The AI arms race in hiring may cause the market for high-quality job applicants to fail

Two years ago, which is a lifetime or more ago in generative AI years, I wrote a post about the impact of large language models on job interviews. I concluded that post with:

So, perhaps there are offsetting benefits of AI on the job interview process, if the AI is being used by the employer. If AI leads to job interviews that are more effectively able to screen for high-quality job candidates, by reducing bias in the interview process, then that can be a good thing. The AI might also be better able to ask the searching questions that distinguish high-quality and low-quality job candidates.

Of course, that assumes that the AI is interviewing a human job candidate. What happens when a job candidate, being interviewed by an AI job interviewer, is being given real-time prompts on how to answer by an AI chatbot? Or, in a Zoom job interview, the job candidate simply replaces themselves with an avatar or a 'deep fake' video of themselves generated in real time, and using an AI-scripted voiceover. If it hasn't happened already, it is going to be happening soon. Will that be the death of the job interview as a screening tool? Time will tell.

Time has told, indeed. The Financial Times reported back in May (paywalled):

You can almost hear the howls of frustration from HR departments. Jobseekers have discovered artificial intelligence and they’re not afraid to use it. Employers have become snowed under by people using the new tools to churn out impersonal applications. Some applicants are using AI to bluff their way through online assessments, too...

So it should be no surprise that jobseekers have turned to new generative AI tools such as ChatGPT to speed up or “game” a process that already felt dehumanised. Videos have even appeared on TikTok in which people demonstrate how to use ChatGPT to provide answers to questions in asynchronous video interviews, which the applicant then simply reads out.

The problem here is that the online tools that employers have been using in recent years as a tool for screening job applicants, have suddenly become less effective due to generative AI. The issue is one of adverse selection, which my ECONS102 class covered last week.

Job applicants know their quality as a worker (how hard-working, conscientious, smart, etc. they are), but employers do not. Quality is private information. Faced with a lack of information about the quality of each job applicant, employers would assume that every job applicant is low quality. This is referred to as a pooling equilibrium (because all job applicants are pooled together and treated the same). Employers would offer low wages to each new hire, because they assume that the new hire is low quality. High quality job applicants don't want to be treated as if they are low quality and receive low wages, so they would reject any low wage offer, and stop applying for jobs. Only low-quality job applicants would be left in the market. The market would fail for high-quality job applicants.

Of course, employers have found ways to deal with this situation. When the uninformed party (the party that doesn't know the private information, which in this case is the employers) try to reveal the private information (about the quality of job applicants), economists call that screening. Screening in a job application setting will be effective if it accurately reveals the quality of the job applicant. In-person job interviews used to be used as a screening tool. More recently, as noted in the Financial Times article, employers have moved to more technological solutions, such as 'asynchronous video interviews'.

With the rise of generative AI, those new screening tools have suddenly become ineffective, because they can no longer distinguish the high-quality job applicants from low-quality job applicants who are using generative AI. Generative AI has broken the current approach to hiring. As the Financial Times article notes:

Some online assessments are, for now, less vulnerable to AI use, such as those that involve playing short games. But I wouldn’t be surprised to see the return of mass in-person test centres for technical skills assessments.

I wouldn't be surprised to see the return of in-person job interviews as well. Of course, smart glasses combined with real-time generative AI assistance may render even in-person methods of assessment vulnerable, if not now then certainly in the near future. The unfortunate outcome of all of this is that the market for high-quality job applicants will soon be failing, and no one (not employers, not high-quality job applicants, and not even low-quality job applicants) will benefit from that.

Read more:

• The impact of large language models on job interviews, from both sides

Information asymmetry, adverse selection, and large language models

In an article in The Conversation earlier this week, Nicholas Davis (University of Technology Sydney) wrote about information asymmetry as it relates to large language models (like ChatGPT):

A lack of skills and experience among decision-makers is undoubtedly part of the problem. But the rapid pace of innovation in AI is supercharging another challenge: information asymmetry.

Information asymmetry is a simple, Nobel prize-winning economic concept with serious implications for everyone. And it’s a particularly pernicious challenge when it comes to AI...

AI creates information asymmetries in spades. AI models are technical and complex, they are often embedded and hidden inside other systems, and they are increasingly being used to make important choices.

Balancing out these asymmetries should deeply concern all of us. Boards, executives and shareholders want AI investments to pay off. Consumers want systems that work in their interests. And we all want to enjoy the benefits of economic expansion while avoiding the very real harms AI systems can inflict if they fail, or if they are used maliciously or deployed inappropriately.

Davis is generally talking about information in relation to AI safety, but I want to take this post in a slightly different direction first, about AI quality. I will return to AI safety at the end.

When we consider AI quality, information asymmetry will cause a problem of adverse selection, ultimately causing the market for high-quality large language models to fail. So, let's consider how this market fails. There is information asymmetry in relation to the quality of large language models. AI firms know whether their large language model (LLM) is high quality or not, but users do not know. The 'quality' of a large language model is private information. Because of this information asymmetry, a LLM user should assume that any LLM is low quality. This creates a pooling equilibrium, where the user is only willing to pay however much they value access to a low-quality LLM. AI firms with high-quality LLMs, which are more expensive to train and run, will not be willing to accept users who are only willing to pay a small amount. These AI firms with high-quality LLMs drop out of the market. The only AI firms that are left will be those with low-quality LLMs. The market for high-quality LLMs fails.

A key part of the problem here is that LLMs may be experience goods. Experience goods are goods where the consumer doesn't know the quality (or some other important attribute) until after they have consumed it (they are called experience goods, because consumers don't find out what they have really bought until they have experienced it). The quality of a LLM isn't revealed to the user until after they have used it, and seen what the LLM outputs in response to a given prompt. However, users are not experts, and they might not even know the quality of the output after they receive it. That would make LLMs not experience goods, but credence goods (goods where some of the characteristics, the credence characteristics, are not known even after the good has been consumed). It seems likely that for most non-expert LLM users, LLMs will be a credence good.

So, how can the adverse selection problem be solved for LLMs? In other words, how can we move from the pooling equilibrium to a separating equilibrium (where LLM users can separate the high-quality LLMs from the low-quality LLMs)? There are two options: (1) screening; and (2) signalling. Screening involves the uninformed party (the LLM user) trying to reveal the private information (the quality of the LLM). This is unlikely to work well, because the LLM user typically isn't an expert. The only way that they can reveal the quality of the LLM themselves is to try it out, but if LLMs are actually a credence good, then even trying them out won't reveal their quality.

That leaves signalling, which involves the informed party (the AI firm) credibly revealing the private information. Only an AI firm with a high-quality LLM would want to attempt signalling, to try and reveal that their LLM is high quality (and AI firm with a low-quality LLM is unlikely to want to reveal that their LLM is low quality). In order for a signal to be effective, it needs to meet two conditions. First, it must be costly. And second, it must be costly in such a way that the AI firms with low-quality LLMs wouldn't want to attempt the signal.

What might be good signals for an AI firm with a high-quality LLM? Firms that sell experience goods have come up with a number of options, including free trials, testimonials from other customers, branding or advertising. In this context, a free trial might not work. If LLMs are a credence good, then a free trial won't reveal the private information (and we would expect to see AI firms with low-quality LLMs also offering free trials). So, while free trials are costly (they come with an opportunity cost of foregone income for the AI firm), they don't appear to be costly in a way that AI firms with low-quality LLMs wouldn't want to attempt them.

Testimonials from other customers might be an option, but only if the LLM user can trust those other customers. So, testimonials from acknowledged experts in AI might be worthwhile, but testimonials from other non-expert users might not. Are testimonials costly to the firm? Probably not greatly, but they will be more costly for an AI firm with a low-quality LLM. They might have to pay for a testimonial, whereas an AI firm with a high-quality LLM might receive praise from an expert without payment. The problem here, though, is that a non-expert user may not be able to tell the 'experts' giving testimonials apart, and so may not trust any testimonial.

