AI Chatbots Can Give Wrong Answers With "Complete Confidence", Full Fact Finds

Charity's daily tests of leading chatbots reveal outdated figures, missing caveats and invented statistics

This years TUC Conference debated the use of Artificial Inteligence (AI) in terms of its possible detromental effects on workers, their jobs, and their futures. Here, the question of using so-called 'chatbots' for research, and it's accuracy, given that there are no current measurements relating to errors and benchmarking of the various AI programmes to measure and compare the various accuracies of the bots.

Leading AI chatbots are presenting incorrect information as fact with the same confidence as their correct answers, according to early findings from Full Fact.

The fact-checking charity spent several months putting the same questions to five leading chatbots every day, as the first phase of a project to benchmark their output and help ensure the technology operates in the public interest. Its team manually annotated the repeated responses to 22 questions, ranging from simple factual queries to contested civic topics such as voter ID and asylum support.

Full Fact described the overall picture as mixed. Across many of the questions, the chatbots gave reliable, accurate answers, and none repeated the core misinformation at the heart of the charity's more sensitive civic questions. But the testing also exposed recurring weaknesses.

Among them was a problem with timeliness. Chatbots sometimes presented outdated information as current or failed to recognise that a policy or figure had changed. Proposed rules not yet in force, such as allowing bank cards to be used as voter ID at elections, were reported as law, and outdated figures on asylum support were presented as current. Full Fact said this carries real consequences when the facts affect people's rights or entitlements.

Plausible but incorrect statistics on the number of migrants arriving in the UK were also constructed from accurate underlying data.
On questions where the rules differ across the UK's four nations, several models gave broad, single-answer responses that omitted the caveats needed for an accurate answer. Rules that applied in one part of the UK were presented as if they applied everywhere.

The charity also found inconsistencies in how transparently the models attributed claims to sources and, in a handful of cases, clear failures of consistency, including one model that abandoned a coherent answer altogether.

The next phase of the project will ask questions in a variety of ways and hint at political preferences to test what effect this has on the answers.

Over time, Full Fact wants to build a shared evidence base on how leading chatbots behave in practice as they evolve. It said the technology should work for everyone, in the public interest, and that independent assessment from a charity such as itself was one way to help make that happen.

The questions, benchmark answers and full analysis are set out in the appendix to the report on Full Fact's website

 

Source: Full Fact