I have recently seen several risks contributions which assumed that humans are the cause of most system accidents and that if the human was somehow replaced by a computer and not allowed to override the computer (i.e. to mess things up), everything would be fine. The issue is too complicated to cover adequately here. But before rushing off to replace human controllers with computers, at the least consider the following: ** Most accidents involve multiple failures of different components of the system. It is rarely possible to pinpoint one particular failure as the sole cause. (e.g. Three Mile Island involved at least four or five different types of mechanical failures. Who got the blame?) ** There are often powerful and compelling reasons for wanting the blame placed on the human. For example, Babcock and Wilcox can be sued for billions if there is something wrong with the design of their nuclear power plants — how much can you collect from some poor operator? ** The human is often called in to save the day after chaos has already begun and then expected to come up with a miracle. If he does not save the day, then the blame is often placed on him/her instead of the initiating mechanical failures. ** Most accidents result from unanticipated events and conditions. Thus it is doubtful that computers will be able to cope with emergencies as well as humans do. Expert systems do not help in coping with unanticipated events or conditions. ** There are many examples of accidents which were averted by a human overruling an errant computer. If the operator had not intervened at the Crystal River Nuclear Power Plant, for example, a catastrophe might have occurred because of the computer error. The hype about "expert systems" and "artificial intelligence" may be very dangerous. There are reports that commercial pilots are becoming so complacent about automatic flight control systems that they are averse to intervene when failures do occur and are not reacting fast enough (because of the assumption that the computer must be right). The problem is just not that simple that the answer "replace the human with a computer" will solve it. Nancy
As an owner of Eastern Airlines stock (fell from $11 to $5 right after I bought it), I'm particularly upset by this. I don't know the details; I hope someone with more knowledge can fill them in. According to my stock broker (Disclaimer: I don't have any hard documentation, and I'm not a Wall. St. expert), one of the major blows to the already troubled company was a bogus earnings report issued on a Dow Jones computer (something like 20 cents instead of $1.50). The mistake was corrected within the hour, but in that hour, portfolio managers had dumped Eastern stock, and the price fell $3, and never recovered. I think this happened around early September.
There seems to be some misunderstanding about computerized stock trading. First, "programmed buys" and "programmed sells" really have nothing to do with computers. All "programmed" transactions could be done by hand but typically they are extremely complex, so a computer is needed. Programmed trading only occurs when special intermarket conditions are present. Program trading consists of arbitrageurs who use the spread between the value of stocks on the New York Stock Exchange (NYSE) and the Chicago Board of Options Exchange (CBOE) in Chicago. Occasionally other markets are used. Intermarket arbitrage adds to market volatility but not in a negative sense. The infamous "Triple Witching Hour", a time four times a year of extremely volatile trading, is a direct result of this intermarket arbitrage. Eric Nickell in his note compare the market to a feedback system that oscillates - something like a forcing function with resonance. Well not at all true. The market cannot get really swamped because something will "break-down first". In the case of the NYSE - the "market makers" will have an "order imbalance" preventing further trading.
In RISKS Vol 2, Issue 16, Kurt Hyde write: > There are many documented cases of accidental miscalculation in computerized > vote tallying equipment. The reasons why such errors were discovered is > because reconstruction and recount was possible. Investigators > reconstructed by gathering the machine-readable ballots. They were then > able to recount by machine or by hand. Such reconstruction is impossible > with the current state of the art in computerized voting booths because no > physical ballots are created. Recounts in such cases are wholly dependent > upon the software to have stored each vote in its proper storage location at > the time of voting. While the risks would not be entirely removed, and regardless if any fraud or error is suspected, there could be a standard practice initiated whereby a sample from each election is validated by follow-up phone call or physical notification. Privacy could be somewhat maintained by automating this process, e.g., immediately after the polls close, the computer randomly selects some small sample and sends a letter saying, "Citizens Jones, according to our computer voting system, you voted thusly:...." The citizen then returns the card validating or invalidating his voting record. A box could be checked for him to indicate that he would rather not acknowledge via mail or not at all; the percentage of such respondents would probably be low. Also, since some people may goof or maliciously be inconsistent, the final validation would not have to be unanimous; some standard percentage of validation would pass as I believe it does today in a recount. If delegating the follow-up procedure to a computer is the start of a new computer risk, then it could be done manually, but I believe this kind of check-back mechanism would significantly reduce the risks involved in computer voting to the point that it could gain approval. Larry Polnicky, Honeywell Information Systems, McLean, Virginia.
