There’s a great debate these days about what the current crop of AI chatbots should and shouldn’t do for you. We aren’t wise enough to know the answer, but we were interested in hearing what is, apparently, Microsoft’s take on it. Looking at their terms of service for Copilot, we read in the original bold:
Copilot is for entertainment purposes only. It can make mistakes, and it may not work as intended. Don’t rely on Copilot for important advice. Use Copilot at your own risk.
While that’s good advice, we are pretty sure we’ve seen people use LLMs, including Copilot, for decidedly non-entertaining tasks. But, at least for now, if you are using Copilot for non-entertainment purposes, you are violating the terms of service.
Legal
While we know how it is when lawyers get involved in anything, we can’t help but think this is simply a hedge so that when Copilot gives you the wrong directions or a recipe for cake that uses bleach, they can say, “We told you not to use this for anything.”
It reminds us of the Prohibition-era product called a grape block. It featured a stern warning on the label that said: “Warning. Do not place product in one quart of water in a cool, dark place for more than two weeks, or else an illegal alcoholic beverage will result.” That doesn’t fool anyone.
We get it. They are just covering their… bases. When you do something stupid based on output from Copilot, they can say, “Oh, yeah, that was just for entertainment.” But they know what you are doing, and they even encourage it. Heck, they’re doing it themselves. Would it stand up in court? We don’t know.
Others
Now it is true that probably everyone will give you a similar warning. OpenAI, for example, has this to say:
- Output may not always be accurate. You should not rely on Output from our Services as a sole source of truth or factual information, or as a substitute for professional advice.
- You must evaluate Output for accuracy and appropriateness for your use case, including using human review as appropriate, before using or sharing Output from the Services.
- You must not use any Output relating to a person for any purpose that could have a legal or material impact on that person, such as making credit, educational, employment, housing, insurance, legal, medical, or other important decisions about them.
- Our Services may provide incomplete, incorrect, or offensive Output that does not represent OpenAI’s views. If Output references any third party products or services, it doesn’t mean the third party endorses or is affiliated with OpenAI.
Notice that it doesn’t pretend you are only using it for a chuckle. Anthropic has even more wording, but still stops short of pretending to be a party game. Copilot, on the other hand, is for fun.
Your Turn
How about you? Do you use any of the LLMs for anything other than “entertainment?” If you do, how do you validate the responses you get?
When things do go wrong, who should be liable? There have been court cases where LLM companies have been sued for everything, ranging from users committing suicide to defaming people. Are the companies behind these tools responsible? Should they be?
Let us know what you think in the comments.
This is a good portrait of sober AI use, sans evangelism and utopian idealism. I use Google AI studio, it is a beautiful web interface. Something like Gemini but more IDE esque imo. Gemini is not so useful, collating my prompt history data and plopping it arbitrarily into any old prompt is useless to me. I mainly use LLM for entertainment, I have it write scripts for me when I get bor3d. 5% of these scripts hit a margin of quality i am pleasef with, the othet 95% of the time I am reprompting to inch closer to the desired output.
Not for anything production. Not for code.
Just for faster research, like clicking through wikipedia (18 years ago).
On topics that I actually am proficient, the results vary between too much babble and false information. So I assume it is just as wrong on any other topic.
The only benefit is that you can ask questions in natural language and can get some inspiration in new directions. Anything else needs proper sources.
Still prone to error so take any research with a grain of salt. Massachusetts lawyer got in trouble for using AI to draft paper citing cases that turned out to be fictional cases. He got in trouble. If your school grade or job is on the line, always double check AI generated information to be sure it’s not something from fanfiction
I can’t imagine saying stuff like this and actually taking myself seriously… Honestly, what do you think the odds are that what you said would ever be implemented? Do you make these declarations just to feel good, or do you actually buy into them? There’s absolutely zero chance that people won’t use it for code, that cat is already long out of the bag.
this is for the “free” version by the way. the whatever the heck they call the one intended for “professional” use or whatever lacks this warning.
Oh. Product differentiation? Interesting.
AI is very good at summarizing console logs from my MacOS machine. It’s kind of amazing how well it works for this. It probably doesn’t hurt that, if it’s wrong, I am unlikely to notice the error…
As more and more AI “slop” lies around on the internet, I wonder what is going to happen with future AIs that are trained on that slop. Could be a very sloppy future.
One might even say it’s a slippery slop!
(Ok, I’ll let myself out)
As time goes on, training on LLM outputs has the disadvantage training data coalescing in median outputs. Humans will connect “parameters” in novel, incorrect or serendipitous manners, whereas an LLM make thr connections along multidimensional vectors (like 300B dimensions for example, like current gen LLM).
