AI was trumpeted as the breakthrough that would transform our lives. The aim was to produce a machine able to analyse facts and reason for itself. It was believed that computers could use their speed and problem-solving ability to match human expertise, and that the knowledge of an expert could be recreated within a machine using a system of rules and facts. Unfortunately, theory then bumped into real life.
Now, in the Nineties, rather than trying to out-perform human beings in problem-solving skills, there are signs that AI applications are earning their keep by processing routine tasks, enabling people to concentrate on more demanding work.
Why did AI fail to live up to expectations? The researchers had underestimated how difficult it was to encode expert knowledge and everyday life into rules. Also, they failed to recognise how time-intensive and expensive it would be to keep such devices up to date, and people were reluctant to rely solely on a machine for important answers.
Although such expert systems are successful in specialised fields of engineering and finance, forecasts that we would all be using them in everyday life have never come true. Similarly, predictions made decades ago that computers would soon translate speech from one language to another in seconds failed to take into account the complexity of everyday talk.
Researchers this year finally produced a system that was able to recognise and translate speech from Japanese to German or English - but only if the speakers restrict their discussion to the booking of international conferences, as the machine's vocabulary is limited to that topic. And they must speak slowly.
If AI does come into our everyday lives it will be as an ever-present aide, rather than as a genius in a black box.
Typical of the new approach is work by Dr Herb Doller at the Veterans Administration Medical Center in Dallas, Texas, to produce a cost-effective assistant for doctors in specialist clinics. The aim is not to increase expertise but to cut costs and time.
Dr Doller claims the system reduces the cost of care by 40 per cent and that this can be applied to any clinic. He describes the potential effect on medical practice as 'explosive'.
The latest device from the hospital's AI laboratory is called Epileptologist Assistant, and is used by nurses in a follow-up clinic to question epileptic patients before their examination by a doctor.
'If you ask a physician or any other expert how they spend their time, they will say that 60 to 80 per cent is spent on things that could be considered relatively trivial,' Dr Doller says. 'About 20 per cent of time is spent on interesting and exciting cases and just 1 per cent on treatment that needs special research or consultation with a colleague. Rather than concentrate on that 1 per cent, like traditional expert systems, we set out to provide the best assistance to manage the 60 to 80 per cent of routine work.'
The Assistant prompts the nurse to ask between 30 and 50 questions from a repertoire of 300. Each line of questioning is dictated by more than 200 rules, and at the end it delivers a full report of patient responses as well as suggesting changes in treatment which the doctor can alter or adopt during consultation.
Dr Doller estimates two nurses and one doctor can now handle a caseload that previously needed three doctors. This lets the rest of the team spend more time on complicated cases and research. The system also allows nurses to take more responsibility for giving primary health care, and trains them in questioning epileptic patients; a skill that used to take six to nine months to acquire.
There is still enthusiasm for expert systems. Delegates at the Healthcare Computing conference in Harrogate last week heard how expert systems could help doctors to diagnose Hodgkin's disease and train staff. But Steven Hand, working on the Hodgkin's disease system, found it impossible to produce one universal version, as treatment varied from hospital to hospital, drug regimes changed regularly and new techniques made its suggestions out of date.
Business requirements work against complicated systems that need continual maintenance, too. Already banks and building societies are using low-key expert systems that can run on an office PC and can easily be set up to assist in mundane tasks such as assessing loan applications. These also enable junior staff to take on more responsibility and receive training in the process.
Finance houses are using neural networks, which excel at pattern- matching, to judge whether customers are good or bad credit risks by reading through application forms. The networks achieve this 'training' by looking at the forms of known bad credit customers, and then at good. The self-adjusting network analyses the forms time and time again until it has learnt to sift good from bad. It then uses this learning to assess new applications.
A team at Portsmouth University recently announced a prototype system, called Mycroft, that uses rules about criminal cases, derived from evidence submitted by Sussex detectives, to analyse clues and suggests how investigations should proceed.
Another form of AI allows a person accessing a database to ask for information by typing in plain English, and receive a similar answer, without realising that the English words have been turned into computer commands behind the scenes to speed the search. Some large databases can already be used in this way.
What cannot be foreseen is whether such innovations will become a part of everyday life, like the fax machine, or remain a technological oddity that never spreads far from the laboratory or finance house.
A rash of personal digital assistants, billed as complete personal organisers, are due soon from manufacturers such as Apple, using AI techniques to recognise the user's handwriting. The idea is that you can scribble 'Lunch with Bob on Tuesday' and it will automatically make the diary entry, perhaps pausing only to query which Bob. If such assistants catch on, AI will finally have hit the streets.
Perhaps in a couple of years the police constable flagging you down for speeding will be prompted in his note-taking by Beat Assistant, running on his Personal Digital Assistant.
Or more likely, he'll still be using a notebook and slightly chewed Biro.