These questions are at the forefront of   research and were the topic of discussion at his recent talk as part of the IP Osgoode Speaks Series.  Prof. Pasquale brought with him a simple message to his talk: the law ought to be “A Rule of Persons, Not Machines.”

Correcting Bias with Bias

To begin, take the pinnacle actor of our legal system: the judge. Judges are human, after all, and they bring human biases with them to the court room. Studies demonstrate that judicial outcomes can depend on variables such as the performance of the  or whether the decision was made before or after the . In contrast, computer algorithms can produce more consistent legal outcomes across a given set of cases. Humans are biased, after all, so why not replace them?

Algorithms have biases too, answers Prof. Pasquale. Their outcomes depend on the humans that develop their code. Which factors should be given heavier weight in a computer’s decision? How much room should be carved for the protection of constitutional rights? What if the software contains undiscovered errors? How should the system import contemporary societal values into its decision? In considering these questions, Prof. Pasquale shows that greater consistency does not equate to fewer biases.

In addition to these coding difficulties, computer algorithms are much better at evaluating backward-looking inputs than solving forward-looking problems. They may therefore be unable to effectively replace the law-making role of the judiciary. Take, for example, a situation where a computer must decide if constitutional rights should be extended to novel situations. Analysing historical data to determine the likelihood of a human judge allowing such an extension is not difficult, but fully predicting the extension’s effects on society is. It’s almost near-impossible: there are simply too many variables to consider, many of which, such as personal values, cannot be reduced to simple metrics.

Finally, Prof. Pasquale contends that algorithms lack transparency. Human-made judgments reduce legal logic into an intelligible, written form. This writing may be applied, built upon, or criticized by subsequent thinkers in the legal system. Algorithmic judgments do no such thing. Perhaps the coding and mathematics used to process inputs are intelligible, but they provide no guidance as to how the law is, was or ought to be applied.

Blockchain and Property Law: A Legal “Wild West”

Prof. Pasquale turns to  concerns with emerging blockchain technologies like  to demonstrate the real-world legal problems that arise when cold, efficient machines replace human judgement.

Institutions operated by human beings have traditionally been responsible for tracking and facilitating the exchange of goods and services for currency, at least for big-ticket items. If X purchased a car from Y, the bank would ensure a precise number of dollars transfer from X’s account to Y’s account, and the vehicle would be registered with the state in X’s name. Money is traded for ownership, and each step is traceable.

This is not the case with blockchain.

Instead of relying on trusted, human-run institutions to maintain the records, blockchain technology relies on its users – the public at large – to track one-another’s balances on their computers. A transfer’s legitimacy is verifiable because each user holds a secret digital key that they use to authorize the movement of their funds or property. This key is nothing more than a sequence of numbers – it has no name attached to it, and the only requirement to use it is knowledge of its sequence.

Although blockchain is revolutionizing transactions by cutting out the “middle man” and making them instantaneous and secure, they cause some serious legal headache. First, losing your authentication key means that you lose access to its value stored on the blockchain. If your car title is recorded on the blockchain in association with your digital key, losing your key means you’ve lost the only proof you have that you own the car. Compare this to losing your paper title to the car: you’d simply need to stop by your local registry with identification and pick up a new one.

Second, imagine your digital key gets stolen by a hacker who hops in your car and drives off into the sunset. Not only do you lack proof that he stole it, but, for all intents and purposes, the blockchain now considers him the true owner. The law cannot help you.

Finally, consider the law’s benefits and protections conferred on you by virtue of you owning your car. If the car is defective, you can return it to the seller. If your friend doesn’t return the car after you lend it to them, the law can force them to give it back. Consumer protections like these and many others are simply unavailable to you if the law does not recognize your ownership.

Prof. Pasquale makes it clear that although blockchain may become useful for certain applications, it illustrates the danger of replacing human-run systems with purely algorithmic ones. Blockchain is not going to make property law go away.

Algorithmic Accountability

It is a grim future, argues Prof. Pasquale, where machines replace the role of human judgement in its entirety. What role, then, does he see computers filling in the legal world? Artificial intelligence should enrich professional judgement, he suggests, not replace it. What constitutes a “better way” to do law must be decided with a broad, social understanding of the systems that the law serves; not through applying technology for the sole sake of efficiency. There is massive room for computers to improve the way that we do law, but it must be approached through a critical lens.

Principles of algorithmic accountability must be interwoven into these systems. Models must be transparent, data must be unbiased and algorithms must be applied to appropriate tasks. Leading thinkers and the public alike must be able to critique these systems and lend their voices to the software’s development. Technology is strengthened when subjected to an open and free exchange of ideas and criticisms, so legal innovation must evolve with public accountability.

