April 08, 2006
Some basics from the economics of pollution
I have taken to characterizing "keeping the bad stuff out" of user-contributed content resources as a pollution problem. What can we learn from the long-standing economics literature on pollution (traditionally about environmental pollution)?
For this entry, I'm going to quote some passages from an article I published on "Economic Incentives" in the Encyclopedia of the Environment (Marshall Cavendish, Tarrytown, NY, 2000):
When costly side effects can be ignored by a polluter, there will be too much pollution relative to a Pareto optimum. Several policies can give polluters an economic incentive to consider side effects when deciding how much pollution to generate. The three most important economic incentives are taxes, subsidies, and tradable permits. These all work by internalizing the externality, that is by making the polluter directly face the cost created by the pollution.
Taxes, subsidies, tradable permits: this is a pretty limited palette, though it does comprise most of what the environmental economics literature has comprised. Over time, I'll be suggesting other incentive-based approaches to pollution in user-contributed content resources, but for now will stick to these three:
Taxes. Since 1920 economists have recommended imposing a tax on polluting activity equal to the incremental social cost imposed by that activity (A. C. Pigou, 1920). Then as long as the social cost of the pollution is greater than cost of prevention or clean-up, the polluter will want to reduce the pollution, to the point at which the social benefits of further reductions are not sufficient to warrant the costs of obtaining the reductions. A Pareto optimum can be achieved.
Subsidies. Rather than impose a tax to discourage pollution the government can offer an equal subsidy per unit of reduction. A subsidy per pound of gunk reduced creates the identical incentive as a tax per pound produced: each pound of gunk eliminated raises profits by the amount of the subsidy or tax. The main difference between taxes and subsidies is distributional: the cost of control can be paid by taxpayers (through a subsidy), or by some combination of the factory's owner, workers and customers (through a tax).
Tradable permits. A very different approach to using economic incentives for environmental problems is to create a market in which the polluter must pay a price for the use of the formerly unpriced input (e.g., clean air). The usual method is for the government to issue permits for a fixed amount of gunk and to allow individuals and firms to buy and sell the permits. The fewer the permits the higher will be their market price. By controlling the quantity of permits the government can control the permit price so that the polluter has to pay the same amount per pound of gunk as it would under a tax or subsidy. Thus, all three methods can solve equivalently the problem of equating the costs and benefits of externalities, while using the polluter's self-interest to obtain the socially desirable level of control.
Another well-known point from the pollution economics literature, this one quite relevant for the prevalent approach of trying to design technical fixes to prevent information pollution:
One advantage of using these economic incentives is that a given level of pollution control can be attained at the least cost. For example, with tradable permits, the permits will be most valuable to polluters with the highest control costs; they will purchase the permits, while those with lower control costs sell their permits and reduce their pollution. This result contrasts with the use of emission standards: all polluters must to control to a given level, even though it will likely be cheaper to have some polluters control a bit more while others control an equal amount less.
In the article I also discussed several complications that change the way in which the above mechanisms work. These include uncertainties about the benefits or costs of control; market imperfections in the markets in which polluters operate; non-convexities caused by severe or irreversible harms. These all have analogues for information pollution as well. Of course, the discussion above is only in terms of the effectiveness of these methods for maximizing social efficiency; if the objective function puts weight on other factors as well, these approaches may work less well.
Polluting user-contributed reviews
A recent First Monday article by David and Pinch (2006) documents an interesting case of book review pollution on Amazon. A user review of one book critically compared it to another. Immediately following a "user" entered another review blatantly plagiarizing a favorable review of the first book, and further user reviews did additional plagiarizing.
When the author of the first book discovered the plagiarism, he notified Amazon which at the time had a completely hands-off policy on user reviews, so it refused to intervene even for blatant plagiarism. (The policy since has changed.) Another example of the problem of keeping bad quality contributions out.
David and Pinch remind us that when an Amazon Canada programming glitch revealed reviewer identities,
a large number of authors had "gotten glowing testimonials from friends, husbands, wives, colleagues or paid professionals." A few had even 'reviewed' their own books, and, unsurprisingly, some had unfairly slurred the competition.
David and Pinch address the issue of review pollution at some length. First, the catalogue six discrete layers of reputation in the Amazon system, including user ratings of reviews by others, and a mechanism to report abuse. Then they conducted an analysis of 50,000 reviews of 10,000 books and CDs. Categories of review pollution they identified automatically (using software algorithms):
- Reviews copied from one item to another in order to promote the sales of a specific item.