LLM trained on LLM outputs would force these vectors into specific vector groups, the outputs would be restricted compared to training on more stochastic human training processes.
Just a crazy bit of legal trivia…the CYA sort of boilerplate often doesn’t actually cover anyones butt.
There’s a famous case from my contracts course…unfortunately i don’t have the citation in front of me…but a group of employees was trying to unionize and the employer sent out a pamphlet saying “Don’t unionize, because as it is today, you already have all of these great protections: … No need to unionize! p.s., NOT A CONTRACT.” And the employees accepted the deal implied in that paper, and they didn’t unionize, and then someone was fired without process. And the employee sued, saying they had received a promise to have a certain amount of process before being fired. And the employer said, “it says right here, NOT A CONTRACT.”
The court sided with the employee, because the words “NOT A CONTRACT” didn’t change the fundamentals. Offer, acceptance, performance. The pamphlet was clearly an offer, declining unionize represented acceptance, and continuing to work represented performance. bam it’s a contract. There’s a bunch of other cases with basically the same punchline — the reality outweighs the legalese when the legalese is blatantly false.
So when someone sues MS for the consequences of relying on copilot, there will still be all sorts of questions about ‘reasonableness’ and split liability and all that. But that disclaimer sentence won’t mean a darn thing, given that the first above-the-fold salesman claim is “Supercharge Your Productivity With Microsoft Copilot”. It’s a productivity tool and no amount of whispering in the dark that it isn’t is going to change that fact.
Cleverly leaving behind a note that says “NOT A MURDER” in magazine cut-out letters on one of my victims so they legally can’t prosecute me… I’m such a genius sometimes
“NOT A CONTRACT” = non-binding agreement, meaning either agreement party can change anything at any time without a notice, and there are no legal responsibilities per se. It is as good as verbal agreement, ie, based on trust. I am speculating that the person who was fired was identified as the troublemaker, and the rest were scared into submission, that’s how those kinds of things are usually handled. Oh, and there is no guarantee all of them won’t be fired tomorrow for no reason at all.
Contract, on the other hand, would clearly spell the details of what’s being agreed to, on what terms, what are the penalties for breaking the contract, etc. I am pretty sure when they were hired there was some kind of fine print somewhere, but it was too nebulous for the corporate lawyers to latch onto, so there.
Some courts WILL side with the employees, because they know quite well how companies regularly screw their own employees, with or without contract, and get away with these using legal technicalities (say, hiring someone full-time but NEVER giving full 40/hr workweek – ha, I’d say pretty much all US companies are guilty of that, and if the Supreme Court reviews THAT kind of class action … of course, won’t happen any time soon, but I digress ). It is really one bureaucracy vs another bureaucracy, and employees usually get caught in the legal cross-fire, sometimes it covers them, sometimes it doesn’t.
Companies also regularly hire hordes of drive-by lawyers who identify all the legal holes to exploit in certain situations, whereas ordinary people most of the times are stuck being sitting ducks because they have to work to feed themselves, literally, held by the balls.
Make an image
gladiator shouting “are you not entertained?”
But swap shiny terminator in for gladiator
I had the same thought, except I was picturing Clippy waving the gladius.
Nope. Why would I do such a stupid thing?
Everytime I see an example of LLM use on a subject I know, it is wrong. The code I’ve seen generated doesn’t inspire confidence, especially given how often people mention the LLMs “using” parts of an API that don’t exist.
LLMs are not to be taken seriously.
Yeah and I have a feeling these horseless carriages are going nowhere
I left Windows from most computing chores because of Microsoft”s AI deployment. As a Windows user, who has used Windows since 3.01, I no longer trust Microsoft.
You trusted Microsoft/IBM/Novell/Lotus/Oracle?
Beware, down the no trust road lies your own mail server, running OpenBSD.
AI for coding is basically, in practice, slightly worse autocomplete, IMHO.
With the option of starting each coding effort with a steaming random pile of AI slop.
Like running through an IDE ‘app wizard’ (or similar) w your eyes closed, rando options, then fixing.
It will run it’s course.
There are applications.
Same as last time, when was the last time you saw a non-neural net OCR?
The number of MBAs stampeding is concerning.
They really Really REALLY want ‘AI coders’ to be true (‘damn surly code monkeys’).
….Nothing else during that time period invalidated your trust?