Prof. Pasquale left the room with a parting thought: the legal profession needs to carefully shape the algorithmic systems it deploys, or else those systems will shape the legal profession. Using the words of Douglas Rushkoff, Prof. Pasquale warns,

“Program or be programmed.”

Mike Noel is an IPilogue Editor and a JD Candidate at Osgoode Hall Law School.

The post Algorithmic Accountability: Prof. Frank Pasquale’s Thoughts on Artificial Intelligence in the Law appeared first on IPOsgoode.

Intellectual property law is meant to provide incentives for creators to both engage in useful innovation and make their creative works available to others. In order to do this well, there ought to be an understanding of the kinds of incentives that encourage creativity. “IP law has to have a theory of why people create, how they interact with the things that they create, and how the markets for creativity and innovation are likely to work,” Prof. Buccafusco tells us.

When dealing with this theory of IP law, most of the legal literature and scholarship derives its assumptions on the basis of Rational Choice Theory (RCT). RCT postulates that people weigh the costs and benefits of each decision, and they’ll make decisions that maximize their self-interest by giving them the greatest satisfaction. Prof. Buccafusco then poses the critical question: “are any of these assumptions that underlie the nature of intellectual property law in fact correct?”

His following  seek answers to this question, and they look to do so by finding out which biases cause creators to deviate from the cold, calculated behaviour of RCT.

The Endowment Effect

The first bias Prof. Buccafusco’s examines is the endowment effect – that is, the hypothesis that people place more value in a good when their property right to that good has been established.

He beings by describing an in which coffee mugs were randomly distributed to half of an undergraduate law and economics class at Cornell University. The students assigned a mug each quoted the lowest price that they’d be willing to accept to sell the mug, while the students not assigned a mug gave the highest price that they’d be willing to pay to buy the mug. Under RCT, one would expect that these prices should be about equal. But they weren’t. By simply being endowed with the mug, the sellers’ prices were much higher than the buyers’ prices on average. As a consequence, the number of completed sale transactions was less than what RCT would predict.

So how might this bias play out in an intellectual property market? This is where Prof. Buccafusco’s experiment comes in. They set up a contest in which ten student painters submitted their paintings to compete for a $100 prize. These “Painters” were told that they would be matched with a second group of “Buyers” who would make the painters a cash offer to purchase the right to the Painter’s prize if their painting won the contest. The Painters were asked to write down their lowest price they’d sell this right for and the Buyers the highest price they’d pay. If the prices met, then a transaction would occur; the Buyer’s cash for the Painter’s chances of winning.

In addition to these two groups, an “Owner” group was each randomly assigned ownership of one of the paintings and stood to win the $100 prize if their painting won. Similar to the Painters, they were asked to write down their lowest price for which they’d sell their chances of winning the prize to one of the Buyers. There being a 1 in 10 chance of winning $100, the rational prices of all three groups (ignoring the perceived quality of each painting) shouldn’t deviate much from $10.

The actual median prices were as follows:

Buyers: $17.88

Owners: $40.67

Painters: $74.53

What do these staggering differences tell us? Not only did the endowment effect create a gap in the valuations of the Owners and the Buyers, but the act of “creating” caused an even bigger gap between the Painters and the Owners. “Initial endowment may be really sticky” explains Prof. Buccafusco. His research tells us that the intellectual property market may not be as efficient as RCT might predict: fewer transactions occurred as the result of this bias than may be ideal.

Content Attribution

Prof. Buccafusco tells us the second experiment’s goal of determining “the extent to which people are willing to trade off real dollars for an opportunity to have their name attached to a work that they’ve created.” How much do creators value receiving credit for their work?

In this experiment, photographers are told that a graphic designer is looking for a photograph to use as the background of an image that she is submitting to a contest for a $1000 prize. They’re told that she is going to see the photographer’s photo plus four others and will buy the rights to use one of them. The photographer would get cash in this transaction, but no prize money from the contest that the designer is entering. However, they’re also told that their photo will appear on a major website if the designer wins the contest.

The photographers are randomly split into two groups. Group 1 is given no default attribution – analogous to the system in the United States. That is, they’re told that they will not be given credit for their photograph’s appearance on the website. Group 2 is given default attribution – analogous to the system in Canada. Here, they’re told that they will be given credit for their photograph’s appearance on the website. Both groups are first asked how much money they’re willing to accept to sell the designer their rights to use their photos in this contest: the first group without credit, and the second group with credit.

