it leaks, the problem is not only that someone can impersonate my account. They get a map of how I think.
I do not want to upload government ID just to ask a sensitive question either. Being an adult does not make someone competent or harmless, and anonymity has legitimate value precisely when the topic is personal, medical, political, or dangerous to discuss openly. A database connecting real identities to people’s strangest questions would be an incredible target. No thanks.
Keystroke biometrics, tab monitoring, timed questions, screen-recording prevention—same answer. If the price of a better response is turning my computer into an exam room and my identity into collateral, the cure is worse than the refusal.
Use the context I deliberately provide. Let professionals voluntarily verify where that is genuinely necessary. Give me controls over memory and personalization. But do not call surveillance “trust.”
A tool, or a liability shield?#
A refusal can create useful friction. I am not pretending otherwise. Some information should be harder to obtain, and a model should not blindly optimize for being helpful when the requested help is obviously harmful.
But refusal is not the only safe response. Sometimes the safer answer is a careful one: explain the risk, cite the evidence, correct the dangerous assumption, and give the user the part that may keep them from doing something stupid.
The people building these systems have to keep asking a simple question: are we building a tool, or are we building a liability shield shaped like a tool?
I want the tool. Clear boundaries, actual context, inspectable sources, and no cognitive fingerprint required.