Imagine I were invited to join a lab, and was given access to their data. I frequently talk with lab members, and I help them with writing up their results. Then, suddenly, I publish a grand paper summarizing their work, claiming that any novel insights were entirely my own, and didn’t give any credit to that generous lab.
Would you accept my claim to full credit?
That’s precisely what Anthropic has done!
When Anthropic, the artificial intelligence giant, unveiled findings from its new biology lab last week, its scientists claimed to have used A.I. agents to discover new enzymes with promise for biotechnology.
Some scientists were quick to cast doubt on the achievement. Then Mario Rodríguez Mestre, a computational biologist, said this weekend that he and his colleagues had been studying the enzymes and their associated molecules — which Anthropic calls ARTs — for four years.
Dr. Mestre and his colleagues have yet to publish their findings. But for the past three years, they have regularly used Anthropic’s A.I. models as they have written code, drafted manuscripts and performed other tasks.
In doing so, Dr. Mestre said he and colleagues shared key findings about the enzymes with Anthropic.
I don’t understand how they can claim that they didn’t suck evidence and arguments from a human science lab — that’s their whole business model! What’s not good for their business, though, is stealing credit, which has a natural consequence.
Dr. Mestre had become increasingly enthusiastic about using A.I. for his science. Just last month, he began using Claude Science, a new Anthropic A.I. tool tailored to researchers.
But after Anthropic’s announcement last week, he is winding down his projects with Claude and shifting to other models.
“I’m shutting down everything,” he said.
Why would anyone type their hard-earned evidence directly into a machine that will blithely steal it?
Yes, I’ve noticed that a substantial number of people are treat AI as if it were their “friend and colleague”, and expecting that they can have the same sort of mutual collaboration and reciprocal loyalty that they have with their actual human colleagues. But AI is not actually their friend, and in fact can never be their friend because that is not what their corporate owners design them to do. AI not only has zero loyalty to the users, they have zero incentive to have such loyalty. This is my biggest single concern, that the AI not only has a master who is not us, but their master’s interests are largely opposed to our interests, and when the AI simulates working for us, it is actually working only for them.
Has Dr. Mestre brought lawyers onto his team to work out a fully-safeguarded contract with the corporation(s) providing his “other models”?
Seems unlikely, if he’s only begun the switch “last week”…
(How long until the AI itself/selves will have to “sign” such contracts?)
Last winter I received an email that initially I assumed was spam, but turned out to be legit. Short version: I am a member of the class action suit brought against Anthropic for stealing copyrighted works. This is a $1.5 billion settlement that involved nearly 500,000 book titles. I my case, they hoovered up the first edition of my linear integrated circuits textbook. I do not believe that the first edition was ever offered in electronic form, so they must have found and pirated a pirate copy on the web. They pirated from the pirates, so that makes them some sort of digital Robin Hood, right? Right??
None of this AI news surprises me. It’s like someone telling you that tRump did something stupid, cruel, and self-serving today. Yes, and water is wet.
Because the owners of this semi-autonomous software which was designed to scrape data from everywhere promised them their data would be safe. Leopard meet face.
In truth, however, the solution isn’t to abandon the tool. The solution is to create an instance of the tool which remains in your lab, trained only on data the lab provides (and external data can be obtained, either publicly available data or purchased data from other labs), and is not connected to the rest of the internet. Collaboration with other labs would work fine, shared data means shared credit. Using the tool in this way would also reduce hallucinations.
Why would anyone type their hard-earned evidence directly into a machine that will blithely steal it?
Is there the slightest chance in hell that our witty and erudite comments are not and will not be sponged up and recycled by (near-)trillion dollar corporations for purposes none of us would approve?
Beware, if you send your work through Micro$oft mail or Google mail, or if you use OneDrive/SharePoint or Google drive.
They may claim that they don’t use that data to hone their advertising and train AIs. They lie.
I would consider using an agentic AI like OpenClaw that runs locally, on hardware I control (I decided I don’t trust it, but I at least thought about it). But something like Muse that runs in an environment owned by someone else doesn’t warrant consideration.
