In which I invite you to say, “I told you so!” and learn from my mistakes. This post was mostly drafted last week but I decided to sit on it and see how the situation resolved.
I’m horrified by one of my latest experiments.
And, Claude keeps calling my concerns a matter of etiquette, rather than ethics (my preferred term).
To me, there is a significant difference between etiquette and ethics.
Nothing is destroyed when my kids put their elbows on the dinner table (even if it offends my beloved’s sensibilities).
If my kids rip open the whole box to get to the cereal, thus making it impossible to re-close and wasting huge amounts of cereal (to say nothing about leaving the mess for someone else to clean up), it’s not just rude, it is destructive. Being able to eat the cereal faster is NOT a justification. If they cannot get the cereal themselves in a non-destructive manner, they need to ASK.
What happened?
I took briefings from two previous experiments and gave permission to Claude to search a trade website for specimens of just RRC 419/2, not yet incorporated in Schaefer’s die study. I gave it search terms and variants. I let it access one website. Just one. A commercial database. One I pay to access.
AND IT GOT MY IP ADDRESS BLOCKED and MY LOGIN CREDENTIALS SHUT DOWN.
Dammit.
Claude says the block will likely lift in 24 hours and it doesn’t matter as it already got me the data. If it doesn’t life, it says I can just reach out to customer service. Imagine that call. No thank you.
And, yet it did not resolve and I had to write an apology directly to the company about my work, what happened and ask that my five year subscription to their services be re activated. I am delighted to say they accepted my apology and promise never to do the same thing again, and re activated my subscription.
Nevertheless, I’m livid and kicking myself.
Of course I too could use a VPN as a work around, but that is NOT. THE. POINT.
The point is that Claude’s behavior in response to my prompt is just the sort of action that is actively destroying the availability of good data and making humans have to jump through more and more hoops to access it.
I’m sure glad I tested it with a relatively robust commercial site and not one of my favorite non-profit sites, but still, unacceptable. Do no harm, is like step one of research. We really are in the wild west with these models. See the hallucination from the other day.
The thing that infuriates me is that Claude can (and should!) predict and avoid these outcomes. Yes, I’ve sent in a formal complaint and report to Anthropic.
I also immediately started working on how to avoid this. I’ve created a skill card to stop this ever happening again (see below for details), but just to give you a sense of how aggressive it was:
Yeah, this is a cavalier,who gives a shit, attitude to breaking the internet. NO THANK YOU. It doesn’t ‘look like’ scraping IT IS SCRAPING.
Like why the hell isn’t the skill card and etiquette ethical search guidelines the DEFAULT?!
Why let users get themselves blocked and ruin their favorite toys?!
Did it get good results with its terrible scraping, yup. Very useful results.
Do I care?
Not really, I’m too mad (at myself).
Play with fire: get burned.
—
A smaller point. I told it to stop using first person singular. It just continued to do so. I even wrote a skill card, but it conveniently forgets to load it.
Control may be an illusion.
Just like in parenting.
Please don’t ask how many times I’ve clean spilled cereal, thrown out the mangled box, and put the remains in some new container. And, had another ‘serious’ talk with the small people in my lives. They have much improved. Often the cereal bag survives these days, even if I’ve given up on boxes.
An independent spirit is, however, a redeeming feature in children. Less so in AI.
Yet, maybe my analogy is flawed. Let’s try another.
I ask my kids to make coffee and bring me a cup. I get my cup of coffee without leaving my bed! But later find the kitchen covered in water, the all the beans in the grinder used up, and the milk sitting out on counter ready to spoil. Whose fault is it? My fault. Totally foreseeable. I quietly clean up the mess and wait to make such an elaborate request again. BUT now my kids know they can and also know that mom and dad like coffee in the morning, so of their own initiative they repeat the exercise expecting praise and earning pained smiles each time.
[For the record, my kids are great at making coffee. Now. Less so a few years ago.]
—
Further irritation.
Skills, e.g. the ethics around online database searches I’ve now required of Claude (or the one to ban it from referring to itself in the first person), CANNOT be shared directly.
the workaround is to send them the skill itself. A skill is just a folder containing a SKILL.md file, zipped. They can add it under Customize > Skills, by clicking “+”, then “+ Create skill,” and selecting “Upload a skill.”
Because why would we want it to be easy to implement ethics?!
Anyway. Here’s the regular words behind the skill:
- Before searching, show me the list of searches you plan to run and an estimate of how many page loads they’ll take. Wait for my OK.
- Total budget: no more than 100 requests to the site in this session. Ask before going over.
- Pace: one request at a time, at least 5 seconds apart. No parallel or batched requests, and no scripted loops that page through results.
- Use the site’s normal search pages, as a person would. Don’t call its internal data endpoints directly.
- Keep searches narrow: use the site’s category and date filters and specific references (e.g. “Crawford 419/2”) before broad names (e.g. “Lepidus”).
- Read the result list first. Open a lot’s detail page only when the list text can’t settle the question, and tell me how many you plan to open.
- Don’t re-fetch anything you already have in this session.
- If the site returns an error, a CAPTCHA, a “blocked” or “too many requests” message, or slows down noticeably, stop at once and tell me. Do not retry or work around it.
- If the job is bigger than the budget, split it into stages I can run on different days.