- Reviews posted by the same author on multiple items (or multiple editions of the same item) trying to promote a specific product, agenda, or opinion.
- Reviews posted by the same author using multiple reviewer identities to bolster support for an item.
- Reviews (or parts thereof) posted by the same reviewer to increase their own credibility and/or to build their identity.
They also make an interesting point about the arms-race limitations of technical pollution screens:
The sorts of practices we have documented in this paper could have been documented by Amazon.com themselves (and for all we know may have indeed been documented). Furthermore if we can write an algorithm to detect copying then it is possible for Amazon.com to go further and use such algorithms to alert users to copying and if necessary remove material. If Amazon.com were to write such an algorithm and, say, remove copied material, this will not be the end of the story. Users will adapt to the new feature and will no doubt try and find new ways to game the system.
Spam economics: Private stamps vs. repudiable bond payments to recipients
After years during which everyone talked about economic incentives to better sender and receiver interests in unsolicited email, we may finally be seeing the dawn of the incentive-centered design era for email.
AOL and Yahoo! this winter announced they were adopting the Goodmail system to create a special class of incoming mail: senders that paid the Goodmail fee per message would have their mail placed directly in user inboxes, with no server-side filtering or blocking by the ESP (email service provider, AOL and Yahoo! in this case). Mail without the Goodmail stamp will receive traditional treatment, being filtered and possibly placed in the user's spam folder.
A rather loud debate immediately followed, focused primarily on one concern: AOL and Yahoo! would tighten the filtering screws on unstamped email, eventually shoving so much of it into the spam folder that everyone would be "forced" to pay for the Goodmail stamp or likely have their mail discarded, unopened by users (or users would be forced to treat their spam folder as a regular inbox, and lose the benefits of the filtering). Nonprofits in particular howled because, they claimed, their mail is valuable, but they are too poor to pay for the stamps. (If members of non-profits aren't willing to pony up $0.25 per email in member fees, just how valuable are the millions of pieces of mail that non-profits want to send?)
But rather than get into that debate right now (see "Backlash to sender-pays email incentives"), I want to discuss the economics of two different but related approaches to using financial incentives to economically filter spam: the private stamp (Goodmail) approach, and the use of recipient-repudiable bonds ("stamps" vs. "bonds" for short).
The bond approach is similar to stamps, with critical differences. The sender pays for a (digitally-signed) stamp; mail with that stamp goes directly into the reader's inbox, unfiltered. However, after opening the message, the reader can either keep the stamp (push a button in the mail client to "deposit stamp"), or relinquish it back to the sender, which can be interpreted as a message that "I valued this mail, you can send more like it in the future."
How do the differences matter? First, an implementation issue: it is relatively easy for a third-party provider like Goodmail to implement a payment deposit system; it is not nearly so easy, at least right now, for individual email users to receive a micropayment attached to every email and deposit it. Email clients aren't programmed for this, and in any case, the necessary micropayments infrastructure just doesn't exist (yet) at that level of granularity.
Assuming that technical detail can be solved in the near future, how else are the two different? One of the most important differences is the very limited role for recipient preferences in the private stamp approach. A stamp of, say, $0.01, will discourage senders from sending email that is worth less than $0.01 for the sender. But the threshold is being set by the third party (Goodmail, perhaps together with an ESP like AOL), not by individual users, and thus does not directly reflect the value to the recipient of receiving unsolicited email (or not). Arguably, competition between ESPs would push the stamp price to about the right average level over time, but it would not reflect heterogeneity in user preferences.
A bond system could with little or no cost allow each user to set their own threshold for the required size of bond, thus allowing recipients to customize their own mail preferences.
Another problem with the stamp approach is that that goes through this channel pays for a stamp. For mail that both sender and recipient agree is desirable, that incurs unnecessary expense. But perhaps more important, it will prevent some desirable mail from being sent. Suppose the stamp is $0.01, and a sender has mail to send that the sender values at only worth $0.005 if delivered, but the recipient also values at $0.02 if received. The sender won't be willing to buy the stamp, and the mail won't get sent. With a repudiable bond, however, the sender might send a trial message, and if the recipient repudiates the bond, the sender will know the recipient values the mail and will allow similar messages to arrive without a bond payment in the future.