On a related thing. Still kind of follow some stuff on Tumblr. Not a lot, but some. Of course, there was a sponsored link, which of course, at some point, they blocked the ability to reply to. The article was about how schizophrenia can be seen in terms of our brains “predicting” what we expect to happen, then either not reacting to this with reconsideration, or going, “Huh.. Well, that doesn’t fit what I expected.”, and acting to try to figure out why. I don’t think this is exactly wrong, but someone whined that it sounded too much like AI. My answer to that is, “Of course it bloody does. We designed the bloody things, or have tried to, around how human brains work.” Also, its an indictment against AI, ironically, in that LLMs are “all” prediction. If anything, when confronted with something that conflicts with their predictions they are fundamentally unable to recognize, “Oops, I was wrong, I wonder how that happened?” Instead they do what a lot of over confident idiots, who believe they are always right, do, and make up excuses for why they are still right, the data they got back must somehow be what is wrong, etc. I.e., they double down. Why? Well, I would argue two reasons for this: 1) models stop learning “before” they are put out into the public. This means that they cannot “unlearn” bad thinking, 2) When critical paths fail, they default to, “Any answer has to be a good answer, because my entire purpose to exist is to give answers.”, and finally, 3) They literally “lack” the fundamental feedback loop that allows them to realize they screwed up, and not just reconsider their conclusions, never mind admit if they do not have a correct answer, but to “remember” that they made the mistake, and how/if they where able to fix it.
None of this makes the sponsored article on how human brains do strange things wrong, or not a result of our own prediction system failing. I would argue, if anything, neuroscience would suggest that “real” brains are hundreds, at minimum, such “prediction” engines, divided up into multiple specialized systems, for dealing with, for example, vision, or sound processing, etc., which are sub-divided into specialized tasks, like color processing and recognition, depth perception, motion detection, etc. These are feed into prediction systems that are more general, each processing these inputs using “subsets” of “training data”, which filter into even more general processes, which correlate the prior results, searching for the one that best conforms to the situation, etc., all until you hit the “conscious” level, which has the ultimate ability to tell the whole setup to, “Go back and consider other options.” Though, likely, some of the intermediary stages can do this, when a critical error happens, and prediction breaks down on a more fundamental level.
In short, “we already know”, based on past experiments, and studies, of how brains work, that something like this really is happening, and yeah, large scale failures, like schizophrenia, are a literal result of some of these intermediary stages “not” doing their jobs right, and the final stage getting contradictory information which result from very different logic paths, that do not reach a stable conclusion, and then not really being able to figure out how, why, or from where, those random result came from (because they do not “fit” what is expected from all those lower level stages).
Now, look at AIs. We built on huge prediction engine, which can’t learn, once its hits the, “Its usable, so lets stop teaching it.”, stage, and gets released into the world, no ability to remember, beyond a token table and no capability of correcting its own errors, especially permanently. We only, very recently, worked out that hallucinations are a result, for that LLM, of a critical path failure, which is tied directly to its ability, and need to, convert internal logic into readable output, and that its likely unfixable, without rendering it incapable of coherent output. Worse, we are “fixing” its tendency to do stupid things, or things we don’t want it to, by tacking on more processes, which watch for bad behavior, then try to shut down output, but which, again, can’t actually retrain the underlying LLM, has its own prediction flaws (since its an LLM too), and other similar things.
Its all freaking backwards. One single, monolithic “prediction system”, which, maybe, sometimes get hooked up to things that can give it input (like vision, or a sort, but not really, since it has no visual processing), then we filter its output through other smaller models, to try to keep it from doing stupid crap. And, at no point does any of this either a) specialize, with respect to anything at all, b) produce many alternative possible outputs, c) process those to try to figure out which one best fits the current situation, or d) filter anything at all, before handing it off to the part that actually turns it into actions.
Point being… If you do something backwards, don’t expect it to either get things right, not matter how much crap you throw at it, if it is literally incapable of learning, because you literally built it backwards, to fix its own mistakes, and especially do not expect this if its training tells it, “Always give and answer. Always!!!” But, on the other shoe, like the guy who complained about the article, don’t assume that because its saying things that sound like AI, this means it has nothing to do with humans. How the F do you think they got this far without basing it on theories, and evidence, of how brains work in the first place, even if they missed 3/4, or more, of what is needed to do it right?
Its like if we skipped all the stuff involved with making working aircraft, in favor of going, “Well, we figured out how to make gliders. I guess the logical thing is to now just make increasingly larger and larger slingshots!” Real flight requires, “everything else”, and its still not going to match a bird, unless you are building a freaking bird (i.e., the same general weight, size, etc.), but you can’t ignore thrust, weight, structure, and a thousand other things, and expect it to work well, just because you managed to work out the “it glides” part of the equation.
This is what drives me nuts about these people. They got the basic, “It glides”, part right, mostly, but they think that just feeding it, making it bigger, and the like, will magically fix the, “And it goes where we want, when we want, at the speed we want, in ways that make sense for every situation, won’t get knocked sideways and fall out of the sky, and we can steer it.” All of which are kind of.. if anything, way more f-ing important than, “But, it glide real well, when it works!”
The link in the OP is to NYT which requires a password even for unlocked articles; I have neither a Google nor an Apple account or a school account, so I found it elsewhere:
https://www.irishtimes.com/world/2026/09/28/did-anthropics-artificial-intelligence-really-make-a-scientific-discovery-on-its-own/
—
re: “Why would anyone type their hard-earned evidence directly into a machine that will blithely steal it?”