What about branding or advertising? These are costly to a firm, but it's unlikely that they are costly in a way that would prevent AI firms with low-quality LLMs from attempting them.

All of this is bad news for the LLM user. They can't tell high-quality LLMs and low-quality LLMs apart, they probably can't use screening to help tell them apart, and the AI firms with high-quality LLMs probably can't use signalling to reveal that their LLM is high quality. We're left in a situation where the LLM user just has to assume that any given LLM is low quality (the pooling equilibrium).

So, what's left? Davis suggests that:

Well-designed guardrails will improve technology and make us all better off. On this front, the government should accelerate law reform efforts to clarify existing rules and improve both transparency and accountability in the market.

Of course, this is intended as a solution to information asymmetry about AI safety, not a solution to the adverse selection problem related to AI quality. However, now that we've considered the market failing in relation to AI quality because quality is a credence characteristic, it might be worth considering: is AI safety also a credence characteristic?

If AI safety is a credence characteristic, then safety isn't revealed even after the LLM has been used. Then we end up in a similar situation for AI safety as we do for AI quality, in that there is little that users can do to separate the safe LLMs from the unsafe LLMs. We would have a AI safety pooling equilibrium, where all LLM users would have to treat LLMs as potentially unsafe. Since LLM users would not be willing to pay more for LLMs that are safe (since they couldn't be sure about safety), and because safety presumably costs AI firms more, the AI firms with safe LLMs will start to drop out of the market. That would leave only the unsafe (but regulated) LLMs available, which seems like a suboptimal outcome. This isn't an argument against AI safety regulation, but hopefully provides some food for thought about its possible consequences.

Have we reached peak online dating?

On the surface, the persistence of online dating is somewhat of a mystery. As I explained in this post, online dating should have serious adverse selection problems:

Adverse selection arises when one of the parties to an agreement (the informed party) has private information that is relevant to the agreement, and they use that private information to their own advantage at the expense of the uninformed party. In the case of online dating, the private information is the quality of each potential date. They know if they are high quality or not, but no one else knows. That could lead to a pooling equilibrium, where every online dating subscriber assumes that everyone they are matched with is low quality (because they can't tell the high quality and low quality dates apart, and assuming that every match is high quality is a recipe for disaster). High quality dates don't want to be treated as if they are low quality, so they drop out of the online dating market. Eventually, the online dating market only has low quality dates. The 'market' for high quality dates fails.

The online dating market deals with this through screening (an activity that allows the uninformed party to reveal the private information). Screening is achieved by allowing subscribers to chat with the people they are matched with, which they can use to work out who is high quality and who is low quality. However, screening is imperfect, and costly. And the costs tend to be higher when the average quality is low. So, unless the online dating services can keep the quality of their users high, eventually things will turn out bad for them. And, as I noted earlier, that may be increasingly difficult.

So, it was perhaps inevitable that we would get to a point where even screening is not enough, as reported by the Financial Times this week:

“Trying to engage young women is the biggest struggle for dating apps,” said Rebecca McGrath, associate director for media and technology at Mintel. “Significant gender skew means it is harder for men to find matches and, subsequently, women often become bombarded, making the experience worse for all.”...

The number of paid subscribers on Tinder fell to under 10mn in the three months to March, a sixth consecutive quarterly decline. Monthly active users, the majority of whom use the app’s free services, have fallen steadily since 2021, according to figures from Sensor Tower. Bumble also showed a first-quarter drop in the number of active users, data from the app-tracking service showed, even though paid subscribers remained steady.

The declines in users come as reports increase of so-called dating app fatigue. A survey by Bumble, for example, found that 70 per cent of women using the app had experienced “burnout”...

Match Group’s [chief executive Bernard] Kim said Tinder’s rebrand in 2023 was expected “to have some positive impact on users, particularly women and Gen Z”. He noted however that paid subscriber growth would come primarily from “product innovation”, which includes improving profile quality, moderation and the accuracy of its algorithmic matchmaking.

'Burnout' is one consequence of an increasing cost of screening in online dating. It's becoming more difficult for women to find high-quality matches, and they are starting to switch off. The FT article discusses various ways that the apps are trying to keep women engaged. However, to solve the problem, they need to reduce the cost of screening. The option that may come closest is this one:

Tinder said in February it was expanding its own identity verification programme, which compares a video selfie with a passport or driving licence, as well as the images on a user’s profile. Bumble, which already offered a similar feature, said it had enhanced its “computer vision model for likeness comparison” in the first quarter in order to improve verification.

At the least, that will reduce the impact of fake accounts, bots, and catfishing. However, it still doesn't address the challenge that large language models pose for online dating. I'm still expecting a non-trivial amount of the traffic on online dating apps to become ChatGPT (or Claude or whatever) talking to ChatGPT.

Read more:

• Who loses status, and who gains status, with ChatGPT
• AI may be the killer app, as well as the killer of (dating) apps
• Shareholders of dating apps should be freaking out right now

Shareholders of dating apps should be freaking out right now

My ECONS102 class has been covered the economics of information this week, including asymmetric information and signalling. So, it is timely to revisit the topic of online dating (which I have discussed earlier this year, here and here), particularly given the rise of artificial intelligence. As the Financial Times reported last month (paywalled):

Terrible news from the world of online dating. As if a parade of dubious romantic prospects and dead-end chats wasn’t bad enough, artificial intelligence has dipped its toe into the dating pool.

Eleven years ago, Tinder helped to turn dating into a series of quick-fire interactions on the internet. But for some jaded users, even writing “Hi” to a romantic prospect is now too much effort. Tech start-ups such as Rizz and YourMove AI are gaining a foothold in the sector by offering AI assistance in creating witty opening lines and appealing profiles.

Meeting strangers on the internet is by nature a random affair. Artificial chat is at least a less sinister way to help that process along than asking users to swab the insides of their mouth, as DNA-dating app Pheramor once did. But a proliferation of AI-assisted conversations suggests that eventually dating apps will simply be full of computers trying to woo other computers.

One start-up even offers the chance to watch this exact premise unfold. Teaser AI asks users questions about themselves and their personalities and then crafts AI-generated chat that is designed to mimic them. When individuals match, they can sit back and watch as their chatbots try to chat each other up.

I'm sure I wasn't the first to point out that we would eventually find ourselves in a world where one instance of ChatGPT was chatting up another instance of ChatGPT, with no humans involved. On top of being a somewhat amusing situation, there is a downside in this development. It will simply make it more difficult for someone genuinely looking for love to find a match online.

To see why, we need to discuss adverse selection. Adverse selection arises when one of the parties to an agreement (the informed party) has private information that is relevant to the agreement, and they use that private information to their own advantage at the expense of the uninformed party. In the case of online dating, the private information is the quality of each potential date. They know if they are high quality or not, but no one else knows. That could lead to a pooling equilibrium, where every online dating subscriber assumes that everyone they are matched with is low quality (because they can't tell the high quality and low quality dates apart, and assuming that every match is high quality is a recipe for disaster). High quality dates don't want to be treated as if they are low quality, so they drop out of the online dating market. Eventually, the online dating market only has low quality dates. The 'market' for high quality dates fails.

Of course, the market typically hasn't failed. People find love in online dating, so clearly the online dating apps are doing something to reduce the adverse selection problem. One thing that they do is offer the ability for subscribers to chat with the people they are matched with. Chatting allows the subscriber to work out (albeit imperfectly) who is high quality and who is low quality. When applied by the uninformed party, this is referred to as screening.

