Artificial intelligence is worse than humans in every way at summarising documents and might actually create additional work for people, a government trial of the technology has found.
Amazon conducted the test earlier this year for Australia’s corporate regulator the Securities and Investments Commission (ASIC) using submissions made to an inquiry. The outcome of the trial was revealed in an answer to a questions on notice at the Senate select committee on adopting artificial intelligence.
The test involved testing generative AI models before selecting one to ingest five submissions from a parliamentary inquiry into audit and consultancy firms. The most promising model, Meta’s open source model Llama2-70B, was prompted to summarise the submissions with a focus on ASIC mentions, recommendations, references to more regulation, and to include the page references and context.
Ten ASIC staff, of varying levels of seniority, were also given the same task with similar prompts. Then, a group of reviewers blindly assessed the summaries produced by both humans and AI for coherency, length, ASIC references, regulation references and for identifying recommendations. They were unaware that this exercise involved AI at all.
These reviewers overwhelmingly found that the human summaries beat out their AI competitors on every criteria and on every submission, scoring an 81% on an internal rubric compared with the machine’s 47%.
Are we talking 10% worse and 95% cheaper? Or 50% worse and 10% cheaper? Or 90% worse and 95% cheaper?
Because that last one is good enough for fiscal conservatives. Hell, the second one is good enough for fiscal conservatives.
The linked pdf lists the deficiencies of the LLM responses. They are varied and it sometimes misses the mark completely or cant grasp vital context.
Still pretty useless comparison, they testet 10 university level humans against Llama2-70B. The model has fallen out of use completely by now and was never really great at summarization. The study didnt fine tune it either, so this isnt really representative of the current situation.
There are far better models out, that were either especially trained for summarization or can be easily fine tuned to excel at it. Not to mention the Llama3 and 3.1 series, with the crazy 405B model.
Knowing this it seems like a very low quality study. They should probably redo this with multiple conditions.
I’ve heard this refrain a few times. Still waiting for it to pan out.