After stating this price they’re each then told that they have the opportunity to switch the status of their attribution. Group 1 is asked how much money they’d be willing to accept to sell their photo if the designer chose to give them credit. One would expect that this second price would be lower than the first price if content creators value having their names attached to their work. Meanwhile, Group 2 is asked how much money they’d now be willing to accept to sell their photo if the designer decided to remove the photographer’s credit from the website. Because, again, of the value creators place in their names, one would expect that this second price would be higher than the first price.

In essence, Group 1 was now buying their right to attribution and Group 2 selling their right to attribution.

By looking at the average difference between the non-credited and credited prices for each group, it can be determined if the attribution rights were valued more in the no-default or default system. The results were as follows:

Group 1 (No-Default Attribution) Difference: $3.61

Group 2 (Default Attribution) Difference: $14.77

The results are significant. The data shows us that the act of being endowed with attribution caused Group 2 to value being credited for their work over four times as much as Group 1. Because of the endowment effect, people seem less willing to sell their credit than they do to buy it.

What does this tell us in practical terms? Again, “the default creates a kind of stickiness,” says Buccafusco. Being given attribution rights by default might make the access to the works of creators especially sticky, he explains, and this makes the markets for them deeply inefficient. The American system appears to make more sense here: if we want more transactions for creative works to occur thereby increasing the number of works in use, then we may not want to be starting creators off with the right to credited use by default. For efficiency's sake, perhaps we should allow them to bargain for the right at a cheaper price through contracts.

Borrow or Innovate?

Third and finally, Prof. Buccafusco’s seeks to determine how people make decisions to create in markets characterized by sequential innovation. Innovation is built on innovation. Buccafusco illustrates this: we first had Dracula the book, and then we had Dracula the movie. Then came the Twilight novels, the Twilight movies, the Twilight Fan Fiction, and 50 Shades of Grey. Innovation almost never happens in isolation.

There are two choices for a creator in a market with existing works: borrow or innovate. To borrow is to pay for a license, and to innovate is to work around the existing IP. Both of these options have costs. Borrowing has licensing fees, and innovation has risks of uncertainty. As existing rights fill up the innovation space and there’s less room to innovate, we’d expect this innovation would decrease as it becomes harder to do.

The experiment focused on this borrow/innovate decision. In it, subjects are given Scrabble letters with which to earn points by forming words. They are given two options: they can borrow from a pre-existing solution that is first shown to them, or they can innovate and create a solution from scratch. Innovating gives them bonus points, while borrowing does not.

The subjects are split into three groups based on the quality of the pre-existing solution shown to them without being told this quality: 60% of the best solution, 80% of the best solution, and 100% of the best solution. If the subjects were to act rationally, one would hypothesize that the subjects presented with more complete solutions would be more likely to borrow from the solutions given to them (and less likely to innovate).

The results are unexpected:

60% Group: 46.8% innovators

80% Group: 40.9% innovators

100% Group: 85.7% innovators

We saw a drop in the innovators between 60% and 80%, but why did we see an increase in the number of innovators between 80% and 100%? Prof. Buccafusco thinks that the subjects were influenced by an “innovation heuristic.” Subjects didn’t look at the value of the solutions in terms of the points they were worth; they simply considered how easy it would be to come up with some new words that weren’t in the pre-existing solution. Since the 100% solution had quite difficult words, the subjects had little trouble coming up with alternative words of their own without regard for the fact that their low values make them terrible Scrabble words. And they lost a lot of points because of this.

So what does this mean for innovation? “We tend to think about innovation as always good,” Prof. Buccafusco continues. The bias revealed in this experiment shows that creators may be predisposed to spend energy coming up with solutions that already exist for license. “Innovation is often terrible,” Buccafusco concludes. It often leads to “producing stuff that we already have, but expending substantial amounts of resources to do so.”

Prof. Buccafucso's research shows us that creators' decision-making in IP markets appears to be significantly altered by these three biases. Perhaps the resulting market "stickiness" is even causing the decisions of these creators and their welfare to diverge. If we want our intellectual property law to be effective when it provides incentives for creators to engage in useful innovation and make their creative works available to others, then it's clear that we need to take into consideration biases like these and others.

Mike Noel is an IPilogue Editor and a JD candidate at Osgoode Hall Law School.

[1] "Valuing Intellectual Property: An Experiment" is available at SSRN:

The post (Ir)Rational Choice Theory: Prof. Chris Buccafusco’s Search for the Biases of Creativity appeared first on IPOsgoode.