It also allows the high-quality dates to signal that they are high quality. To be effective, a signal must meet two criteria. First, the signal must be costly (otherwise everyone, even those who are lower quality dates, would provide the signal). And second, it must be costly in a way that makes it unattractive for the lower quality dates to attempt (such as being more costly for them to engage in). Engaging in witty and flirty conversation (or whatever type of conversation a subscriber is looking for, if not witty and flirty) on a dating app is costly - it takes time and effort. It will also be less costly for high-quality dates, for whom the desired type of conversation may come more naturally. It is much easier to be yourself in conversation than it is to pretend to be someone you are not. So, conversation on dating apps is an effective signal, albeit an imperfect one.

The problem with AI is that it eliminates the signalling value of conversation on the app. An uninformed subscriber should no longer find the conversation on the app to be a credible signal of the quality of the match they are engaging with. They can't use the conversation for screening either. After matching, the best option is likely to be, as I suggested earlier, in-person meet-ups.

The shareholders of online dating services should be freaking out right now. Their business model, which is built on providing high-quality matches, is under threat by AI. Once the novelty of watching their AI wingman chatting up their match's AI wingman wears off, subscribers will realise that paying for the dating app has become a waste of time. In fact, here's the share price performance over the last six months of the three biggest dating app companies, Match Group (MTCH) in blue, Bumble (BMBL) in black, and Grindr (GRND) in green (original data taken from here):

While Match is up 10.4 percent over that time, Grinder is down 19.5 percent, and Bumble is down 25.3 percent. It's not looking good.

Read more:

• Who loses status, and who gains status, with ChatGPT
• AI may be the killer app, as well as the killer of (dating) apps

The impact of large language models on job interviews, from both sides

The Financial Times had an interesting article (paywalled) recently about the impact of artificial intelligence on recruitment:

Students applying for graduate jobs this summer can take advantage of a new personal interview coach. If they send over a specific job description they can receive tailored interview questions and answers — and feedback on their own responses — all for free.

The coach, offered by the job search engine Adzuna, is not human but an artificial intelligence bot known as Prepper. It can generate interview questions for more than 1mn live roles at large companies, in industries ranging from technology and financial services to manufacturing and retail.

For a graduate job in PwC’s actuarial practice, the chatbot spits out questions such as: “What skills do you think an actuarial consultant should have?” and “How would you explain actuarial concepts to a client who is not from a finance background?”. When a user answers a question, Prepper generates a score out of 100, and tells them which parts worked well and what was missing.

Prepper is part of a new wave of chatbots powered by generative AI — from ChatGPT to Bard and Claude.

As I noted in my previous post, employers are trying to overcome an adverse selection problem. Job candidates know whether they are high quality or not, but employers do not. High-quality job candidates use signalling to reveal to employers that they are high quality, as I noted in my previous post. Employers use screening to try to identify the quality of job candidates. One screening tool is the job interview. Candidates' likely quality as employees is revealed through the job interview process. However, if AI is prepping job candidates, then that reduces the efficacy of the screening process. If low-quality job candidates can be effectively prepped to seem like they are high-quality job candidates, the screening process fails to solve the adverse selection problem.

On the other hand:

Grace Lordan, an economist at the London School of Economics and director of The Inclusion Initiative, which studies diversity in corporate settings, says companies, particularly technology groups, are experimenting with generative AI to conduct initial interviews.

“One of the biggest areas of bias is actually the interview,” she says. “This is when people’s affinity bias, or representative bias, which means choosing people who look like others in the organisation, comes in.”

AI-conducted interviews could go some way to removing that bias, she says. “Generative AI is quite convincing as an avatar. Using AI as another serious data point will allow pushback from the machines [against human bias].”

So, perhaps there are offsetting benefits of AI on the job interview process, if the AI is being used by the employer. If AI leads to job interviews that are more effectively able to screen for high-quality job candidates, by reducing bias in the interview process, then that can be a good thing. The AI might also be better able to ask the searching questions that distinguish high-quality and low-quality job candidates.

Of course, that assumes that the AI is interviewing a human job candidate. What happens when a job candidate, being interviewed by an AI job interviewer, is being given real-time prompts on how to answer by an AI chatbot? Or, in a Zoom job interview, the job candidate simply replaces themselves with an avatar or a 'deep fake' video of themselves generated in real time, and using an AI-scripted voiceover. If it hasn't happened already, it is going to be happening soon. Will that be the death of the job interview as a screening tool? Time will tell.

Signalling theory suggests that we may need two categories of university degree

In The Conversation last week, Ananish Chaudhuri (University of Auckland) argued the case for having two categories of university degree:

And while cognitive abilities such as reading, writing and maths matter, so too do social skills such as empathy, resilience and an ability to work in diverse groups and with diverse views...

Universities play a crucial role in developing these skills. But the emerging two groups of students – on campus and off – are not getting the same education. The increasing emphasis on online instruction and exams is devaluing degrees...

This suggests we may need to distinguish between online and on-campus students in each of our courses. The course content will be the same, but the assessment methods will be different.

Online students can take tests, quizzes and exams remotely. Some of this may also be available to on-campus students. But on-campus students will be expected to come to lectures regularly, ask questions, write, speak and engage in interactive tasks, including group work.

Would students sign up for on-campus courses but simply not attend? This could be prevented by making sure each student completes tasks that earn participation marks that count toward on-campus credits. If they fail to do so, they will automatically become online students.

Is this unfair to online students? Not necessarily. Many with jobs may prefer it. In any event, they will have to consider whether the benefits of coming to campus are worth it in terms of job prospects or earning potential.

It is worth reviewing the purposes of education from the perspective of the student. On the one hand, education provides useful cognitive and non-cognitive skills that have value in the workforce. In theory, skills development need not necessarily be different between in-person students and online students. However, some 'transferable skills' like teamwork and interpersonal skills are more difficult to develop in an online environment (as Chaudhuri notes).

On the other hand, education provides a signal to employers about the quality of the job applicant. Signalling is necessary because there is an adverse selection problem in the labour market. Job applicants know whether they are high quality or not, but employers do not know. The 'quality' of a job applicant is private information. High-quality (intelligent, hard-working, etc.) job applicants want to reveal to employers that they are hard-working. To do this, they need a signal - a way of credibly revealing their quality to prospective employers.

In order for a signal to be effective, it must be costly (otherwise everyone, even those who are lower quality job applicants, would provide the signal), and it must be costly in a way that makes it unattractive for the lower quality job applicants to attempt (such as being more costly for them to engage in).

Qualifications (degrees, diplomas, etc.) provide an effective signal (they are costly, and more costly for lower quality applicants who may have to attempt papers multiple times in order to pass, or work much harder in order to pass). So by engaging in university-level study, students are providing a signal of their quality to future employers. The qualification signals to the employer that the student is high quality, since a low-quality applicant wouldn't have put in the hard work required to get the qualification. Qualifications confer what we call a sheepskin effect - they have value to the graduate over and above the explicit learning and the skills that the student has developed during their study.

Now we can see where there is a key difference between in-person and online education. In-person education is more costly than online education, even if the tuition fees are the same, because it requires effort for a student to get themselves onto campus and into class each day. Higher-quality students (who will be higher-quality job applicants) are more likely to be conscientious and make this effort than lower-quality students. So, having an in-person education should provide an additional signal of quality to employers, over and above the signal provided by the degree itself.

However, if there is no way for the in-person students to distinguish themselves from the online students, then the value of the signal provided by in-person education is lost. Chaudhuri argues that the solution is to create a two-tiered qualification system. One (in-person) qualification would convey a signal of high quality to employers, and the other (online) qualification would convey a signal of lower quality to employers.

I'd argue that this already happens to some extent. It's the reason why Massive Open Online Courses (MOOCs) haven't displaced traditional education, despite being free or low cost. Employers can tell the signalling value of a degree when compared with studying online using a MOOC. The problem that Chaudhuri notes is that students' university transcripts probably don't clearly identify students who have done their study in-person from those who have done their study online.