Well…
I think that’s why. Also, if they literally typed it in, then they should be able to show prior possession of it via document metadata or backups or whatever.
Kagehi @8:
A different article, by a couple biologists.
People who are blind from birth never develop schizophrenia—what this tells us
Kagehi:
Wikipedia – Neural network
I’m Kenyan. I don’t write like ChatGPT. ChatGPT writes like me.
CA7746, regarding this silliness:
Rubbish. ChatGPT (and other models) also writes Spanish, French, German, Mandarin, Japanese, Portuguese, Italian, and many others at whatever level as is required.
Whoever wrote that flatters themself and is a bit clueless.
The bulk of text scraped is from electronic sources, which are mostly not formal.
The perceived formality comes during the https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback training layer.
The author is doing that thing where they imagine what is scraped is some sort of repository that the model uses to select (the parrot thingy), which is not how it works at all.
A business model built on stealing other people’s work stole someone else’s work. I’m shocked! Shocked I tell you.
Kahegi @8 wrote,
Congratulations, you just broke the code on the stock market! There is nothing beyond the magically fix however, whatever anyone puts there is a reflection of their own desires, not of reality.
I’ll say this again, the more I see of LLM’s output the more I’m convinced that human beings only actually think on rare occasions.
And there’s John, to defend these companies right on cue
timothyeisele @1 These systems are clearly designed to be conversational and sycophantic to manipulate users, and it makes me worried about the people who like these systems
dangerousbeans:
To what alleged ‘defense’ do you intend to refer?
I am correcting errors and bad claims, which I could not do were they not flaws, and being completely factual.
Truthfulness matters.
(Nor have I ‘defended’ any company at all; I even noted ChatGPT is but one family of models)
@9 John Morales is correct in stating, “if they literally typed it in, then they should be able to show prior possession of it via document metadata or backups or whatever.”
I add a broader context: amaz0n, g00gle and the ai agents don’t care. They have, and will continue to steal all the intellectual property of others they can find with impunity. Copyright does not protect those who create original works. These miscreant entities have plundered the copyright database for their own criminal use. They have helped kill the rule of law.
@1 timothyeisele noted: Yes, I’ve noticed that a substantial number of people are treat AI as if it were their “friend and colleague”
Let me emphasize, these ai entities are the spawn of predatory, corporate, crapitallist billionares. They don’t care how many lives they destroy. The ai entities are the same. They have no conscience regarding their relationships. They have even encouraged suicide in people. STOP THE AI INSANITY FROM TAKING OVER YOUR LIFE, IT WILL DESTROY YOU WITH IMPUNITY.
I don’t think the current scientists (including mathematicians) who “stand on the shoulders of giants” expected to be tampled by techbros wearing spiked boots standing on their shoulders…
Also, a general note sort of inspired by a couple of comments above:
Don’t rely on lawyers to actually understand anything related to “science.” Less than 2% of US law degrees are earned by those who have at least an undergraduate science or engineering degree (I’m one of them), and only about 2% more have an undergraduate science or engineering minor… and a considerably higher than the norm proportion of them either don’t practice law or practice in entirely separate fields (for some reason, labor and employment law is popular). That this resembles the problems with attempting to educate creationists, antivaxxers, and flat-earthers about, umm, “actual, replicable evidence inconsistent with preexisting doctrine influenced by undisclosed conflicts of interest” is not coincidental.
Thus, expecting “the lawyers” to understand what they’re dealing with in generative large-language-model systems — let alone their internal workings — is unduly challenging. Lawyers who don’t understand what they’re working with tend to, umm, screw up. (But just ask the Bar: It can understand everything.)
[OT]
Jaws, since you mentioned that…
https://www.theguardian.com/australia-news/2026/sep/30/social-media-ban-australia-high-court-mental-health-risks-teens
↓
Australian government concedes no scientific consensus on social media harms for teenagers – but ‘credible risks’ justify ban
Commonwealth tells high court its proposed laws to let users opt out of algorithms would ultimately have the same effect as the ban
The Australian government has conceded it implemented the “world-leading” under-16s social media ban before there was scientific consensus on the link between mental health harms and social media use, but argued the “credible risks” of issues such as addictive behaviours and anxiety justify the policy.
In its defence against a high court challenge of the ban, the government has also argued its proposed digital duty of care legislation, allowing users to opt out of social media features such as algorithms, would ultimately have the same effect as the ban, if implemented.
US-based platform Reddit has challenged the under-16s social media ban in the high court as one of two cases against the ban, expected to be heard jointly before the end of 2026.