Kurt Hyde's reference to the Phillipine elections and the security of computerized vote-counting systems reminds me that the issue of computer security is artificially narrow. If I am a criminal, and you confront me with an unbreakable computer security system, I will simply direct my attention elsewhere. Attacking strong points went out with World War I (or, to maintain the underworld analogy, with Machine Gun Kelly). The most elaborately password-protected system is easily cracked if the passwords are transmitted over telephone lines, or if people leave their passwords lying about on scraps of paper. That may fall outside the venue of computer science, but not outside the venue of reality. In the case of the Phillipine elections, it didn't matter how well the vote-counting computers were programmed; there were soldiers at the polling places threatening to shoot voters. Ballot boxes were opened to reveal twenty thousand ballots marked in the same handwriting for Mr. Marcos. The computer operators were being told what numbers to enter. I guess there's not much you can do about risks outside your direct control. My point is not to get too focussed in our concerns. [As noted many times in RISKS, any single weak link may represent a vulnerability. In systems designed not to have single weak links, there are weak combinations. Thus we must be concerned with ALL of the weak links. PGN]
Today I attended an IEEE videoconference on "Applications of Artificial Intelligence" with Drs. Tom Mitchell (CMU/Rutgers), Alex Pentland (SRI), Peter Szolovits (MIT) and Harry Tennant (Texas Instruments). Aside from some overdriven graphics such that it interfered with the audio, it was an excellent intro to AI (for those concerned with the medium AND and the message). I asked the question, asked here and elsewhere by others, about the potential legal responsibility of authors of AI software, the most obvious example being medical diagnosis. The answer from the panel was that most AI work now has been done under very controlled conditions, responsibility has never been tested in a court case, and that (possibly) the law applying to publishers of reference books might apply also to AI systems (that is, willful deceit would be punishable but typos and other innocent mistakes would not make a publisher accountable). But according to one of the panel members some AI researchers ARE in fact taking out insurance against possible suits but (paraphrase) "the insurance companies look upon this problem as something of a lark and the insurance emiums are low now" although the same panel member said that (paraphrase) "this may become a very important problem in the future". I originally phrased the question to ask whether the implicit threat of possible suits against artificial intelligence applications might have a chilling effect on research and development of interesting applications (that is, those involving human life and property), but as it was not asked it was not answered. My own (legally uninformed) feeling is that AI by its very nature spreads around the concept of "volition" such that the present legal system might have a difficult task in assigning responsibility in a damage suit (and these doubtless will come down the pike someday).
There's an interesting article in the 2/24 issue of InformationWEEK concerning the DES. Apparently, DES was up for voting to become an international encryption standard sanctified by ISO. The NSA (National Security Agency) was lobbying very strongly within ANSI (the United States' representative within ISO) to have DES disapproved... the apparent reason being that wide standardization of DES, and its routine use, would make it substantially more difficult for NSA to monitor overseas voice and data communications. IBM pushed very strongly within ANSI for a "yes" vote within ISO (DES is already an ANSI standard, and its details have been readily available to anyone for the past five years or more). In the end, IBM won and NSA lost; ANSI abstained from voting, which had the same net effect as a "yes" vote. Have any studies been done concerning the risks of having, or not having a secure data-encryption scheme to guard the integrity of one's data?
Please report problems with the web pages to the maintainer