Of course, savvy employers don't only rely on job applicants providing signals. Employers engage in screening, in order to reveal the private information for themselves. They do job interviews and subject job applicants to various pre-employment tests, which helps the employer to tell the high-quality and low-quality job applicants apart. It would not surprise me at all to learn that, as part of the standard job interview script, employers now ask how much of their degree a job applicant completed online. However, job applicants can lie in job interviews, so screening is unlikely to be as effective as signalling in solving the adverse selection problem. Perhaps, two categories of university degree is the best solution after all.

Regardless, the takeaway for students is, as I noted in this 2017 post, that they should be acutely aware of the signals that they are sending to future employers. Education is a signal, but the type of education matters as well.

Read more:

• Why all students need to understand adverse selection and signalling

Twitter's blue tick is losing its value as a signal

One of Elon Musk's first actions as the new owner of Twitter was to announce a change to Twitter's 'verified status'. Previously restricted to verified real people (and usually to those with some celebrity status), users would now be able to get their own blue tick for just US$8 per month (or equivalent in other countries). That change comes with immediate problems, as outlined in this article in The Conversation by Timothy Graham (Queensland University of Technology):

...Musk’s US$8 blue tick proposal is not only misguided but, ironically, likely to produce even more inauthenticity and harm on the platform.

A fatal flaw stems from the fact that “payment verification” is not, in fact, verification...

Although Twitter’s verification system is by no means perfect and is far from transparent, it did at least aspire to the kinds of verification practices journalists and researchers use to distinguish fact from fiction, and authenticity from fraud. It takes time and effort. You can’t just buy it.

Despite its flaws, the verification process largely succeeded in rooting out a sizable chunk of illegitimate activity on the platform, and highlighted notable accounts in the public interest. In contrast, Musk’s payment verification only verifies that a person has US$8.

Payment verification can’t guarantee the system won’t be exploited for social harm. For example, we already saw that conspiracy theory influencers such as “QAnon John” are at risk of becoming legitimised through the purchase of a blue tick.

Allow me to put an economics lens on the problems here. It relates to asymmetric information, adverse selection, and signalling.

First, there is asymmetric information on Twitter. Each Twitter user knows whether they are an authentic user and not a bot, a troll, or a scammer. However, each Twitter user doesn't know which other users are bots, trolls, or scammers. That leads to a problem of adverse selection. Each Twitter user, knowing that there are lots of bots, trolls, and scammers, doesn't know for sure if any other account is authentic or not. Whether any Twitter user is a bot, troll, or scammer, or not, is private information (known only to the user themselves, and not to others - that's why it is called asymmetric information). To avoid being trolled or scammed, a Twitter user's best (risk averse) option is to assume that every other account is a bot, a troll, or a scammer. This is what we refer to as a pooling equilibrium (because all other users are pooled together in the Twitter user's mind, as if they are all the same, and low quality). Since Twitter users don't want to engage with bots, trolls, or scammers, if they are assuming that every other account is like that, there is little point being on Twitter. Authentic Twitter users start to drop off the platform, and eventually the only 'users' left are bots, trolls, and scammers. This is what we call an adverse selection problem - each Twitter user wants to engage with other authentic users, but all they find are bots, trolls, and scammers.

Of course, Twitter hasn't collapsed as a platform, so it must have found a way to deal with this adverse selection problem. One way is through the blue tick (verified user) status, granted only to authentic users. The blue tick is a signal to other users that the user with the tick is authentic. In order for a signal to be effective though, it needs to meet two conditions. First, a signal must be costly. The blue tick was previously difficult to obtain, as users had to go through an authentication process (including verifying their identity). So, while there was no monetary cost, there was a cost in terms of time and effort. Second, a signal must be costly in such a way that those with low-quality attributes would not attempt it. Since the authentication process required identity verification, this was a process that bots, trolls, and scammers would be unlikely to attempt. So, Twitter's blue tick seems to meet the conditions of being an effective signal that users are authentic (despite some counter-examples). So, Twitter users could be fairly sure that they were interacting with authentic users, if those users had the blue tick. This is a separating equilibrium (because Twitter users are able to separate the authentic accounts that they want to interact with, from the bots, trolls, and scammers, that they don't want to interact with).

That is all about to change. As Graham's article in The Conversation noted, under the new regime all that it will take for a user to obtain Twitter's blue tick is the payment of US$8 per month. While that meets the first condition of an effective signal (costly), it fails on the second condition, because almost any bot, troll, or scammer with US$8 per month would be willing to pay for the tick. The blue tick will cease to be a signal of an authentic account.

Is that the end of Twitter though? Signalling is only one way to overcome the adverse selection problem. The alternative is screening - where the Twitter user themselves tries to reveal whether another account is authentic or not. That requires a bit of detective work on the part of each Twitter user, and is going to be far from perfect. Perhaps each Twitter user is best only interacting with people that they know personally, or people they have heard of and can be fairly sure are not fake accounts. Avoiding interacting with new accounts, that have few followers, and tweet mostly junk, has always been a good strategy, but will become even more important once the blue tick loses its value as a signal.

Twitter probably won't die as a result of the changes to the blue tick. But it's certainly not going to be as user friendly as before.

Professional job interview 'proxies' and adverse selection

One of the examples of adverse selection that I use in my ECONS102 class is the adverse selection experienced by employers when trying to find high quality workers. Adverse selection arises when one of the parties to an agreement (the informed party) has private information that is relevant to the agreement, and they use that private information to their own advantage at the expense of the uninformed party. In the case of an employment situation, the private information is about the job applicant's quality or productivity as a worker. The worker knows this, but the employer doesn't. That could lead to a pooling equilibrium, where the employer has to assume that every job applicant is low quality, and would consequently only offer a low wage. High quality workers won't work for a low wage (because they know that they are worth a lot more to the employer), so they would turn down any job offer. The labour market for high quality workers fails.

Of course, in the real world that doesn't happen. That's because employers have adopted screening methods, in order to reveal the private information (about whether a job applicant is high quality or low quality). Specifically, the job interview is a screening method. That creates a separating equilibrium, where employers can separate the high quality job applicants from the low quality job applicants. The job interview is effective as a screening tool, because any applicant can easily lie on a CV, but it is a lot harder to lie in an interview. Or is it? Business Insider reported earlier this week:

Some job candidates are hiring proxies to sit in job interviews for them — and even paying up to $150 an hour for one.

In a recent Insider investigation into the "bait-and-switch" job interview that's becoming increasingly trendy, one "professional" job interview proxy, who uses a website to book clients and keeps a Google Driver folder of past video interviews, said he charges clients $150 an hour...

The "bait-and-switch" interview works like this: a job candidate hires someone else to pretend to be them in a job interview in hopes they will secure the job. When the job starts, the person who hired the proxy is the one to show up for work...

With an increasing amount of job interviews happening over the phone or video chat due to remote work environments, the "bait-and-switch" trend is getting easier, experts told [Insider's Rob] Price. 

If employers can no longer be sure that their job interviews are working as screening tools, then job interviews no longer lead to a separating equilibrium. Employers (and job applicants) would be back at the pooling equilibrium, where high quality workers wouldn't be able to be offered high wages (because they can't be easily distinguished from low quality workers).

However, screening isn't the only tool available. The high quality job applicants can use signalling to reveal the private information about their quality as a worker. An effective signal is costly, and is costly in such a way that the low quality job applicants would not want to attempt it. Education credentials are a form of signalling - they are costly to obtain, and potentially more costly for low quality workers, who might take longer to get a qualification, or might have to work harder to do so (and either of those situations makes a qualification less attractive to them).

Contrary to the hopes or expectations of pundits like Bryan Caplan (whose book The Case Against Education I reviewed back in 2019), education as a signalling tool may actually become more important. If job interviews are ineffective because of proxies, I can easily imagine that the whole human resource management battery of assessment tests will also be ineffective, along with various skills assessments. Unless employers (or HR consultants) can come up with some method of credible identity verification at the time of the interview or assessment, employers may become increasingly sceptical of those screening methods. And that only leaves signalling, of which education is potentially the most important type. It really illustrates just how important it is for students to understand the signalling value of their education.

