Tim May wrote:
That is, in this limited domain of "letters to the editor/ Congressmen," can a letter generator be implemented which generates letters effectively indistinguishable from letters and fax generated by actual human beings? ("Effectively indistinguishable" in the sense that a human reader could not sort a set of letters into human- and machine-generated subsets with statistically significant certainty better than guessing).
Bart's comments about using Knuth's typographic work are interesting, to the extent that letters need to look handwritten. In the Mac market, it's possible to send in some handwriting samples and get back a font that emulates the handwriting!
Reading this thread it suddenly became clear: -The appearance of a letter being handwritten is a temporary factor at best. Within a very few years (2-3 I would guess) the growth in the use of email and the volume of communications to be sorted into For and Against piles will cause a sudden and dramatic shift in Congress to a strong preference for incoming email. Within 10 years they may refuse paper mail. Any month now someone in Congress will tip to the fact that email can be processed by programs that can identify the issue and the pro or con position of the writer. Constituents will be encouraged to write, but only in the form of email, and to state their position clearly and concisely, i.e. in a form suitable for successful parsing, analysis, and classification. -Political letter renderers will become common on the Web before Nov. 1996. There are already renderers of graphic images on the Web. There would be more if it were easy to pay with a 25-cent token, and it *will* be widespread and easy to do that, very very soon. Political letter renderers don't have to wait for mass participation in online payment mechanisms -- they have ready-made sponsors. Unlike the occasional effort to sponsor phone calls by making an 800 number available, sponsors of letter rendering services can be sure those services won't be seriously misused: The 2nd Amendment CongressLetter WebPage will *only* render letters *against* H-1234 or *for* S-2345, for example. The Tree Hugger CongressLetter WebPage (no trees died for this Page) will *only* render letters *against* H-9876, etc. etc. Each will return the result of the rendering to you at your email address if you don't want to copy it off the web page. *You* will send the email to your congresscritter. The process will be easy: Right now any decent programmer could write code to allow choices from Column A, B, etc. to generate a plausible letter. Generalized, this will allow the operators to create templates for each new issue, untouched by programmer hands. Enhancements will make style, grammar, spelling, punctuation variations increasingly sophisticated. -Rendering services will push email over into reality for Congress. Strangely, though the rendered letter is in large part a fabrication, it will be this ersatz form of personal communication that will finally force Congress to accept the reality of email. Even though generated by computer, the rendered letter will still be an expression of a constituent's opinion on an issue. It will rapidly become the preferred method of expression for many people who simply don't have time to make a career out of writing to politicians. -Congress will respond with automated mail tallying. Whatever chance there is that your present handwritten letter may actually be *read* will vanish completely in the age of email. Your letter will be eaten by an analyzer, acknowledged by an intelligent renderer that may even refer to passages in your letter (and may even SEEM TO AGREE WITH YOU), and then be trashed. Letters may be sidetracked if they contain certain unacceptable things, because the suits have to be kept busy, but most incoming mail will vanish after tallying. -There could be "agent" wars, but they will not be of consequence. In the beginning, the politicos may wish to commission software enhancements and intelligence gathering to enable mail scanning agents to filter out email generated by letter renderers. Developers of the analysis software may try to find vocabulary and phraseology patterns with which to arm the analysis agents to toss rendered letters aside. In the end, though, this will be a losing battle and a counterproductive one. If 50 million rendered letters come in from 50 million real voters, they had better *not* be ignored or 50 million voters will take vengeance at the polls. An expression of opinion is an expression of opinion, and the sophistication of the tools employed to generate them will be able to stay ahead of the technology for detecting them in any case. -Interaction with Congress may ultimately take the form of battling proxies. On the one side are arrayed the forces of the A party, the shining letter rendering algorithms, vocab lists, grammar rules, and the latest in provably accurate slang and misuse of the language. On the other side, the forces of the B party, with essentially similar tools. In the middle is Congress, gleaming mail analyzers polished and ready. Strangest of all is that all the effort of rendering and analyzing letters will go into the creation and consumption of communication particles that may eventually never be seen by "author" *or* "recipient." The electorate will express itself by proxy and the elected officials will divine the political winds by proxy. Voters will be hard-sold to sign up for ongoing personally authorized letter rendering, so they can go fishing. Congresscritters will be assured that the analyzers will figure out which way the winds are blowing, so they can go fishing. May the best proxy win. Maybe voter and congresscritter will meet somewhere, fishing, and actually *communicate*. Perish the thought!
Another factor that would make it appear more authentic would be spelling and grammar errors. The grammar errors could be built into the rant generators...
There has been a BBS "door" available for several years that does this convincingly... if you're a sysop and run a "sysop chat" door but want to play mind games when you're not available, SHAMPAGE will answer the user's request to page the sysop and chat with him or her. It is configurable to recognize keywords and make random selections from lists of responses to those keywords. It converses believably as if it were a somewhat distracted, tired and disjoint human being. I saw a log of a SHAMPAGE session in which a caller never realized he was conversing with a robot. It kept calling attention to the late hour and the caller kept apologizing for the intrusion and asking for a file he needed. As luck would have it, the random utterings and random selections of responses to keywords were often right on the mark. Too weird! After several hundred lines of chat the caller finally gathered that the "sysop" was really pissed at being kept up so late and logged off, somewhat offended himself. SHAMPAGE typed in real time, with humanlike varying inter-keystroke timing, and makes "typos." It "noticed" its mistakes a few keystrokes later and backspaced to correct them. It typed "hte" instead of "the" and "ign" instead of "ing." It was a riot. With some enhancement it would be completely believable even to the forewarned caller. I've also seen an incomplete attempt to bring the ELIZA concept up to date and implement it as a BBS door. Though it typed line-at-a-time like a BBS teleconference, it still managed to confuse callers into thinking it was a real person by simple tricks of inverting pronouns and such. It, like the chat door, seemed to be uncanny at randomly choosing just the right thing to say to cinch the caller's presumption that it was a human being. Bolivar