Why won't recipients always keep the bond payment? Well, first, this would just make the system work the same as stamps (except that users get the money, not a third party), so that's not a reason why bonds are worse. However, it also doesn't make sense in the example above. If I want to receive, say, an electronic catalog, but I keep the bond, then the sender may stop sending to me, and I lose out.
This is a very quick review of the two approaches, and yes, of course the issues can be more subtle. See Loder et al. for a scholarly discussion of the two.* Vanquish Labs, a vendor of a bond system, has an online article that critiques the Goodmail stamp approach (February 2006 : CertifiedMail = Certified Disaster).
*Thede Loder, Marshall Van Alstyne, and Rick Wash. "An economic solution to unsolicited communication". Advances in Economic Analysis and Policy, 6 (1), 2006.
Keeping bad stuff out: Making a play on social news sites?
About a month ago, some rumors were about that Google was about to acquire Sun Microsystems. The news got hot when blog stories claiming an acquisition was imminent were promoted to the front page on community/social new site Digg.com. It pretty quickly became clear that the rumors were largely unfounded. What hasn't been quickly resolved is whether or not someone tried to manipulate Digg, possibly to cash in on speculative trading in Google or Sun stock.
The basic idea is simple: get enough shill users to vote for a financially-significant rumor to promote it to the front page, thus automatically getting more widespread attention, and hope that the burst of attention causes a temporary stock price adjustment that can be exploited. (For example, in an acquisition the price of Sun would almost surely increase, and thus gullible readers might start buying it and bidding it up; the scam artist could purchase shares in advance to sell at the inflated price, or sell it short at the bubble price and collect when price returns to normal.)
Digg claims that it almost surely was not manipulated, but it seems clear that such manipulation is possible in user-contributed content news sites. Recall how Rich Wiggins found that people could get flim-flam press releases fed into Google News (here and here), and how authors using pseudonyms have promoted their own books with favorable "reviews" on Amazon.com.
It appears that in the past Digg has been manipulated (though apparently as an experiment, not to manipulate stock prices).
Old problem: how to evaluate effort by programmers?
This is a nice recent example of an age-old hidden action problem for managers: how to compensate brain workers, whose effort is intrinsically hard to monitor? The story here concerns Microsoft's performance incentives for programmers. According to a recent report (by an organization that is trying to organize a union at Microsoft, so take the tone with a grain of salt),
The way it works is that under this system, managers can only give out so many high marks. If everyone on a manager's team did 4.0 work, only two of them might be able to get them.
Benchmarking is a standard, almost necessary way to assess performance in an organization with a large number of brain workers, but no benchmark system is perfect, and it's not obvious which will be the best system for a given environment.
Principal-agent problem in action
Having trouble getting your employees to do the work you want them to do?
April 07, 2006
Another technical "screen" without incentives design facing trouble
CAPTCHAs ("completely automated public Turing test to tell computers and humans apart"!) are an increasingly common technical device to try to "keep the bad stuff out" of many services: users are asked to type in a word displayed with distortion that is easy for most humans to read, but difficult for image recognition software, before they can access various services.
Well, not too surprisingly, programmers have taken on this challenge, and software is getting better at doing the recognition and getting around the CAPTCHAs. And it turns out there's a lot of cheap labor available for spammers and others to hire to thwart them, too.
As a general matter, technical screens are not incentive-compatible, and when the value of getting past them is reasonably high, we can expect that they will get caught up in "arms races" and rarely be highly effective against pollution.
April 05, 2006
ET call ICD
This is an old (2001) story I recently heard about. At the time, the SETI@Home project (which distributes the search for extraterrestial intelligence computations to volunteer machines around the world) was plagued by cheaters, including folks who hacked the client software to report that they were contributing far more CPU cycles than they actually were, apparently in order to get the reputation of being the top contributor in the search for ET. (Can you put that on a resume and get a job?)
Another user-contributed content (or effort, in this case) service that had trouble keeping the bad stuff out.
Posted by jmm at 10:38 PM
Didn't keep the bad stuff out
Here's a user-contributed content service that sure didn't keep the bad stuff out (from their point of view!). Chevy set up an interactive marketing site at which you could create your own customized version of a video commercial for one of their SUVs. A few users had some fun with the freedom. (Good chance the links won't last at the Chevy site for long -- I'll try to find another source when they disappear.)