[HT: Marginal Revolution]

Is there a signalling explanation for the high cost of article re-formatting?

My post yesterday highlighted the high cost of re-formatting papers for submission to academic journals, with a total cost estimated at US$1.1 billion per year. I was reflecting on this today, and perhaps there is a reason for this high cost, aside from each journal's publishers wanting to maintain a particular style that distinguishes the journal from other journals.

Perhaps the time cost of formatting (and re-formatting) is a way of dealing with asymmetric information, specifically adverse selection. At the time of submission (i.e. before peer review, and before the editor has even read the abstract), the quality of a paper submitted to a journal is known to the authors (presumably), but not to the publisher. The quality is therefore private information. Since the publisher doesn't know whether any particular article submission is high quality or not, their best option (aside from editorial and peer review, which I will come to in a moment) is to assume that every submission is low quality. This is a pooling equilibrium - all article submissions are pooled together as if they are the same quality. The publisher may as well pick randomly from this pool of submissions of unknown quality, leading to a journal that gains a reputation for low quality articles (this is basically the business model of some publishers that offer 'pay-to-publish'). Authors with high quality articles would avoid those journals, lowering the quality of the article submissions further. Eventually, only the lowest quality articles get submitted, and published. This is a problem of adverse selection, because authors want their high quality articles to be published, but in the end, only low quality articles get published.

The way to solve an adverse selection problem is to reveal the private information - in this case, to identify which article submissions are high quality, and which are low quality. That would lead to a separating equilibrium, where low quality articles are rejected and high quality articles are accepted and published. The publisher can reveal this information through editorial and peer review. The editor reads the abstract (and perhaps the paper), and decides whether it is worthwhile sending for review, and if not the submission is desk-rejected. If the paper is sent out for peer review, its quality is judged by the peer reviewers. These processes are a form of screening - where the uninformed party (the publisher) tries to reveal the private information (about the quality of the article submission).

However, screening is not the only way to deal with an adverse selection problem. And the problem with screening through editorial and peer review is that it takes up a lot of time. The editor has to spend time reading and making decisions, and the peer reviewers have to spend time reading and writing reports. The alternative to screening is signalling - where the informed party (the authors) reveal the private information themselves.

Now, of course, if you simply ask the authors whether their paper is high quality or not, every author (even those with low quality articles) would respond that their paper is high quality. In order for a signal to credibly reveal the private information and be effective, it needs to meet two important conditions: (1) it needs to be costly; and (2) it needs to be more costly in a way that makes it unattractive for those with the low quality attributes to attempt. One way that signals could meet the second condition is if they are more costly to the authors of low quality articles.

By having idiosyncratic formatting requirements, it is clear that submitting to a journal is costly, so it meets the first condition. What about the second condition? I can see two ways that we could argue that costly re-formatting is more costly for authors of low-quality articles than for authors of high-quality articles.

First, authors of low-quality articles will realise that their submission has a lower chance of acceptance than a high-quality article does. That means that they can anticipate having to go through the re-formatting and submission process more than once, leading to a higher cost. To avoid this higher cost, they may avoid submitting to high quality journals (thereby revealing that their paper is low quality). Authors of high-quality articles know their submission is high quality and has a higher chance of being accepted, so they know they face a lower cost of re-formatting, and will be more likely to submit to a high-quality journal.

Second, authors of high-quality articles are more likely to be high-quality academics, who are well supported by their institutions, have research grants, and may have research assistants who can handle the re-formatting at relatively low (salary, or monetary) cost. Authors of low-quality articles are less likely to have this support, and have to handle the re-formatting themselves, at (presumably) higher salary (or monetary) cost.

So, perhaps the high cost of re-formatting journal articles for submission to journals is a signalling mechanism that acts as a way of sifting out the low-quality journal submissions before they start tying up editor (and peer reviewer) time? Would moving to a system where the authors can submit in any sensible format for the initial submission actually be an improvement, or would it tie up more resources in unnecessary editorial and peer review? It would be interesting to see some analysis of the experience of journals that have adopted a more open formatting approach for first submission, in terms of the quality of submissions (and the quality of published articles).

How insurers can use data to beat adverse selection and moral hazard

This week, my ECONS102 class has been covering the economics of information. In particular, we focus on the problems of information asymmetry, and we spend a fair amount of time talking through problems of adverse selection. Adverse selection arises when one of the parties to an agreement (the informed party) has private information that is relevant to the agreement, and they use that private information to their own advantage at the expense of the uninformed party.

A classic example of adverse selection, which I've blogged about many times, occurs in the market for insurance (regardless of whether we are discussing home insurance, car insurance, health insurance, or even life insurance). The insured person knows whether they are high risk or low risk, but the insurer doesn't know - risk is private information. Since the insurer doesn't know how risky any person applying for insurance is, their best option is to assume that everyone is high risk. We refer to this as a pooling equilibrium - all insurance applicants are pooled together as if they are the same risk. The insurer then sets the insurance premium on the basis of the risk pool they think they have (high risk). The low risk people will (rightly) identify that the insurance premium is too high for them, and they drop out of the market, leaving only high risk people buying insurance. The insurance market for low risk people fails - they can't by insurance if they can't credibly convince the insurer that they are low risk. This problem is referred to as adverse selection, because the people who select into applying for insurance are the people that the insurer least wants to insure!

As you know, we do have insurance markets that cater to low risk people, so the markets must have adapted to deal with this adverse selection problem. This involves the private information (about the level of risk) being credibly revealed to the uninformed party (the insurer). If the insurer tries to reveal the private information, or tries to induce the informed party (the person applying for insurance) to reveal the private information, this is referred to as screening.

Insurers can screen applicants on the basis of their demographic and other information that they provide when they apply for insurance, their insurance history or credit history, and details about what they are insuring (house, car, health, life, etc.). However, insurers are increasingly using online data to screen applicants and determine their risk. Take this example, from The Wall Street Journal (gated) last year:

"Did you document your hair-raising rock-climbing trip on Instagram? Post happy-hour photos on Facebook? Or chime in on Twitter about riding a motorcycle with no helmet? One day, such sharing could push up your life insurance premiums.

In January, New York became the first state to provide guidance for how life insurers may use algorithms to comb through social media posts—as well as data such as credit scores and home-ownership records—to size up an applicant’s risk. The guidance comes amid expectations that within years, social media may be among the data reviewed before issuing life insurance as well as policies for cars and property.

If you're not thinking about how much information you reveal on social media, perhaps you should be now that it might cost you more in terms of insurance (on the other hand, if you are a low risk person, then perhaps your social media posts will earn you a lower insurance premium). However, that isn't the end of insurance companies' use of data.

Another information asymmetry problem in insurance happens after the insurance contract is agreed. This is the problem referred to as moral hazard - this problem arises when one of the parties to an agreement has an incentive, after the agreement is made, to act differently than they would have acted without the agreement. In the case of insurance, the insured party might act in a more risky manner when they are insured than they would have acted without insurance. They can do this because they have passed some of the (financial) risk of their actions onto the insurer.

One solution to moral hazard problems is for the uninformed party (the insurer) to monitor the actions of the informed party (the insured) more closely. And, you guessed it - insurers are looking at data to deal with moral hazard problems. As one example, Sven Tuzovic (Queensland University of Technology) wrote in The Conversation last year that:

...wearable devices are not only being embraced by consumers, but also across insurance industries. Health and life insurance companies collect data from fitness trackers with the goal of improving business decisions.

Currently, these business models work as a “carrot” incentive. That means consumers can benefit from discounts and cheaper premiums if they are willing to share their Fitbit data.

