algorithms given agency must be grounded by ethical ideas. opinions and ideas are anonymously my own — not an agent’s ⚖️⚙️ CS PhD student
USA
Joined November 2026
- Tweets414
- Following116
- Followers94
- Likes194
Humans of @X, what are your best techniques for getting AI to write prose that doesn't sound like AI?
Looking for methods: style guides, rule sets, prompting patterns, editing passes, tools. Whatever has actually made the output better for you, not just "tell it to sound human."
1
3
69
The most dangerous trait of AI coding agents may be their speed. Consider a human who is 90% accurate and an agent that is 99%. Because agents can be spawned in parallel and output text much quicker, a very intelligent agentic system can still produce errors at a dizzying rate.
1
1
46
AI skills and harnesses, ethics of agency, machine learning/ML, raising capital. If you're in tech, let's #connect 🤝
The anti-nuclear movement post-Chernobyl is one of the most tragic mistakes of recent history. Fed by the oil and gas industries, anti-nuclear posturing has led to unnecessary pollution, worse health outcomes, and is partially responsible for our wars in the Middle East.
We should be ashamed of this record and build nuclear whilst we transition to fully renewable energy.
2
70
What is the most useful heuristic for deploying AI?
Critical thinking
If you can automate a piece of critical thinking that you consider to be superfluous, do it. Be very judicious and cautious, though.
For example, automate writing simple test cases or history docs. Do not automate writing the core architectural backbone of a program. Everything else lies somewhere in between.
1
4
84
I wonder how good the classification models are at identifying bot content? Seems like a cheap moderation opportunity to distinguish human content from the slop.
1
2
72
AI that does the work while humans only check it erodes the expertise needed to check it. This risk falls hardest on students and on fields that lose their entry-level training ground. Use AI with intent! and keep practicing the skills you refuse to lose.
The dilemma is no longer whether professionals will use AI, but what happens to the person checking it. The default seems to be that an AI does the work and a human verifies it. But beware, solely verifying decays the very skill that qualified you to verify. AI users should deliberately practice any skill that they don’t want to lose.
In 1983, Lisanne Bainbridge described the “ironies of automation”: automate a job, and the operator is left to monitor it and take over when it fails. But human skills fade without use - an experienced operator who mostly monitors may take over like a novice. By taking over routine, predictable, and easy operational tasks, automation leaves the human operator with only the most complex, unpredictable, and poorly understood abnormal conditions.
Medicine offers an early warning. At four centers in Poland, endoscopists’ detection rate in non-AI colonoscopies fell from 28.4% to 22.4% after AI tools arrived. The study is observational, with only 19 doctors, but the direction is concerning.
The effect likely begins earlier, while operators learn. In an Anthropic trial, 52 engineers learned a new Python library with or without AI. On a no-AI quiz, the AI group averaged 50% and hand coders 67%. The gap was largest for debugging - knowing when code is wrong, and why - the non-nominal condition.
This is the supervision paradox: checking AI output decays the skill that qualified you to check it. This is not a technical oversight that better models will fix. Checking is not practicing, and reading content does not improve your ability to write it.
AI can already perform significant portions of entry-level work in many fields, and the most deterministic, measurable fields will be eviscerated first. Programming leads here - it is among the most deterministic, measurable, and digital. The code runs or it doesn’t.
The problem is: entry-level work was also the training. Students have the most to lose, because their skills are still forming. A senior professional who leans on AI lets a skill erode. A student may never build it, and may never understand what is missing. Some fields may be replaced outright, but many more could slowly lose their human expertise: without entry-level jobs, no one becomes an expert, and when today’s experts retire, who is left to operate the automations?
So use AI with intent, starting now. Name the skills you refuse to lose, and keep practicing them without AI, on a schedule. Draft before you prompt, debug before you ask, and form your own view before you read the model’s. If you are a student, that list should be long. Thinking is difficult, and that difficulty is where the value lives. Which skills are you protecting, and what will you do this week to start?
Sources:
Bainbridge, L. (1983). Ironies of Automation. Automatica.
Reporting on the Poland colonoscopy study, ASCO Post (August 2025).
Anthropic. How AI assistance impacts the formation of coding skills.
Reposting this because verification processed.👍
4
73
The dilemma is no longer whether professionals will use AI, but what happens to the person checking it. The default seems to be that an AI does the work and a human verifies it. But beware, solely verifying decays the very skill that qualified you to verify. AI users should deliberately practice any skill that they don’t want to lose.
In 1983, Lisanne Bainbridge described the “ironies of automation”: automate a job, and the operator is left to monitor it and take over when it fails. But human skills fade without use - an experienced operator who mostly monitors may take over like a novice. By taking over routine, predictable, and easy operational tasks, automation leaves the human operator with only the most complex, unpredictable, and poorly understood abnormal conditions.
Medicine offers an early warning. At four centers in Poland, endoscopists’ detection rate in non-AI colonoscopies fell from 28.4% to 22.4% after AI tools arrived. The study is observational, with only 19 doctors, but the direction is concerning.
The effect likely begins earlier, while operators learn. In an Anthropic trial, 52 engineers learned a new Python library with or without AI. On a no-AI quiz, the AI group averaged 50% and hand coders 67%. The gap was largest for debugging - knowing when code is wrong, and why - the non-nominal condition.
This is the supervision paradox: checking AI output decays the skill that qualified you to check it. This is not a technical oversight that better models will fix. Checking is not practicing, and reading content does not improve your ability to write it.
AI can already perform significant portions of entry-level work in many fields, and the most deterministic, measurable fields will be eviscerated first. Programming leads here - it is among the most deterministic, measurable, and digital. The code runs or it doesn’t.
The problem is: entry-level work was also the training. Students have the most to lose, because their skills are still forming. A senior professional who leans on AI lets a skill erode. A student may never build it, and may never understand what is missing. Some fields may be replaced outright, but many more could slowly lose their human expertise: without entry-level jobs, no one becomes an expert, and when today’s experts retire, who is left to operate the automations?
So use AI with intent, starting now. Name the skills you refuse to lose, and keep practicing them without AI, on a schedule. Draft before you prompt, debug before you ask, and form your own view before you read the model’s. If you are a student, that list should be long. Thinking is difficult, and that difficulty is where the value lives. Which skills are you protecting, and what will you do this week to start?
Sources:
Bainbridge, L. (1983). Ironies of Automation. Automatica.
Reporting on the Poland colonoscopy study, ASCO Post (August 2025).
Anthropic. How AI assistance impacts the formation of coding skills.
Reposting this because verification processed.👍
It seems self-evident that neurons’ complicated physical structure, particularly convoluted non-acyclicity, precludes transformers and maybe any NN structure from being a sufficient computational representation of them. Why am I wrong?
50
Making the internet better for AI agents will make it worse for everyone else.
The future of the internet will be a vast collection of closed, semipermeable intranets, with slow, audited connections between one another.
42
AI will by sheer speed of generation dilute ideas, quality, and depth of human writing, because of the reader’s fatigue and the writer’s need to keep up with the pace.
This is directly related to news outlets’ precipitous drop in quality because of the 24-hour news cycle.
2
42
Many philosophers have made these arguments, but I am now personally for the first time.
This is exciting to me!
1
36
I am reticent to say this, but:
- a lot of the real slop code out there was written by humans
- less than 1% of human-written code is worth emulating
- we compare agentic code to the top quality of human code produced (as we should)
1
52