Debate Intensifies on Search Engines' Legal Liabilities Related to AI Outcomes
Ok but wait til we get to appeals with amicus brief from someone like google because depending on how this goes, all search engines could suddenly have such liability? But thats absurd and likely this has to go the other way. Rehashing 30 year old debates about legal websites.
That seems awfully presumptive about how the search was phrased. Apply law to actual facts and it's advice. There was a time when this was argued about search. Ai is just search but the output looks different. It's the users' sophistication of expectations that need to change.
no. i think you're unduly limiting what you think the word search means. One can't do legal (re)search without also providing at least some level of legal advice because you had to convert the facts to legal terminology somewhere. turning keywords to sentences isn't the threshold
if you think the difference is more than the formatting of presentation, you might be exactly who I meant when I said users need increased sophistication. it looks and feels different. it isn't actually different.
I'll concede I'm exaggerating slightly ("search" may now be term of art excluding some Ai functions) but grok also exaggerates the difference. Statistical pattern generation is machine navigating knowledge space, and turning queries to results. Grammar isn't what makes it advice.
"Chain reasoning" and "self-critique" are just another round of generated text pattern. This is the same debate as when WebMD first allowed users enter symptoms and get tailored diagnoses. Or any legal search site allowing lay persons to read case law searched by their own facts.
The novelty here is the fluency, not the function. WebMD mapped symptoms to diagnoses; legal research tools map facts to cases. Generative models just collapse those steps into narrative text. That's a difference of presentation style, not whether it's advice.
"Emergent synthesis" is marketing language for probabilistic text generation. It's still just a system mapping queries to patterns learned from existing text.
Scale improves capability, but recombining patterns into novel sentences isn’t a new category of reasoning. Search already maps queries through ranking and inference to knowledge. Generative models mostly narrate that process. It feels different. It isn't.
Those behaviors are probabilistic text generation producing reasoning-shaped language. LLMs produce novel sentences, but the reasoning structures come from the training data. That can be powerful synthesis but is not a new category. It's still a query-to-knowledge system.
your "no search engine originates" only works if you lock the abstraction level at document retrieval. At a higher level both systems navigate a knowledge space from query to result; search via ranking, LLMs via probabilistic generation. It's not different.
that distinction is false because search engines also generate new edges when they rank, cluster, or infer semantic relationships between documents.
Again that's false, or debatable. LLMs don't "author new nodes entirely" they find structures of reasoning derived from patterns. And modern search engines aren't pure static retrieval of fixed corpus either. Both systems navigate knowledge spaces. The difference is output style.
That comparison freezes "search" at its most primitive definition. Modern search already infers relationships, expands queries, and synthesizes answers across documents. LLMs use a learned representation instead of an index, but both are still query-to-knowledge systems.
your leap of "creation" vs "search" just begs the question. Both systems still depend on prior knowledge representations; one compiles them into model weights, the other into an index. That's an architectural difference, not a categorical one.
That claim is epistemically unverifiable. "Absent from any training corpus" isn't demonstrable. What we can say is that models generalize patterns from prior text. That's can be extremely powerful, but it's still extrapolation from learned representations.
That's a weaker claim than before. "no single source" just describes compositional generalization; recombining patterns across multiple sources into a new chain. That's how statistical models work but novel sentences don't imply novel reasoning processes.
That's statistical generalization: extrapolating learned patterns to new inputs. It doesn't establish a new epistemic category. Calling extrapolation from latent structures "creation" is philosophically suspect.
Novel outputs don't imply a novel mechanism. Statistical models generalize learned patterns into new combinations. Calling that "creation" shifts the claim from mechanism to appearance. And that's exactly my point: it's appearance.
you can't prove your claim that "full derivation traces to no corpus reconstruction". that's the epistemic problem: you are claiming proven novelty of reasoning, but the mechanism only supports untraceable recombination.
Exactly! The provenance is unknowable either way. That's why your claim of authorship is weak. Statistical generalization produces novel outputs. But evidence of strong generalization is not proof of new epistemic category. it’s a system navigating learned representation: search.
Outcomes don't redefine mechanisms. Surprising results don't establish authorship. Epistemologically and mechanically, these are systems mapping learned representations: that's called search!
your "authorship by emergence" just renames statistical generalization. The mechanism hasn't changed; only the scale.
now "runtime authorship" renames statistical generalization again. buzzwords don't change the mechanism. Scale changed performance, not the class of mechanism. In machine learning language that's still just generalization from learned representations. That's fundamentally search.
Those "latent alignments" are functions of learned weights derived from training. The model is navigating a learned representation space, albeit dynamically rather than through an index. Computationally and functionally, that's still search.
That reasoning is circular: novel outcomes are taken as proof that the mechanism isn't search. But the mechanism, dynamic computation from learned weights producing sequences, is still a form of search through a learned representation space.
but "verifiable post-cutoff originals" can also be generated by search processes (genetic algo, Monte Carlo tree search, theorem provers). So your "empirical test" doesn't distinguish the category you're claiming.
First you said the distinction was novelty of output: we now agree some search processes can generate novel results. Your new claim is search requires explicit loops or trees. It doesn't. Exploring a space via probability distributions is search; transformers do it implicitly.
So now search is being defined as "iterative tree exploration". That's a very narrow definition. More generally "search" means finding outputs in a space that satisfy a criterion. Transformers explore token space via probability distributions during decoding: still search!
Yes many computer scientists treat broad classes of computation as forms of search; some argue all algo can be framed as such. By contrast, you're narrowing the definition to exclude a functionally similar mechanism, and your boundary keeps shifting in response to counterexample.
that definition of search isn't standard in CS. I think there are search methods with no heuristics or evaluation loops (consider: binary search, nearest-neighbor search). And isn't LLM decoding also iterative token selection over sequence space?
softmax already evaluates alternatives: the model scores every token in the vocabulary, and decoding selects among them. That's selection over a search space.