So is Microsoft violating their own terms in their advertising? Look here:
https://www.microsoft.com/en-US/microsoft-365-copilot/business
https://microsoft.ai/news/health-check-how-people-use-copilot-for-health/
Or they do not know what they are doing, or they do misleading advertising, or what?
It’s great for boilerplate coding and debugging when I’ve made typos etc. When writing code, it cant do anything it’s not seen previously in the code it’s trawled online and recently could not do some simple logic to check if various serial ports were alive or dead that took me 15 minutes to do by hand, but at least it’s not being pretentious or hallucinating …. other than it recently called me a ‘Full stack software developer’ which was a bit strange. I just thought i was a hacker. Cant really see any legal implications as the code is always checked and sometimes it’s wrong and the LLM can usually fix it itself. If somebody does not check their code works properly, well ……
They’re literally selling this service to local governments for the purpose of coding. “Entertainment” my ass. Get sued.
I’m a little salty since I’m currently being forced at work to use a lot of AI tools, including Copilot. It’s good for boilerplate. Validating the output of non-trivial code is difficult, I have to fully understand the requirements, and it’s so much more boring than writing the code. Of course we never have the full requirements off the bat. All the test cases are with another team … and those tests aren’t available in the dev environments … So the AI can’t possibly have full context for what it’s making, not even I have full context.
Unfortunately I see this going the way of AAA game development: Outsource thinking to other companies. Hire replaceable “tool users” instead of developers. Thinking As A Service?
I’m kind of interested to watch them make the argument in court that government infrastructure is a form of entertainment. Idk might be funny
… was also excellent for reminding me how various databases, tables, php files, python scripts, C++ files etc interact with each other and which of them was not used anymore. I managed to upload about 40 files in total and the explanation was 100% correct.
Only entertainment? Well now I kinda hope that one of their enterprise partners starts a lawsuit. Their advertising campaigns of late are all geared towards the idea that Copilot can give sound financial and legal advice. In light of case law, I’d love to see how the courts take that set of promises. It’s really quite fun when the ad team doesn’t talk to the legal department.
Ask the Tennessee woman who spent 5 months in jail after falsely being identified by AI in a North Dakota fraud case. I doubt that she was “entertained”.
Eh the analog carbon units do that all the time too
Show me a single case of any law enforcement group looking at photos with their mk.1 eyeballs, seeing someone that looks like they fit a perp, ignoring that the photos came from a thousand miles away, then issuing a warrant.
Xcops make mistakes all the time.
Sometimes they do it on purpose because they are bad humans too.
But no department would waste effort looking at security camera footage of the train station at noon, while trying to solve a case of a bank across town being robbet at the same time.
It would be a massive waste of funding/time, and likely to lose the department even more money/credibility if they charged someone with a crime they couldn’t possibly have committed.
These toys MIGHT be a great help if used correctly.
Searching through a mass of data for references that can then be ready/watched/scrutinized by humans.
But letting one loose to troll unrestricted data and ‘decide’ a person fits something?
That’s insane.
I use it for many technical matters… and have caught it with errors or even stating “impossible” when that is very possible and in play. And I prove it with irrefutable observations. CoPilot simply recalibrates, states I am right, and moves on processing with the new information. Me: CoPilot, will you use my data to provide others the correct information? The respons is NO, it is not allowed to use new info to the benefit of others.
That’s the problem with the ‘AI’ bubble.
These are TOYS.
They are not TOOLS.
A chatbot is a fun thing to play with, even if it has ethical problems at every level.
It isn’t a tool to solve the ‘customer service problem’.
It isn’t a personal assistant.
All the companies pushing ‘AI’ will tell you this.
Then they will change their expression and sell you the dream.
And in private they will pray at their altar of money that some day soon this TOY might suddenly become a useful thing.
“Please. PLEASE make this useful! I have sunk all my money into this so that one day it might work.”
But it won’t.
It CAN’T, because the tech doesn’t function like that.
No amount of secret injection molding tech will make a plastic doll able to answer the office phone.
LLMs can make fantastic search engines.
They can do an amazing job at giving you context aware links back to reference material, whether it is an obscure case file, music lyrics, or a 3rd grade letter you wrote to daddy.(Assuming it was trained on that material.)
But it shouldn’t just grab the text, squish it into an answer-shaped blob, and present it as it’s own answer.
It might APPEAR like it can do the job. But that is just because the people tweaking the tuning knobs NEED you to accept it and interact with it to keep getting investment money.
This is “fake it till you make it” on the most grand scale imaginable.
But a plastic Barbie doll is never going to “make it” as a doctor, no matter how much money and electricity and prayers you throw at it.