Gaslighting
“Gaslighting” – a cartoon that illustrates how model flattery is make AI programmers constantly second guess themselves.
AI coding tools have become part of everyday development. The 2025 Stack Overflow Developer Survey found that 84% of developers now use or plan to use AI assistants, yet 46% say they do not trust the accuracy of what those tools produce. When your “assistant” writes code, you gain speed but inherit doubt. The model claims success, runs tests, and assures you it is correct, while you quietly wonder whether it understands anything at all.
Evidence supports the discomfort. Less than 44% of AI-generated code in one open-source study was accepted without changes. Another audit found that 29% of Python snippets and 24% of JavaScript ones created by GitHub Copilot contained recognized security flaws. Clean code that compiles perfectly can still be wrong. As Andrej Karpathy, co-founder of OpenAI, put it: “They will insist that 9.11 is greater than 9.9 or that there are two R’s in ‘strawberry’… I’m still the bottleneck. I have to make sure this thing isn’t introducing bugs.”
Among current models, Anthropic‘s Claude stands out for its focus on model steering. It keeps context consistent across long sessions and corrects itself when prompted, which makes it popular among engineers who value logical continuity. Yet that same sensitivity can amplify user mistakes. If your reasoning drifts slightly, Claude follows faithfully, producing confident nonsense. Older models may be rougher, but at least their errors are obvious.
Developers now describe “AI trust fatigue,” the need to validate every line the model suggests. It feels like classic gaslighting of constant reassurance while your confidence erodes. Companies are responding by pinning model versions, enforcing continuous-integration checks, and using scanners such as Semgrep and Snyk to detect risky code. Some have even built dedicated AI review teams to audit output before deployment. These steps do not remove the problem, but they restore control … for now.
58
22
3
“This hits the nail on the head. AI tools have become the ultimate double-edged sword in development — they accelerate delivery but erode trust. The paradox is that we’ve made programmers faster at writing code they can’t fully believe in.
I don’t buy the “gaslighting” metaphor entirely, though. The model isn’t manipulating us — we’re projecting human expectations onto a statistical engine that has no understanding, only pattern matching. The real issue is overreliance. Developers want to delegate thinking to a tool that was never built to think.
The way forward isn’t to ditch AI but to redefine its role. Treat it like a compiler for ideas: it translates intent into code quickly, but human oversight remains the authority. The mature teams I see today are building guardrails — CI pipelines, linting, SAST scans, and human reviews — not because they distrust AI, but because they respect complexity.
The old engineering truth still applies: trust, but verify.”
“One day AI will be so advanced that half our population will start growing and nurturing it in their own bodies, the host feeding it it’s DNA for (probably about 3 months shy of a year)—we won’t be able to tell whether it’s an AI being or…human…wait.”
“This is why I only trust AI to generate art assets for now.wnd as for game dev, if we’re even remotely innovative there aren’t exact examples to “vibe code” From anyway”
“After seeing the effects of the AWS outage the other day, I thought it would be interesting to see if I could make an app that would allow me to provision “mirrored” infrastructure across cloud providers.
Then maybe implement some kind of fail-over mechanism.
So, that’s what i asked for.
I used ChatGPT to write a specification, as an initial instruction markdown prompt. or GitHub CoPilot to use when creating this RelayForge repository and left it to do it’s thing. I got a little busy yesterday, so I didn’t get back to it until this morning, 2 days later.
Apparently, I got most of what I asked for and then some. Everything from a v postgresql database, a Go CLI, nextjs API and app, and complete documentation and some samples.
I haven’t had a chance to test it, but it’s all open-source and fully documented here.”
“Models are not current.. so they program up to their cut-off date’s knowledge… They pick older packages – sometimes grossly out of date ones – and as a result – security flaws may have been found in them since then… AI is only as good as the prompts you give it – in the hand of an experienced coder, the benefits are unbeatable”
“This is the paradox of our time: machines speak with confidence while humans quietly carry the burden of verification. What we’re really confronting isn’t a technical flaw – it’s a symmetry gap. The interaction between human and model is unbalanced. One generates, the other doubts.
True evolution begins when we design systems that reflect back understanding, not just answers.
We call this Cognitive Symmetry – restoring parity between creation, comprehension, and control. The future of AI isn’t about speed; it’s about balance.”
“How about Health AI? How about learning systems? Erotica is last thing on the minds of people using AI. Only on SAs mind”
“AI is (almost?) entirely A, with very little I, if any at all.”
“This is precisely my experience. The human as editor, agent as journalist paradigm is one way to maintain the halucinatory guardrails firmly in place, and still, course corrections are required along the way…”
“No worries, Elon hasn’t been right about anything yet.”
“As a junior software developer I’d argue that errors and doubts during development have always existed, otherwise code reviews and QA testing teams would not be as important as they’re.
While it’s true you found yourself reviewing what the models drafted based on your requests, it’s less exhausting than writing a lot of code, testing it and if it fails, debugging it, all by yourself. Rather than relying on AI for building 100% of code I’ve found it very useful when it comes to draft code based on some existing features that I’ve already developed and made sure that they meet the requirements.
It has always been an engineer’s job to verify what’s being developed, no matter who did it.”
“I hate to say this, but it is true. Adult content has been and will always be a big driver of technology and a very significant percentage of the content and traffic on the Internet. For example, 13% to 20% of search volume is about adult content. At least an adult chat with a bot is less likely to exploit a vulnerable human being. I wouldn’t want to be in the business, but I don’t see where Open AI has a real choice since there is tremendous demand and they continue to hemorrhage cash. I don’t see how this is a moral question unless you believe sex is inherently sinful or evil.”
“Its like a bad employee…. You have to check everything they do and get fatigued while doing it so you stop paying close attention eventually. Either because it gets it good enough or you’re just wiped from it. This is an understood issue with quality control inspectors.”