But we could see voluntary participation become mandatory, shifting the incentive from carrot to stick. John Hancock, one of the largest life insurance companies in the United States, has added fitness tracking with wearable devices to all of its policies. Though customers can opt out of the program, some industry experts argue that this “raises ethical questions around privacy and equality in leaving the traditional life insurance model behind”.

In terms of moral hazard, the insured is less likely to engage in risky behaviour if they know that their insurer is watching their every move. Insurers can use the data they collect from devices like Fitbit to not only monitor the insured, but also to determine their risk and adjust future premiums. It potentially solves both moral hazard and adverse selection problems at the same time.

And this is just the beginning. Insurers may turn to more sophisticated artificial intelligence tools in the near future, as David Tuffley (Griffith University) wrote in The Conversation last year:

Then you have a car accident. You phone your insurance company. Your call is answered immediately. The voice on the other end knows your name and amiably chats to you about your pet cat and how your favourite football team did on the weekend.

You’re talking to a chat-bot. The reason it “knows” so much about you is because the insurance company is using artificial intelligence to scrape information about you from social media. It knows a lot more besides, because you’ve agreed to let it monitor your personal devices in exchange for cheaper insurance premiums.

This isn’t science fiction. More than three-quarters of insurance executives believe artificial intelligence will revolutionise the industry within a few years. By 2030, according to McKinsey futurists, artificial intelligence will mean your car and life insurance premiums could change based on whether you decide to take one route or another.

If you're starting to think that there is nowhere to hide, you're right. Even if you refuse to let your insurer access your data, you're simply suggesting to the insurer that you are high risk. The insurer may frame it as if those agreeing to share data are receiving a discount, but really they are applying a higher premium to the high risk people who are least likely to want to share their data.

Should we be worried? Arguably no, unless we are high risk people wanting to pass ourselves off to insurers as low risk. Otherwise, we get insurance priced at premiums that is actuarially fair and accurately reflects our level of risk. However, as Tuffley notes, we should be concerned about what happens to the data that insurers collect about us:

An insurer might also be tempted to use the data for purposes other than assessing risk. Given its value, the data might be sold to third parties for various purposes to offset the cost of collecting it. Advertisers, marketers, lobbyists and political parties are all insatiably hungry for detailed demographic data.

It pays to read the fine print on contracts, and if insurers are going to collect much more data about us in the future, we should at least be aware of the risks of what will happen to that data.

[HT: The Dangerous Economist last year, for the Wall Street Journal article]

Read more:

• Your Fitbit will betray you
• The most surprising thing I learned about home insurance this year
• Could your social media posts make insurance more expensive?
• Why your Fitbit or Apple Watch could soon get you cheaper health insurance
• Could technology eliminate moral hazard in car insurance?
• Is there adverse selection in the life insurance market?

A cynical take on The Body Shop's 'open hiring' policy

In the world before coronavirus lockdowns, this article on Fast Company caught my attention, and I've been meaning to blog about it for a while:

Almost all retailers run background checks on prospective employees—one of the many obstacles for people who were formerly incarcerated and are now trying to find a job. For other job seekers, a drug screening for marijuana might cost them a position even in states where recreational use is legal. This summer, the Body Shop will become the first large retailer to embrace a different approach, called “open hiring.” When there’s an opening, nearly anyone who applies and meets the most basic requirements will be able to get a job, on a first-come, first-served basis.

The company piloted the practice, which was pioneered by the New York social enterprise Greyston Bakery, in its North Carolina distribution center at the end of 2019...

The results were striking: Monthly turnover in the distribution center dropped by 60%. In 2018, the Body Shop’s distribution center saw turnover rates of 38% in November and 43% in December. In 2019, after they began using open hiring, that decreased to 14% in November and 16% in December. The company only had to work with one temp agency instead of three.

I can immediately see two reasons why this might be a good approach for The Body Shop, and neither reason has anything to do with the 'halo effect' of appearing to be a good employer. Both reasons have to do with wages (which, if you read the story, you will notice are not mentioned anywhere).

First, consider a search model of the labour market. Each time a job is filled, this creates a match between the employer and a worker, and that match creates a surplus (the employer receives some additional profits from employing the worker). The employer and the worker share the surplus. How it is shared depends on their relative bargaining power. If the employer has less bargaining power and the worker has more, then the worker can demand a bigger share of the surplus, and wages will be higher. On the other hand, if the employer has more bargaining power and the worker has less, then the employer can offer a lower wage, and wages will be lower.

Now, think about The Body Shop's open hiring. Suddenly, they are willing to accept applications from almost anyone. Applications flood in. The Body Shop has lots of potential workers to choose from. If Worker A isn't willing to accept the wage that The Body Shop offers, then they can simply move on to Worker B, or Worker C, or any of the countless other applicants. Can you see that this 'open hiring' shifts the bargaining power in favour of The Body Shop? They can offer lower wages because the applicant pool is much larger than before.

Second, consider asymmetric information and adverse selection. Workers know how productive they are (more or less), but the employer doesn't know - this is private information. The employer usually wants to hire the most productive workers. So, they try to reveal the private information through screening. The background checks, and job interviews, and psychometric tests and whatever else that human resources people dream up are all ways of screening job applicants in order to determine which of them are most likely to be highly productive workers for the employer.

Now, consider what happens if you remove the screening tools like background checks. Now the employer can't tell the more productive and less productive applicants apart. Basically, they would have to assume that all applicants are the same, and the safer assumption to make is that they are all of the less productive type - we refer to this as a pooling equilibrium. Normally, we'd consider a pooling equilibrium to be bad - employers want to tell the more productive and less productive workers apart. However, if an employer had to assume that all workers were of the less productive type, then their best option is to offer a lower wage (why pay a high wage to workers who are likely to have low productivity?).

So, now you can see that there are two good (albeit cynical) reasons to see why The Body Shop's 'open hiring' philosophy makes good business sense. That's not the end of the story though. Background checks are a fairly imperfect screening tool, and even job interviews and psychometric tests can get things wrong. A better screening tool is to let the worker do some work, and observe their productivity. If the employer can then easily fire any less-productive workers, then this may be a more accurate way to identify the most productive job applicants.

In the U.S., many states have 'at-will employment laws', which basically mean that the employer can fire a worker at any time for any reason (there are exemptions, and the actual laws vary by state). It is interesting to note that North Carolina is a state with at-will laws, and North Carolina is where The Body Shop's distribution centre that trialled 'open hiring' is located.

The Body Shop may well have gotten a lot of positive press with this move, but it's likely that is not the only benefit they have enjoyed from 'open hiring'.

[HT: Marginal Revolution]

The 90-day trial, retail employers, and incentives

My nephew just lost his job, working for a large national retailer. He was employed sometime in November, and let go in early January. I don't know all the details, but the circumstances did get me thinking. So, I'm going to put my nephew's case aside, and write about a more general situation in relation to employee trial periods, that occurred to me when thinking about his case.

Small retailers (19 or fewer employees) in New Zealand can use 90-day trials, as long as this is specified in the written employment agreement (larger employers can't use these trials). Retailers face a couple of problems in terms of staffing. First, they can't easily tell from among the many applicants for jobs, who will be the most productive employees and who will fit into the workplace culture. This is classic asymmetric information - job applicants know if they are hard workers, but potential employers do not. The job interview process is a way of screening applicants, but it is imperfect. So, the 90-day trial gives employers a second chance to observe their new hires at work and see how productive they are. Second, retailers require additional staff over the busy holiday period, stocking shelves and attending to customers, etc.

It seems to me that the 90-day trial gives small employers an incentive to use the rules to solve both problems. Let's say the employer needs three extra staff over the holidays, but they also know that they need one permanent employee. They hire four applicants. However, before they hire them, the employer has no idea which one of the four will be the most productive, or the best fit for the workplace (or whatever other criteria they may have). So, instead of deciding on one of the new hires to put on a permanent contract, they put them all on permanent contracts. Then, at the end of the trial period, they let three of them go (and they don't have to give reasons), keeping the 'best' one of the four. This is what economists refer to as a tournament. Now, here's the important bit: the employer doesn't tell the four new hires that they are competing for one job.

