In the 1980s, when I started looking at algorithms and stuff, one of the tests that I set for whether so-called artificial intelligence works was speech-to-text, also known as speech recognition. IBM lent me a very nice, very fast PC and I had a couple of the leading packages to test. The results were dismal and editing the output took far longer than typing it initially.

The drive to make this work began when doctors in the USA found that it was cheaper and faster to dictate their notes into a computerised recording system rather than an ordinary Dictaphone and send sound files to India where they were transcribed and text files were returned in time for duty the next morning. This was before the internet was widespread and the files were transmitted by ISDN lines which many countries have recently "phased out" having been overtaken by internet connections which are cheaper and faster.

The then nascent voice to text industry tried to do the same job but despite heavy investment and many years, never succeeded in achieving the necessary level of accuracy. It's important: when you are dashing off a quick love letter, it's ok for there to be mistakes. But in medicine or engineering, it's literally a matter or life or death.

OpenAI has, within its suite of programs, one called "Whisper" which, it turns out, doesn't like silence and so it just fills it in, sometimes with content that can have serious, adverse, consequences.

In the summary to the paper, it says "In the summary it says "roughly 1% of audio transcriptions contained entire hallucinated phrases or sentences which did not exist in any form in the underlying audio… 38% of hallucinations include explicit harms such as perpetuating violence, making up inaccurate associations, or implying false authority. "

The specific instance that the researchers worked on is used in transcription of medical notes and the researchers make the point that these transcriptions might be used as evidence in trials or as the basis for future clinical or psychiatric assessments.

So what about the use of such tech in financial crime risk and compliance? I have long warned that computer generated reports - false positives or false negatives - can have serious repercussions for both the customer and for the bank (etc.) and reports should be carefully checked . We know that "generative AI" is touted as a report-writing tool and that internal and ultimately suspicious activity reports might be based on it.

How much effort should busy risk and compliance officers be expected to put into checking that what is in computer generated report matches dictation, for example?

In the reported case, some 30,000 medical personnel use it. The company selling it says it's aware of the problem and is addressing it. Meanwhile, it remains in use.

Do you know exactly what your regtech is producing, on a case by case basis? I very much doubt it.