- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
A new study from Columbia Journalism Review showed that AI search engines and chatbots, such as OpenAI’s ChatGPT Search, Perplexity, Deepseek Search, Microsoft Copilot, Grok and Google’s Gemini, are just wrong, way too often.
They get you 80 to 90 percent close to generally solving most problems asked. Sure they need fact checked as any info does. They are of major use in all areas of study and life. Just not the god everyone wants it to be.
They are less useful than a Wikipedia search and a dictionary. They can functionally replace humans in 0 fields that were not already automatable by machines. They are useless in any situation that warrants any degree of caution about safety.
85-90% is way over-estimated, it gets significantly worse dealing with specific tasks. And even if it was 85-90%, that’s not good enough, even remotely, for just about anything. Humans make errors too, but inconsistently and inversely proportional to experience. This makes no difference to the LLM though, it will always make errors at that exact rate. The kinds of errors it can make are also not just missteps but often pure delusion and very far from what the input was requesting. They cannot reason. They have no rationale. They’re imitation in its most empty form. They cannot even so much as provide information reliably.
They also ruin every single industry they come into contact with, and even worse they have utterly destroyed the usability of the internet. LLMs are a net negative for humanity in so many different ways. They deserve as much attention and investment as chatbots did back in 2005.
Their best use case scenario is in churning out an endless amount of lifeless soleless jpg background noise and word salad articles. Their best use case is in tricking people into giving them money or ad revenue. Scamming is the only thing they are anywhere near functionally useful for.
The problem when you keep moving the bar like that is that you will find yourself supporting stupid gimmicks instead of practical technology.
The problem is that the sales pitch of these AI answering services in search engines is to save you time from having to open search results and read them yourself. The problem with 80-90% accuracy is that if the summaries are hallucinated even once, you can no longer trust them implicitly, so in all cases you now have to verify what it says by opening search results and reading them yourself. It’s a convenience feature that doesn’t offer you any actual convenience.
Sure it’s impressive that they are accurate 80-90% of their time, but AI used in this context is of no actual value.
It’s a real issue. A strong use case for LLM search engines is providing summaries which combine lots of facts that would take some time to compile through searching the old fashioned way. But if it’s only 90% accurate and 10% hallucinated bullshit, it becomes very difficult to pick out the bullshit from the truth.
The other day I asked Copilot to provide an overview of a particular industrial sector in my area. It produced something that was 90% concise, accurate, readable and informative briefing, and 10% complete nonsense. It hallucinated an industrial estate that didn’t exist, a whole government programme that doesn’t exist, it talked about a scheme that went defunct 20 years ago as if it were still current, etc. If it weren’t for the fact that I was already very familiar with the subject, I might not have caught it. Anyone actually relying on that for useful work is in serious danger of making a complete tit of themselves.
Copilot sucks and I totally understand the POV. I stick with GPT, Mixtral. I don’t think their going anywhere anytime soon but they need significant actual refinement.
@EncryptKeeper @OhVenus_Baby I have very much embraced the swearing method to get rid of 50% of my Google result screen being taken up by an untrustworthy statement. Just a waste of space and scrolling time.
yeah my thought was like how often is a web search “right”. To me ai search is just another version of search. its like at first searching gave you a list of urls, then it gave you a list of urls with human type names as well to make it more clear what it was, then it started giving back little summaries to give an idea of what each page was saying. now it gives you a summary of many pages. My main complaint is these things should be required to giver references and their answers should pretty much look like a wikipedia page but the little drop down carats or roll overs (although I prefer a drop down myself).
It never was, but unlike the current batch of LLM assistants that are now dominating the tops of “search” results, it never claimed to be. It was more, “here’s what triggered our algorithm as “relevant.” Figure out your life, human.”
Now, instead, you have a paragraph of natural text that will literally tell you all about cities that don’t exist and confidently assert that bestiality is celebrated in Washington DC because someone wrote popular werewolf slash fanfic set in Washington state. Teach the LLMs some fucking equivocation and this problem is immediately reduced, but then it makes it obvious that these things aren’t Majel Barrett in Star Trek and they’ve been pushed out much too quickly.
What I have seen like duckAI in duckduckgo search it cites references and if you query GPT and other models with the specific data your looking towards it will cite and give links to where it sourced the input from like PubMed. Etc.
For instance I will query with something like give me a list of flowers that are purple, cite all sources and ensure accuracy of data provided by cross referencing with other studies while using previous chats as context.
I find it’s about how you type your queries and logic. Once you understand how the models work rather than blindly accepting them as supreme AI then you understand it’s limits and how to utilize the tool for what they are.
I really feel it should not be necessary to ask them to site all sources though. It should be default behavior.