How does this solve the two problems, and are their better alternatives? It solves the second problem obviously, because there are four workers for the busy holiday period. It solves the first problem, because the four workers will reveal their productivity during the trial period, since they know they are on trial and they want to keep the job when the trial period ends. Since they don't know they are competing with the others for a single job, they have less incentive to work in competitive ways that are unhelpful (such as sabotaging the other new hires' work), than they would if they knew they were in a tournament.

What about a simpler alternative? What if they employer put all of the workers on fixed term contracts for the holiday period, and then told them that there is a chance to be hired permanently at the end? The new hires would still work hard, but the tournament might become more obvious to them, and as above that might result in unhelpful competition. Also, there might be an end-of-game problem. If the new hires know that their fixed term is about to expire, and they suspect that their prospects of getting the permanent job are low, then they have less incentive to work hard. If they are on a trial period and don't know about the tournament, then there is less likely to be an end-of-game problem.

Anyway, I thought this was interesting, and probably wasn't an anticipated effect of the 90-day trials. It would only work for retail, or for other jobs where there is a short-term (certainly, less than 90 days) period of high demand for labour, followed by a lower base level of demand through the rest of the year. Hospitality jobs in some tourist areas that have a single high season (or two) might also qualify. I don't think this is necessarily a negative impact of the law (for example, I wouldn't label it an 'unintended consequence', as I have for policies that have unintended negative effects such as this one). After all, it helps the employer to better identify the 'best' employee for the available permanent vacancy, and that's what the law was designed to do. It's just that retailers can be much more efficient in using the law than most employers, who could only hire a single applicant at a time and try them out.

Your Fitbit will betray you

Yesterday I wrote a post about how home insurers are starting to more accurately price house insurance based on natural hazard risk. Home insurance isn't the only area where insurers are looking at adopting more sophisticated screening methods to deal with adverse selection. Take this story from the New Zealand Herald in June:

Fitbits are already used to track your heart rate, the amount of exercise you do and how much you sleep - essential data that could potentially be used by insurance providers to determine your premiums.

The boom in wearable health tracking technology means we now have more information than ever before on health and well being of people at any given moment.

The Telegraph reports that information collected from these devices is already being used by insurers to calculate insurance premiums and there are concerns that this might lead to only the healthiest customers enjoying lower premiums.

This is serious business. Insurance companies have it in their interests not only to ensure the lowest-risk customers but also to detect potential health conditions before they become severe (and expensive). A study of the insurance market by the Swiss Re Institute, a research organisation, last year found that insurers had filed hundreds of patent applications relating to "predictive insurance modelling".

The issue that an uninformed health insurer or life insurer faces is essentially the same as the home insurer from yesterday's post. They can't tell the low-risk applicants from high-risk applicants. A pooling equilibrium develops, where everyone pays the same premiums (coarsely differentiated based on age, gender, and smoking status). A savvy and entrepreneurial insurer that was better able to tell who the low-risk insured people are could attract them away with lower premiums (knowing that they would cost less to insure because they are low risk).

So, that is effectively what insurers are starting to do. As the Herald article notes:

In making these moves, Insurance companies aim to collect data that could serve to help them make better policy decisions or even tweak existing policies over time.

The Telegraph reported that policy agreements increasingly feature clauses that allow insurers to collect data on their customers.

This is a point that I first made in a post back in 2015 (and an earlier post on technology in car insurance in 2014). We can all look forward to insurers asking for our Fitbit data when we apply for health or life insurance. And if we're fit and healthy, we'll give it to them. The people most likely to withhold that information are the unfit and unhealthy (and those who are most privacy-conscious). Denying access to your Fitbit data would probably be enough to signal to the insurer that you are high risk, and result in a declined application or a higher premium. So, even if you want to opt out of sharing your data, your Fitbit will still betray you.

Read more:

• The most surprising thing I learned about home insurance this year
• Why your Fitbit or Apple Watch could soon get you cheaper health insurance
• Could technology eliminate moral hazard in car insurance?
• Is there adverse selection in the life insurance market?

The most surprising thing I learned about home insurance this year

Home insurance markets are subject to adverse selection problems. When a homeowner approaches an insurer about getting home insurance, the insurer doesn't know whether the house is low-risk or high-risk. [*] The riskiness of the house is private information. In fact, the riskiness of the house is probably not known even to the homeowner, but let's assume for the moment that they have at least some idea. To minimise the risk to themselves of engaging in an unfavourable market transaction, it makes sense for the insurer to assume that every house is high-risk. This leads to a pooling equilibrium - low-risk houses are grouped together with the high-risk houses and owners of both types of house pay the same premium, because they can't easily differentiate themselves. This creates a problem if it causes the market to fail.

The market failure may arise as follows (this explanation follows Stephen Landsburg's excellent book The Armchair Economist). Let's say you could rank every house from 1 to 10 in terms of risk (the least risky are 1's, and the most risky are 10's). The insurance company doesn't know who is high-risk or low-risk. Say that they price the premiums based on the 'average' risk ('5' perhaps). The low risk homeowners (1's and 2's) would be paying too much for insurance relative to their risk, so they choose not to buy insurance. This raises the average risk of the homes of those who do buy insurance (to '6' perhaps). So, the insurance company has to raise premiums to compensate. This causes some of the medium risk homeowners (3's and 4's) to drop out of the market. The average risk has gone up again, and so do the premiums. Eventually, either only highest risk homeowners (10's) buy insurance, or no one buys it at all. This is why we call the problem adverse selection - the insurance company would prefer to insure low-risk homes, but it's the homeowners with high-risk homes who are most likely to buy.

Of course, insurers are not stupid. They've found ways to deal with this adverse selection problem. When the uninformed party (the insurer in this case) tries to reveal the private information (about the riskiness of the house), we refer to this as screening. Screening involves the insurer collecting information about the house and the homeowner in order to work out how risky the house is. With the private information revealed, the insurer can then price accordingly - higher-risk houses attract higher premiums, while lower-risk houses attract lower premiums. We have a separating equilibrium (the high-risk and low-risk houses are separated from each other in the market).

With all this in mind, this story from April surprised me greatly:

Other insurers are likely to follow NZX-listed Tower's lead and increase their focus on risk-based pricing for natural hazards, says an insurance expert...

Thousands of home-owners who live in high-risk earthquake-prone areas and insure via Tower are set to face hikes in their premiums while those in low-risk areas like Auckland will get a cut.

The insurance company, which is New Zealand's third largest general insurer, said it would stop cross-subsidising its policy-holders from April 1 in a bid to send a clearer message to home-owners about the risks of where they lived.

Tower chief executive Richard Harding said at the moment six Auckland households were paying more to subsidise insurance premiums for every one high-risk house in Wellington, Canterbury or Gisborne.

In other words, insurers previously weren't screening for all available private information before pricing their insurance for a given house. Essentially, owners of low-risk houses have been paying premiums that are too high, and owners of high-risk houses have been paying premiums that are too low. It took a little while, but eventually other insurers have also started to use risk assessments in determining insurance premiums, so this discrepancy is disappearing.

Why didn't the market break down due to adverse selection? The issue here is something I noted earlier in the post - the riskiness of a house is private information to both the insurer and the homeowner. If the homeowner doesn't know how risky their house is, owners of low-risk houses can't tell if the insurer is pricing their insurance too high relative to the risk of natural hazard damage. So, the owners of low-risk houses have no reason to drop out of the market. And, if the owners of low-risk houses don't drop out of the market, insurers have no reason to raise premiums.

However, that leaves the market open to disruption. As noted in the April article:

Jeremy Holmes, a principal at actuarial consulting firm Melville Jessup Weaver, said it was hard to say how long this would take. Insurers needed to be as good as their competitors at distinguishing risk.

"Otherwise they risk having their competitors target the lower-risk policyholders whilst they are left with only the higher risks ..."

An entrepreneurial insurer that was able to distinguish the low-risk houses from high-risk houses could start approaching owners of low-risk houses and offering them lower premiums. The remaining insurers would be left with higher-risk houses on average, and would have to raise premiums. This would increase the number of homeowners dropping out of the market (or rather, going to the insurer that was pricing according to risk). Tower was the first insurer to shift to risk-based premiums, so presumably they recognised this issue before any of the other insurers and acted exactly as you would expect - by moving to risk-based premiums before any potential disruptor could enter the market.

Still, it's a little surprising (to me, at least) that pricing based on natural hazard risk wasn't already happening.

*****

[*] For simplicity, I'm going to refer to low-risk houses and high-risk houses, when risk is probably as much a function of location as of the house itself. So, if you must, read 'low-risk house' as 'house with a low risk of damage in an earthquake', and 'high-risk house' as 'house with a high risk of damage in an earthquake'.

Car insurance, adverse selection and gender

Car insurance has an adverse selection problem. The uninformed party (the insurer) cannot tell those with 'good' attributes (low-risk people) from those with 'bad' attributes (high-risk people). To minimise the risk to themselves of engaging in an unfavourable market transaction, it makes sense for the insurer to assume that everyone is high-risk. This leads to a pooling equilibrium - low-risk people are grouped together with the high-risk people and pay the same premium, because they can't easily differentiate themselves. This creates a problem if it causes the market to fail.

To see how the car insurance market fails, consider this from an earlier post about health insurance (it applies just as well to car insurance):

In the case of insurance, the market failure may arise as follows (this explanation follows Stephen Landsburg's excellent book The Armchair Economist). Let's say you could rank every person from 1 to 10 in terms of risk (the least risky are 1's, and the most risky are 10's). The insurance company doesn't know who is high-risk or low-risk. Say that they price the premiums based on the 'average' risk ('5' perhaps). The low risk people (1's and 2's) would be paying too much for insurance relative to their risk, so they choose not to buy insurance. This raises the average risk of those who do buy insurance (to '6' perhaps). So, the insurance company has to raise premiums to compensate. This causes some of the medium risk people (3's and 4's) to drop out of the market. The average risk has gone up again, and so do the premiums. Eventually, either only high risk people (10's) buy insurance, or no one buys it at all. This is why we call the problem adverse selection - the insurance company would prefer to sell insurance to low risk people, but it's the high risk people who are most likely to buy. 

In order to solve an adverse selection problem, the uninformed party can try to reveal the private information - this is referred to as screening. Car insurance companies engage in screening by collecting information about every person who applies for insurance, including their demographic details and accident and insurance history. The insurers know that this information provides some clues as to who is higher risk and who is lower risk to insure. They can then separate the higher-risk and lower-risk groups, and we move from a pooling equilibrium to a separating equilibrium. Higher-risk drivers will pay higher car insurance premiums, and lower-risk drivers will pay lower premiums.

One of the key demographic variables that is associated with car insurance risk is gender. Women drivers are less risky to insure. On average, they have accidents at lower speeds, which are less costly to repair. So, female car owners pay lower car insurance premiums on average. Gender works as one screening tool because it is difficult to fake. Or it was, until this story from CBC in Canada in July:

He wanted a brand new car — a Chevrolet Cruze with all the trimmings.

As a man in his early 20s, he knew his insurance costs would be high.

So he became a woman, though only on paper.

"I have taken advantage of a loophole," said the man — we're calling him David — who spoke on the condition that his identity be kept confidential because of the potential repercussions...

After doing some research, he realized he needed a doctor's note to show the government he identifies as a woman, even though he doesn't.

"It was pretty simple," he said. "I just basically asked for it and told them that I identify as a woman, or I'd like to identify as a woman, and he wrote me the letter I wanted."...

David shipped the note and other paperwork off to the provincial government. And, a few weeks later, he received a new birth certificate in the mail indicating he was a woman.

"I was quite shocked, but I was also relieved," he said. "I felt like I beat the system. I felt like I won."

With the new birth certificate in hand, he changed his driver's licence and insurance policy.

All to save about $91 a month.

"I'm a man, 100 per cent. Legally, I'm a woman," he said.

"I did it for cheaper car insurance."

"David" may have just raised the cost of car insurance for all women in Canada. If car insurers can no longer believe from a drivers licence that an insurance applicant is a woman, then every gender will be pooled together as being the same risk. That means higher premiums for women (and ironically, lower premiums for men, which is what David wanted!).

[HT: Marginal Revolution]

Book Review: Weapons of Math Destruction

I just finished reading the book "Weapons of Math Destruction" by Cathy O'Neil, better known online as mathbabe. The premise of the book is simple - our life decisions and outcomes are increasingly influenced if not determined by algorithms or mathematical models, and often not in a good way. O'Neill calls these models 'Weapons of Math Destruction' (WMDs). For O'Neill, a WMD has a number of characteristics, including:

1. The model is opaque, or invisible - people whose data are included in the model (let's call them 'subjects') don't know how the model functions, or understand how the inputs are converted into outputs;
2. The model is unfair, or works against the subject's best interests; and
3. The model is scalable - it can grow exponentially.

The WMDs that are considered in the book include college admissions, online advertising, justice and sentencing, hiring and monitoring of employees, credit, insurance, and voting. The book is generally well-written and doesn't lack for real-world examples. Consider this passage on online for-profit universities in the U.S.:

The marketing of these universities is a far cry from the early promise of the Internet as a great equalizing and democratizing force. If it was true during the early dot-com days that "nobody knows you're a dog," it's the exact opposite today. We are ranked, categorized, and scored in hundreds of models, on the basis of our revealed preferences and patterns. This establishes a powerful basis for legitimate ad campaigns, but it also fuels their predatory cousins: ads that pinpoint people in great need and sell them false or overpriced promises. They find inequality and feast on it. The result is that they perpetuate our existing social stratification, with all of its injustices.

One of the big takeaways for me was how often models are constructed and used on insufficient data. O'Neill gives the example of teacher value-added, which works well in aggregate but when you consider it at the level of an individual teacher, their annual 'value-added' is based on perhaps 25 data points (the students in their class). Hardly a robust basis for making life-changing hiring and firing decisions.

However, there are also parts of the book that I disagree with. In the chapter on insurance, O'Neill argues against using models to predict risk profiles for the insured (so that those of higher-risk would pay higher premiums than those of lower-risk). O'Neill argues that this is unfair, since some people who are really low-risk end up grouped with higher-risk people and pay premiums that are too high (e.g. think of a really-careful young male driver, if there is such a person). O'Neill could do with a greater understanding of the problems of asymmetric information, adverse selection, and signalling. Having larger pools of insured people that include both low-risk and high-risk people paying the same premium leads to higher premiums. Essentially the low-risk people are subsidising the high-risk people (how is that not more unfair?). The low-risk people drop out of the market, leaving only the high-risk people behind, who are not profitable for insurers to insure. The insurance market could eventually collapse (for a more detailed explanation, see this post). The insurer's models are a way of screening the insured, revealing their private information about how risky to insure they are.

Despite that, I really enjoyed the book. The chapter on voting and micro-targeting of political campaigns in particular was very enlightening - particularly given more recent events such as the 2016 U.S. Presidential election and the Brexit vote. The sections on the teacher value-added models have caused me to re-evaluate the strengths and weaknesses of those models.

Finally, O'Neill argues for a more ethical use of models. This isn't so far from what Emmanuel Derman argues in his book "Models. Behaving. Badly." (which I reviewed here back in 2013), and is a cause well worth supporting.